The Report is Generated by DrillBit Plagiarism Detection Software

Submission Information

Author Name Viswanathan

Title Segmentation and classification
Submission/Paper ID 368241
Submission Date 26-Aug-2021 11:19:10
Total Pages 172
Total Words 42530

Result Information

Similarity 11 %
Unique 89 %
Internet Sources 4%
Journal/Publication Sources 6%
Total content under 'References' 3%
Total content under 'Quotes' 2%

Exclude Information

References/Bibliography Excluded
Quotes Not Excluded
Sources: Less than 14 Words Similarity Not Excluded
DrillBit Similarity Report

A-Satisfactory (0-10%)
B-Upgrade (11-40%)

11 71 B C-Poor (41-60%)
D-Unacceptable (61-100%)
SIMILARITY % MATCHED SOURCES GRADE

Sl.No LOCATION MATCHED DOMAIN % SOURCE TYPE

1. 16 Internet
medium.com 2

2. 34 Internet
www.sciencedirect.com 1

3. 41 Publication
Alzheimers Disease Diagnosis Based on Multiple Cluster 1
Dense Convolu, by Li, Fan Liu, Manhu- 2018

4. 30 Publication
Hippocampus Analysis by Combination of 3D DenseNet <1
and Shapes for Alz, by Cui, Ruoxuan Liu, - 2018

5. 21 Publication
Classification of Alzheimers Disease by Combination of <1
Convolutional by Liu-2018

6. 28 Publication.
A novel statistical technique for intrusion detection <1
systems by Kabir-2017

7. 18 Publication
Automatic Voxel-Based Morphometry of Structural MRI <1
by SPM8 plus Diffe by Matsuda-2012

8. 26 Publication
IEEE 215 International Conference on Advances in <1
Computing, Communi by
9. 10 Publication
A data-driven approach to prostate cancer detection from <1
dynamic contr by Haq-2015

10. 5 Publication
Thesis submitted to shodhganga - <1
shodhganga.inflibnet.ac.in

11. 37 Publication
Improved Highway Network Block for Training Very <1
Deep Neural Networks by Oyedotun-2020

12. 25 Publication
Discriminative analysis of multivariate features from <1
structural MRI a by Li-2014

13. 6 Publication.
A generic post-deblocking filter for block based image <1
compression algorithms by Nelso-2012

14. 27 Publication.
An Ontological Clinical Decision Support System Based <1
on Clinical Guidelines for by Madhusanka-2020

15. 7 Publication
ijece.iaescore.com <1

16. 23 Internet
www.ncbi.nlm.nih.gov <1

17. 63 Internet
www.ncbi.nlm.nih.gov <1

18. 24 Publication.
MRI Segmentation and Classification of Human Brain <1
Using Deep Learning for Diagn by Yamanakkanavar-2020

19. 44 Publication
www.ijitee.org <1
20. 1 Publication
A novel feature selection method for brain tumor MR <1
image classificati by Kaur-2017

21. 35 Publication
cames.ippt.pan.pl <1

22. 15 Internet
www.bitrine.com <1

23. 8 Publication
Comparative analysis of nucleus segmentation techniques <1
for malignancy by Shanthi-2018

24. 17 Internet
worldwidescience.org <1

25. 51 Publication
Topological graph kernel on multiple thresholded <1
functional connectivity network by Jie-2014

26. 36 Internet
moam.info <1

27. 12 Publication
Automatic contrast enhancement of brain MR images <1
using Average Inten, by Isa, Iza Sazanita - 2017

28. 45 Publication
Using the ACO algorithm for path searches in social <1
networks by Jessic-2012

29. 54 Publication ()
Ensembles of Deep Learning Architectures for the Early <1
Diagnosis of t, by Ortiz, Andrs Muni- 2016

30. 52 Publication.
Interval type-2 fuzzy kernel based support vector machine <1
algorithm for scene cl by Zh-2014
31. 65 Publication
Differential evolution algorithm with ensemble of <1
parameters and mutation strate by R-2011

32. 4 Publication
Complex-Valued Neural Networks (Advances and <1
Applications) Quaterni by

33. 14 Internet
www.ncbi.nlm.nih.gov <1

34. 2 Publication
Partial least squares for discrimination in fMRI data by <1
Ander-2012

35. 33 Internet
moam.info <1

36. 56 Publication.
Maximizing the Diversity of Ensemble Random Forests <1
for Tree Genera Classificati by Ko-2016

37. 40 Publication
www.frontiersin.org <1

38. 19 Internet
onlinelibrary.wiley.com <1

39. 20 Publication
Identification of Alzheimers Disease on the Basis of a <1
Voxel-Wise Approach by Beheshti-2019

40. 47 Publication
www.futa.edu.ng <1

41. 38 Publication
www.dx.doi.org <1

42. 58 Publication
www.mdpi.com <1
43. 55 Publication
DeFine deep convolutional neural networks accurately <1
quantify intens, by Wang, Meng Tai, Ch- 2018

44. 42 Internet
journals.plos.org <1

45. 32 Publication
repository.up.ac.za <1

46. 31 Publication
Deep learning neural networks for spatially explicit <1
prediction of flash flood p by Panahi-2020

47. 13 Publication
Artificial Intelligence Methodologies Applied to Prompt <1
Pluvial Flood Estimation by Chang-2020

48. 71 Publication
Thesis submitted to shodhganga - <1
shodhganga.inflibnet.ac.in

49. 39 Publication.
Newly Assembled Microtubules Are Concentrated in the <1
Proximal and Distal Regions by Anthon-1992

50. 3 Publication
Interictal patchy regional cerebral blood flow patterns in <1
migraine patients by L-1994

51. 48 Publication
Detection of magnetite in the Roossenekal area of the <1
Eastern Bushveld Complex, by Twala-2020

52. 43 Publication
www.frontiersin.org <1

53. 46 Publication
Lecture Notes in Computer Science Artificial Intelligence <1
and Soft Computing V
54. 50 Publication
www.ijrte.org <1

55. 61 Publication
Durable power of attorney and informed consent with <1
Alzheimers disease patients by -1997

56. 64 Publication
Structural MRI-based detection of Alzheimers disease <1
using feature ra by Beheshti-216

57. 67 Publication
www.nature.com <1

58. 70 Publication.
Neuropsychological Differentiation of Dementia with <1
Lewy Bodies from Normal Agin by Ferman-2006

59. 66 Publication
repository.up.ac.za <1

60. 11 Internet
www.dx.doi.org <1

61. 53 Internet
link.springer.com <1

62. 57 Internet
moam.info <1

63. 59 Publication
bmcpsychiatry.biomedcentral.com <1

64. 9 Publication
Differences between Gray Matter and White Matter Water <1
Diffusion in Stroke Diff by Mukherjee-2000

65. 60 Internet
www.ijert.org <1

66. 22 Publication
www.frontiersin.org <1
67. 62 Internet
www.ncbi.nlm.nih.gov <1

68. 69 Publication
Thesis submitted to shodhganga - <1
shodhganga.inflibnet.ac.in

69. 29 Publication
European Association of Nuclear Medicine Congress 1014 <1
October 1993 by -1993

70. 49 Internet
geography.exeter.ac.uk <1

71. 68 Internet
www.e-jmd.org <1
 Chapter - 1
1.0 INTRODUCTION

The Machine Learning (ML) and Image Processing (IP) are perfect tools used to extract
the image for tangible information. The two-dimensional view of image representation
1
determines a set of values is called pixel which is rightly concerned as a picture element. Pixel
values represent the gray levels or intensities, colors, heights and similar other things. The
conversions of images are in digital form. The significant motivation beyond Image Processing is
Visualization, Sharpening and Restoration, Image Retrieval, Measurement of Pattern and Image
Recognition [1]. It was a most developing innovation. The picture handling has three steps in
importing the image are image acquisition tools, analyzing and manipulating the image,
depending on image analysis of the image result. The Digital Image Processing (DIP) is being
utilized in numerous fields like medicine, sports, astronomy, remote sensing, industries,
entertainment and defense [2]. It has two methods called analog and digital image processing
techniques. Analog image processing is also called as Visual techniques. It gives a result as
hardcopies format like printouts and pictures. In digital processing techniques, the images are
manipulating as the digital images. The use of picture handling is canny transportation
framework, remote detecting, moving article following resistance observation, biomedical
imaging methods, and programmed visual examination framework [3].

Clinical imaging is a non - obtrusive device. This tool explores the abnormal parts of the
body. Clinical imaging involves capturing images using different applications of instruments. It
is used for detecting clinical abnormalities on anatomy of brain in citizenry. The clinical imaging
strategies are Ultrasound, Computerized Tomography (CT), Magnetic Resonance Imaging
2
(MRI), functional Magnetic Resonance Imaging (fMRI), Positron Emission Tomography (PET),
3
Single Photon Emission Computerized Tomography (SPECT), Magneto Encephalography
(MEG), and similar other things. The recognizable proof of bizarre tissues with high resolution
facilitates medical imaging techniques [6]. A series of images of identical tissues with different
4
contrast visualization through the application of diverse image acquisition protocols and
parameters are the advantages of Magnetic Resonance Imaging.

1
 5
The numerous magnetic resonance images are helpful to find out the whole brain
6
anatomical realities of an equivalent tissue and a lot of contrast techniques help to review the
brain pathology more precisely to researchers [4]. Image segmentation is one among the
7
challenging operations in image processing and machine learning. Image segmentation is a
manual or automatic partitioning of the image. The partitioning image is converted into a group
of relatively identical regions, with similar properties, each of which is labeled consistent with
8
sort of regions within the image segmentation process. The segmentation methods are threshold,
region growing, edge detection and clustering. In the algorithm, less variance of intensity
between abnormal and normal tissues may result in confusion. The lesion happens to be inside
the White Matter (WM), intensity distributions overlap between White Matter Lesion (WML),
9
and Gray Matter (GM). The lesion do not help to discriminate some areas of the White Matter
Lesion from Gray Matter is Non-parametric or multi-parametric statistical classification
techniques. The partial volume pixels which contain a mixture of tissues are known as the border
pixels. The focal point of a brain abnormality is usually different from its edge or boundary in
10
the image intensity. The borders of the brain abnormalities could also be an equivalent as Gray
11
Matter. This appearance is the reason for confusion and misclassification of the bordering or
7
corner regions where the Cerebral Spinal Fluid (CSF) and Gray Matter remain together and may
present the pixel or voxel intensities which are typical to brain abnormalities.
Alzheimer’s disease is Neurodegenerative sorts of Dementia and commonest explanation
for Dementia, constituting sixty percent to ninety percent of all Dementias. Alzheimer’s disease
12
is found to be one among many of striking cause of death in many countries. It causes various
13
problems with memory, critical thinking and personal behavior. It is a special case of dementia.
“A most general term for amnesia and other cognitive abilities serious enough to interfere with
daily lifestyle is named Alzheimer’s malady or Dementia”. The symptoms of Alzheimer’s are
thinking capacity is too slow and other occasional problems with remembering several things.
14
The falling signs of brain cells are amnesia, much confusion and other difficulty changes within
the path of minds work. The earlier symptoms of Alzheimer’s are difficult to sooner memorize
new information. Bewilderment, State of Mind and Behavioral Changes, Confusion about
events, time and place, difficulty in speaking, gulping and strolling are the side effects.
Image segmentation is a commonly used technique in digital image processing and
analysis to partition an image into multiple parts or regions, often based on the characteristics of

2
the pixels in the image. The Image segmentation is done to work in converting an image into a
3
collection of regions of pixels that are represented by a mask or a labeled image. The deep
15
learning for image segmentation using deep convolutional neural networks perform a deep
learning technique called semantic segmentation, it associate every pixel of an image with a class
label. Applications for semantic segmentation include autonomous driving, industrial inspection,
medical imaging, and satellite image analysis. Image classification is rightly refers to a process
16
in computer vision that can classify an image according to its visual content. The intent of the
classification process is to categorize all pixels in a digital image into one of several classes.
Normally, multi-spectral data are used to perform the classification and, indeed, the spectral
pattern present within the data for each pixel is used as the numerical basis for categorization.
The objective of image classification is to identify and portray, as a unique gray level (or color),
the features occurring in an image in terms of the object these features actually represent on the
ground. Image classification is perhaps the most important part of digital image analysis. The
17 13
supervised classification of deep convolutional neural network is a special kind of multi-layer
16
neural networks, designed to recognize visual patterns directly from pixel images with minimal
preprocessing. It is a special architecture of artificial neural networks. The deep convolutional
neural network uses some of its features of visual cortex and has therefore achieved state of the
art results in computer vision tasks. The figure 1.1 shows the Normal Control Brain and
Alzheimer’s disease Brain.

Figure.1.1 Normal Control Brain and Alzheimer’s disease Brain

3
 18
Early diagnosis and medication help the people the status of the disease. It is rising in
adult population in India. The census conducted in 2005 states that at least eighty five million
people with age sixty five and above had this disease. The rate of increase in Alzheimer’s disease
19
in India is higher compared to other developing countries. Moreover, the expansion rate is
20
increasing per annum. The figure 1.2 gives a forecast of increase of Alzheimer’s Disease in near
way forward for India.

Figure 1.2 Forecast of Dementia in India (Source: Alzheimer’s disease Report in World)
It was estimated that there would be a three hundred percent increase in Alzheimer’s
disease within the next forty to fifty years. Mild Cognitive Impairment (MCI) was a preceding
stage in Dementia [5]. If the disease is diagnosed at this stage, treatments are often planned and
thus can prevent the progression to Dementia. Diagnosis of Mild Cognitive Impairment is
therefore vital in preventing progression to Alzheimer’s disease and death due to Dementia. The
main symptom in this prodormal stage is the decline in cognitive skills. Patients are going to be
ready to perform everyday chores but there will be loss of cognitive abilities. The figure 1.3
21
shows the biomarkers for Alzheimer’s disease and their relevance in diagnosis of progression to
Alzheimer’s disease.

4
Figure 1.3 Biomarkers for Disease Progression in Alzheimer’s disease (Source: Alzheimer’s
disease Neuroimaging Initiative - ADNI)

1.1 MOTIVATION
15
In medical imaging the dynamic elements of a patient’s health information are related to
manipulating, handling and processing the details by employing the computers. The Computer
Assisted Radiology Development is used to manage the patient ammunition. Relevant documents
are used to improve future reference for diagnosis. In Magnetic Resonance images, the image
segmentation in biomedical research is confined to some specific tissues or particular region like
hippocampus in ‘S’ size. The medical image segmentation used to locate the hippocampal region
in brain is more accurately to physicians. If the specialist is not available in emergency case it is
a great source of information. Therefore, it is a crucial process in computerized medical imaging.

5
Forthwith, there are many issues in medical imaging based on hippocampus region in Dementia
in cerebrum brain on resonance image. The following are more significant:

 Huge volume of knowledge involved in examing Magnetic Resonance

Image.
 Heavy load of labor in computation process.
 Complex characteristics of gray matter tissues.
 High dimensionality of knowledge.
 Difficulty in classifier design and no prior knowledge of the patient’s
condition.
 Individual patient has large differences among the patients, who has
dementia.
 Establishing a totally automated system in non-feasibility.
 Overlapping the confines between normal and abnormal tissues in brain.

Using the image processing algorithms, the above issues are sorted out. The segmentation
methods may lead to new applications including content based image retrieval in large medical
dataset. The physicians are allowed to seek out similar images in dataset based on hippocampal
region in Alzheimers Disease location, grade, size and similar patterns of growth. Large datasets
22
are needed to make improvements in the processing of the image. Manually or automatically,
this process is done in the brain abnormality or hippocampus detection and segmentation [6].

A few drawbacks of manual segmentation are indicated below:

 In manual segmentation, the abnormal areas are manually located on all contiguous slices
in which the lesion is considered to exist.
 Time consuming is large in manual segmentation, where higher accuracy is required as
demanded by physician.
 Manual segmentation normally suffers from a scarcity of permanent availability,
reliability and reproducibility.
7
 To analyze the hippocampus size and border detection, the manual segmentation is
processed.

6
  The characteristics of gray matter tissues in hippocampus are extracted by using Region
of Interest in automatic or manual segmentation.
 Extracting the cluster in image segmentation is performed through Haar Wavelet
Transform.
 The Inference largely depends on subjective decisions which engage the probability that
23
the observers will reach different decisions on the presence or absence of abnormalities
or hippocampal tissues.
24
The prerequisite for computer assisted brain in Alzheimer’s disease and hippocampus
25
tissues segmentation from Magnetic Resonance Image are to be analyzed. Therefore, the fully or
semi, accurate and robust segmentation become building blocks for excellent research challenge
in the field of medical image analysis. An automatic segmentation determines the identical pixel
properties for its accuracy in extracting the objects from image. The pixel properties like grey
level or intensity, texture, color, brightness and contrast of images forms the elementary
principles for hippocampus segmentation in Magnetic Resonance brain image. In this task, the
26
image owes to high variation in the pixel properties. Improving the great performance of the
overall system’s accuracy, the scientists and practitioners have made tremendous efforts in
devising proficient segmentation approaches and techniques. Hippocampus region, segmentation
process, feature extraction and classification process using different methods and algorithm input
as Magnetic Resonance images are detection of Alzheimer’s disease.

1.2 OBJECTIVES
27
The principal activity of this research is to recognize, fragment and arrange the
Alzheimer’s infection in hippocampal with the help of elite division strategies and order
processing with high pinnacle precision. Recognition and Segmentation of Alzheimer’s disease
from MRI cerebrum pictures is the most critical task which are carried out with the help
28
computerized image processing methodologies. Numerous papers have been published on
Alzheimer’s disease from Magnetic Resonance Image cerebrum pictures with automated or semi
automated detection, segmentation and classification techniques [6]. The Alzheimer’s disease
from Magnetic Resonance Image cerebrum pictures gives a posh characteristic like appearances,
boundaries, and similar other things. From the multi sequence of Magnetic Resonance Image

7
cerebrum pictures, the unique characteristics are to be analyzed and then features are to be
extracted from location and division.

In the quick moving worldwide condition, the principal utilization of the clinical picture
7
is to detect, segment and classify the various sorts of malady which has been gained with large
scale acceptance from research analytics [6]. Majority of the researches have been done on the
field of automatic detection, segmentation and classification of disease from Magnetic
Resonance cerebrum images. In clinical field application, error rates occur at all times. The
29
major aim of the present research work is to design and develop an entirely automated detection,
7
segmentation and classification technique for Alzheimer’s malady from MRI cerebrum pictures.
The objectivies of the thesis are summarized as follows:
1. To conduct a review of the significant researches concerned with the analysis of
segmentation and classification of hippocampus for alzheimers disease data.
30
2. To study the features of 3D patches extraction based on hippocampal segmentation for
the classification task is built up with densely connected neural networks.
3. To asses the functional image scans on manual reorientation, visual reading and semi
quantitative techniques are to generate MRI brain images for better understanding.
17
4. To better understand the ability of machine learning techniques to revolutionize the
analysis of hippocampus region in alzheimers disease.
5. To study the idea of merely fusing the existing classification schemes rather than to
develop the new and refined classification techniques.
6. To validate the developed SVM based classification system on real world datasets so that
it can aid researchers in the automatic classification field to test their algorithms.
7. To develop an efficient system for segmentation and classification of hippocampus in
31
alzheimers disease using deep learning neural networks for MRI brain images.
27
8. To evaluate the developed SVM based classification system on real world datasets so that
it can aid researchers in the automatic classification field to test their algorithms.
17
9. To perform an extensive study over some of the state of the art classification techniques
26
available in the literature.
10. To compare the segmentation and classification accuracies with other competitive
techniques or conventional techniques.

8
1.3 METHODOLOGY

The technique processed for hippocampal region in Alzheimer’s disease in brain is for
32
Detection, Classification and Segmentation are presented and discussed in this thesis. The
experimental results were obtained from various MRI cerebrum pictures for the structural
changes in brain to show the sensitivity, specificity and accuracy of the proposed approaches.
The conjecture has also been validated using distinct data sets of natural images corresponding to
30
various locations of hippocampal regions to demonstrate. It is not an inconsequential task. The
extraction of specific information from MRI cerebrum pictures in hippocampus on Alzheimers
Disease images, particularly if the size of the image increases becomes a complex problem
requiring the application technique based on the scrupulous fundamental of both medical and
machine learning image processing concepts. The data is analyzed to achieve the full potential.

In Medical image or machine learning, formulating a prior knowledge is difficult. It is

subsume directly into automatic or semi-automatic techniques of image processing. In image
processing, the simple structure of discrete pixels represents an image. The problems associated
with the Alzheimer’s Disease in hippocampus image analyses are given below:

 Skewed class distributions hinder the classifier learning task, since the
generalizations of the classifiers are compromised.
 Handling class imbalance has been a big challenge in image mining patterns for
the given data.
 In the multiclass scenario, it is necessary to delineate the boundary between
several classes, which are explored by various sampling and feature selection
techniques in existing methods.
7
 The position of hippocampal region ‘S’ has been changed as per the slice of the
patient. Heterogeneity in size and shape of hippocampus.
 Assessment in Low Intensity and High Intensity pixels.
23
 Diagnosis of AD in hippocampus proves to be detrimental and increases the
morbidity and mortality rate.
33
 Specificity and Sensitivity of the existing methods decreases as the volume of
data increases. The existing statistical methods involves more time.

9
 In the biomedical analysis, the image segmentation is predicted to particular tissues or
particular region like hippocampus present in the magnetic resonance images. The processing of
26
large number of image dataset is to be improved with feasible techniques. In the large number of
databases, the accurate automatic segmentation methodologies are leading to new applications
like content based image retrieval. This research is carried out with Feature Extraction, Feature
Selection, Image Segmentation and Image Classification for detection, division and
characterization of hippocampus bizarre in Alzheimer’s disease and their techniques are listed
below.

Feature Extraction:

 Haar Wavelet Transform

 Discrete Wavelet Transform
 Voxel Based Morphometry (VBM)
18
 Diffeomorphic Anatomical Registration Through Exponentiated Lie
algebra (DARTEL)
Feature Selection:

 Principal Component Analysis (PCA)

 Normalized Mutual Information Feature Selection (NMIFS)
 t-test Feature Ranking
 Fisher Criterion

Image Segmentation:

 Deep Convolutional Neural Network

 Region based Active Contour
 Ant Colony Optimization

Image Classification:

 Deep Convolutional Neural Network (CNN)

 Fuzzy Neural Network (FNN)
 Support Vector Machine (SVM)

10
1.4 CONTRIBUTIONS

The fundamental commitments of proposition are given below:

The first segment of proposition work depends on image segmentation and classification
30 34
with various procedures. Hippocampus is one of the first affected brain regions in Alzheimer’s
Disease. The shape and volume of the hippocampus are often measured using structural magnetic
resonance imaging (MRI). Here multi model deep learning framework based on convolutional
neural network (CNN) for joint hippocampal segmentation and Alzheimer’s Disease
classification in structural MRI data is proposed.
The second piece of theory is handled with pre-preparing dependant on Ant Colony
Optimization (ACO); this is done to expel the high recurrent parts. The component extraction is
performed for picture decay utilizing wavelet change. The following calculation is utilized for
the expulsion of film curios. Here the two dimensional framework is performed to look at the
MRI. The picture characterization is finished by utilizing Fuzzy Neural Network (FNN) for
Alzheimer’s, gentle Alzheimer’s and Huntington sickness. The test brings about this area; the
presentation is worked for upgrading the picture and assessed with FNN.
The third piece of postulation works with Virtual Based Morphometry (VBM) and
18
Diffeomorphic Anatomical Registration through Exponentiated Lie Algebra (DARTEL) to
investigate and approach pre-handling. The element extraction which depends on the crude
element vectors on decay groups works with VBM and DARTEL. The element choice is
performed for decrease by using Principal Component Analysis (PCA). In this 10-crease cross
approval is utilized for estimating the exhibitions of classifiers just as t-test highlight positioning
is utilized to quantify the significant highlights and factors. Fisher Criterion or Irving Fishers is
utilized to decide the top discriminative highlights by lessening the dimensionality of point of
view include vectors. The outcome examination is received in estimation with improved picture
with blunder by utilizing Re-replacement mistake and cross approval mistake. The presentation
investigation acted in crude element vector of decay groups in Linear SVM and RBF SVM.
The last piece of proposition examines all the proposed methods. To play out the
quantitative examination or volumetric investigation on the irregular bit is essential the
35
procedure of segmentation. The strange part is separated. In the image classification, the
highlights are removed from the entire picture. The computerized strategies, particularly in

11
 36
administered techniques Artificial Neural Network (ANN), Support Vector Machine (SVM) is
5
exceptionally effective as far as the exhibition measures. In the equivalent ANN or in the harsh
37
procedure, the central point is the benefit for time accessible. With a view to create programmed
strategies with synchronous benefits of high exactness and assembly time is the focal point of the
26
examination work. In order to show the predominant idea of the current methodology a couple of
regular multi-limits and unpleasant set based districts developing and combining procedures are
additionally executed. The quality, measurements, precision and execution time, harsh set and
38
multi-threshold based seeded locale developing and grouping are seen to be best for MRI
pictures depending on the segmentation analysis.

1.5 BACKGROUND OVERVIEW

The real time applications and the outcome are obtained from picture handling methods.
The development of pictorial data for human understanding and therefore the pre-preparing of
the scene information for self-ruling machine recognition are significant of the Image Processing
stems, two essential application zones; to analyze the irregularities by screening the patients’
clinical imaging is utilized. The crude picture is not straight forwardly appropriate for this
reason. Different picture preparing methods like image enhancement, reclamation, enrollment,
and classification are applied to get great outcome. Digital Image Processing (DIP), hugeness of
Magnetic Resonance Imaging (MRI), Image Structure, Image Acquisition, Image Enhancement,
Image Segmentation, and Image Classification foundation sees are itemized below.

1.5.1 Digital Image Processing

The three sorts of computerized picture handling are: They are low-level preparing,
center level handling, and significant level preparing. Crude activities like pre-handling to
diminish commotion, contrast enhancement and picture honing so as to urge to improve pictures
19
with low-level processing. The inclusion of segmentation, description of those objects to scale
back the dimension to form suitable computer-aided processing and classification of individual
objects, in order to get attributes is mid-level handling. Comprehending outfit of perceived items,
39
as in image analysis, and at the far end of the sector, carrying out the cognitive function typically
related to human visualization; so as to form bode well to perform intellectual capacities are
significant level handling.

12
 Optical imaging (cameras, magnifying lens), clinical imaging (CT, MRI, ultrasound, and
diffuse optical magnifying instruments), galactic imaging (telescopes), radar and hyperactive
ghastly imaging (reconnaissance and remote detecting), pc vision (robots, license plate reader,
and tracking human motion), security and digital rights the board (watermarking, and biometrics)
are the uses of Digital Image Processing [7].
With a view to control the advanced pictures by utilization of personal computer
12
framework is the computerized picture preparing. It is used to improve the pictures by
transforming the sign from a picture sensor into advanced picture. It performs different activities
like picture improving, handling of analog and digital signal, image signals, voice signal and so
forth. Digital Image Processing uses software, and freed from cost, it provides clear images,
image enhancement to remember the info through images, it uses widely everywhere in many
fields, it reduces the complexity, it is used to support a far better experience of life. These are all
the characteristics of Digital Image Processing.
The advantages of DIP are;
1. Image recreation (CT, MRI, SPECT, PET)
2. Image reformatting (multi-plane, multi-see recreations)
3. Fast picture stockpiling
4. Fast picture recovery
5. Fast and great picture circulation
6. Controlled review (windowing, zooming)
In Digital image processing, some of the followings are:

1. Image Acquisition
2. Image Enhancement
3. Image Restoration
4. Color Picture Preparing
5. Wavelets and multi- goals preparing
6. Image pressure
7. Morphological preparing
8. Segmentation methodology
9. Representation and Description
10. Object location and Acknowledgement

13
18
In computerized picture handling, the extent of DIP is the way towards securing a picture of a
book, preparing, removing, portioning, singular characters, and depicting characters, appropriate
for picture handling and perceiving those individual characters.

1.5.2 Magnetic Resonance Imaging

The Computerized Tomography (CT), Position Emission Tomography (PET), CT/PET

23
and Magnetic Resonance Imaging (MRI) are widely used clinical strategies. Requirements to
infuse radioactive medication into human body and the medications will permeate into all the
cells, tissues and organs along the blood are used in the Position Emission Tomography swep.
4
They are administered on equivalent plane to form a fused image with the machine is a
combination of Computerized Tomography and Position Emission Tomography scans. Both
Computerized Tomography and Position Emission Tomography assessment have radioactive
dangers. Position Emission Tomography assessments are excessively costly. X-ray is the savvy
contrast with all imaging modalities. It is used for contemplating the life systems and variations
from the norm. The figure 1.4 shows the MRI unit.

Figure 1.4 View of MRI Unit

14
 MRI is assessed into brain imaging, cardiac imaging, liver imaging; lung imaging and so
on that is predicated on the target position. No chemical is injected into the physical body within
the MRI image processing. During this procedure, there is no radiation harm to the physical body
and therefore, the entire procedure is quite safe. MRI has high goals and exact situating of
delicate tissues and is touchy to the qualities of infections. The MRI machines use magnetic
waves rather than radio waves. MRI is the non-invasive tool which provides direct views inside
the body. The strength of magnetic field measuring units is Tesla and Gauss. The clinical
diagnosis, image acquisition or formation of digital image matrix and certain ultrasonic sensing
devices are employed in it, frequently in the form of 2-D or 3-D array has standard 1.5 Tesla.
40
When the light energy falls on the surface of the sensor, it can observe the objects and can
reproduce them as images.
1.5.3 The importance of MRI cerebrum pictures

The MRI pictures brain tissues are separated into two principle types.

The sorts are

41
1. Gray tissue, white tissue and cerebrospinal fluid (CSF) are Normal tissues.
2. Abnormal tissue alludes the most part of the tumor region.
1.5.4 Brain anomalies

Hippocampal formative variations from the norm are found in large number of patients
with inborn distortions. Concentrating on the morphologic variations from the norm of the HF on
coronal, MR pictures may help within the discovery of infections related with mind oddities,
particularly unpretentious cortical issues. The approach of MR imaging has permitted a way for
anatomic structures of the mind and has added to the examination of mental health. Different
formative mind inconsistencies, similar to a few kinds of cortical digenesis, would now be able
to analyzed utilizing MR imaging. For instance, mesial transient sclerosis has been accounted for
the significant reasons for worldly flap epilepsy. However, the basic variations from the norm
stay undetected in the development of MR imaging [8]. Hippocampal variations from the norm
have been watched in mesial fleeting sclerosis as well as in different illness. Formative changes
in hippocampal arrangements have been accounted in agenesis of the corpus callosum,
lissencephaly, and holoprosencephaly. Such examinations have discovered that HFs was little in

15
62 percentages of cases and that abnormal vertical orientation was available in 82 percentages of
patients.

1.5.5 Planes in MRI images

The MRI images have three planes. There are

1. Axial (Top View)

2. Coronal (Back view)
3. Sagittal (Side view)
23
Within the brain, the axial plane is divided into superior and inferior parts. The interior
parts of the brain contain minimum portion of skull and maximum portion of brain. The figure
1.5 shows the images of axial planes of cerebrum.

Figure 1.5 Imaging Planes (Axial, Sagittal and Coronal)

It is easy to research the abnormalities [9]. It is perpendicular to sagittal and coronal

planes. In the brain the sagittal plane or median plane is divided into left and right portions. In
this plane the minimum portion of the brain is viewed. The sagittal plane image is not taken for
this research. Within the brain the coronal plane or frontal plane is divided into front and back
10
portions. The axial and coronal plane view of the brain is of help to seek out whether the patients
suffer from AD or Dementia. The digital brain image appears black and white [10]. In MRI
25
images, the uncontrolled growth of the white cells in a specific region is the dementia. The brain
tissues of hippocampus are represented by different pixel properties. The range of 0 to 255 is the
gray image range. The intensity range starts at 0.

16
 The background of the image is characterized by a black pixel. The range of 200 to 255
refers to the range of white pixels. The mid range of pixels is mentioned as gray pixels. Both
high and low-intensity pixels are used to view the hippocampus tissues in Dementia in Brain. To
35
locate the right portion of the hippocampus ‘S’, the development of methodology is required.
The background is the surrounding area of the brain. The middle portion is cerebral Cerbero
Spinal Fluid (CSF). The figure.1.6 shows MR images scanning flowchart.

Figure.1.6 MR Image scanning flowchart

1.5.6 Image Enhancement
The purpose of picture improvement method is to develop a picture to get more clarity of
the influenced locale. Approaches are widely categorized into two sorts, namely, spatial
strategies and frequency area procedures. The method of imaging within the image plane (pixels)
42
is spatial method. In light of adjusting the change of a picture is called recurrence method. Both
the techniques are used to solve problems using various combinations of techniques. This
upgrade activities are engaged in edge improvement, honing, histogram evening, differentiate
extending, un-sharp veiling, clamor separating and so on [11]. This process does not affect the
12
originality, but it tries to reinforce for future process. It is going to be administered either locally
or all inclusive. The figure.1.7 shows image contrast view of the diagram.

17
 Figure.1.7 Image Enhancement
1.5.7 Hippocampus

Hippocampus, region of the brain that is associated primarily with memory. The shape of
hippocampus is ‘S’ which seems to be Sea horse in shape. It is placed in temporal lobe. It is most
important in the limbic system because this region regulates memory power, emotional thing,
motivational part and learning skill. It has another major function which is known as cognitive
map function. This function is represented to coding, decoding, storing and recalling functions.
This part is performed with different role in brain for flexibility and changes in aim or goal. This
region helps to change or remodify relational memory part. The figure.1.8 shows the image of
hippocampus region in brain .

Figure.1.8 Image of hippocampus region in brain

18
 43
The hippocampus is a sensitive part of the brain. A range of conditions can adversely
affect it, including long-term exposure to high levels of stress. Several diseases and factors are
30
known to impair the hippocampus. The hippocampus is one of the first areas to be affected by
Alzheimer’s disease. An early sign of Alzheimer’s is when a person begins to lose their short-
term memory. They may also find it difficult to follow directions. As the disease progresses, the
hippocampus becomes harder to function in daily life. The people with severe depression and
stress will loses the hippocampus volume. The shrinking of hippocampus in human brain at the
25
stage of 25 years and 75 years of normal and alzheimers disease brain were shown in figure 1.9.

Figure 1.9 Shrinking of Hippocampus

1.5.8 Image Segmentation

Image Segmentation is the main task in image analysis. The first objective is to partition
41
the image into several regions with strong correlation with the object of the important character
contained within the image. While handling this technique the required image tends to extract
information. The segmentation is often complete or partial. The set of disjoint region can
26
correspond directly or partially to the image objects in the segmentation process. The foremost
inauspicious problem present in digital image processing is autonomous segmentation approach.
8
The Segmentation method process is based on two major aspects, namely, Discontinuity and
Similarity. The change which occurs abruptly in gray level or in image property is a
Discontinuity. Similarity is concerned with the homogeneity among the neighboring region. The
19
figure.1.10 shows various approaches of digital image segmentation process. Dividing the image
into semantically homogeneous regions, or objects to be recognized by future processing is aim
of image segmentation [12].

Figure.1.10 Various approaches of Digital Image Segmentation process

1.5.9 Image Classification

30
Image classification is a processing technique with the various classifiers to categorize
16
the image of different application. Image classification rightly refers to a process in computer
vision that can classify an image according to its visual content. With the increasing volatility,
necessity and applications of artificial intelligence, fields like machine learning, and its subsets,
deep learning and neural networks have gained immense momentum. The training needs
software’s and tools like classifiers, which feed huge amount of data, analyze them and extract
useful features. The intent of the classification process is to categorize all pixels in a digital
image into one of several classes. Normally, multi-spectral data are used to perform the
classification and, indeed, the spectral pattern present within the data for each pixel is used as the
numerical basis for categorization.

20
 The objective of image classification is to identify and portray, as a unique gray level (or
color), the features occurring in an image in terms of the object these features actually represent
on the ground. Image classification is perhaps the most important part of digital image analysis.
Classification between objects is a complex task and therefore image classification has been an
important task within the field of computer vision. Image classification refers to the labeling of
images into one of a number of predefined classes. There are potentially n numbers of classes in
which a given image can be classified. Manually checking and classifying images could be a
tedious task especially when they are massive in number and therefore it will be very useful for
automate the entire process using computer vision.

Structure for performing Image Classification

1. Image Pre-processing:

The aim of this process is to improve the image data (features) by suppressing unwanted
distortions and enhancement of some important image features so that the computer vision
models can benefit from this improved data to work on. Steps for image pre-processing
includes Reading image, Resizing image, and Data Augmentation (Gray scaling of image,
Reflection, Gaussian Blurring, Histogram, Equalization, Rotation, and Translation).

2. Detection of an object:

Detection refers to the localization of an object which means the segmentation of the image
and identifying the position of the object of interest.

3. Feature extraction and training:

This is a crucial step wherein statistical or deep learning methods are used to identify the
most interesting patterns of the image, features that might be unique to a particular class and
that will, later on, help the model to differentiate between different classes. This process
where the model learns the features from the dataset is called model training.

4. Classification of the object:

This step categorizes detected objects into predefined classes by using a suitable
classification technique that compares the image patterns with the target patterns.

21
1.6 ORGANIZATION OF THESIS

The present study consists of seven chapters. The first chapter introduces the frame work
10
of medical imaging and provides an understanding of the significant motivation in image
processing besides the research objectives, methodology and contributions of this work. This
chapter provides the background overview of all the key concepts.
The second chapter describes a literature survey on Medical Imaging data, Alzheimer’s,
Hippocampus region and Cerebrum variation from the normal tissue. This section presents a
brief discussion on the techniques for Picture Upgrade, Image Enhancement, Image
Segmentation, Feature Extraction and Selection, and Classifications.
34
The third chapter deals a multi-model deep learning framework based on convolutional
neural network (CNN) for Alzheimer’s Disease classification using structural MRI data.
30
Hippocampus is one of the first affected brain regions in Alzheimer’s Disease. The shape and
34
volume of the hippocampus are often measured using structural magnetic resonance imaging
(MRI). First, a multi-task deep CNN model is constructed for disease classification. Then,
constructing a 3D Densely Connected Convolutional Networks (3D DenseNet) to learn features
extracted based on the hippocampal segmentation results for the classification task. Finally, the
learned features from the multi-task CNN and DenseNet models are combined to classify disease
status. The method is evaluated on the baseline T1-weighted structural MRI data collected from
97 AD, 119 Normal Control (NC) subjects in the Alzheimer's Disease Neuroimaging Initiative
(ADNI) database. The proposed method achieves a dice similarity coefficient of 87.0% for
hippocampal segmentation.

The fourth chapter studies classification of AD from MRI to highlight extraction using
discrete wavelet change. In Alzheimer‘s ailment conclusion, a straightforward and powerful
grouping approach of Magnetic Resonance Image checks is proposed. The robotized multiclass
conclusion of dementia and the grouping dependent on human cerebrum Magnetic Resonance
pictures utilized Fuzzy Neural Network technique. The Fuzzy Neural Network highlights three
classes that are Alzheimer‘s, Mild Alzheimer‘s and Huntington‘s sickness. This strategy gives
better classification precision with other volumetric techniques.

22
 20
The fifth chapter includes positioning based classification of Alzheimer’s Disease in
hippocampus locale from basic attractive reverberation imaging using bolster vector machine.
The t-Test based component positioning is utilized for highlight the feature selection procedure,
44
where the number of features can be determined using Fisher Criterion. Using Irving Fisher
between Alzheimer’s disease and health control groups is calculated for changing number of
ranked features, the ideal size of fisher criterion of selected features is resolved as top
discriminative features. In the grouping of Alzheimer‘s infection, information combination
methods like source and score fusion are utilized to acquire improved organizing. Exactness,
33
explicitness and territory under the curve are exceptionally serious with the condition of
craftsmanship methods utilizing magnetic reverberation image information.
The sixth chapter demonstrates the materialness to all the methodologies applied to
various MRI groupings of cerebrum pictures. The various strategies are introduced utilizing
comparison of segmentation and classification exactnesses.
The final chapter sums up the entire conclusion and offers a framework of the areas for
future work.

23
 Chapter - 2
2.0 LITERATURE SURVEY

Mind life system in clinical imaging is a non-intrusive innovation [13]. Clinical imaging
has new horizons in analyzing and subtracts signifying the brain structure and its functions. The
way toward utilizing MRI pictures in getting to mind wounds and investigating cerebrum life
systems has gained a gigantic ground. Utilizing the scanner on tissue related information namely
gray level or intensity the primary features of MRI cerebrum pictures are interpreted. The aim of
image processing is to remove essentialness highlight from picture information, from which the
machine can get distinct interpretative or understandable information. A human brain is an organ
which is found in central portion of human systema nervosum. A brain has three major parts
which is cerebrum, brainstem and cerebellum parts. Cerebrum is the largest part which is
assessed to two cerebral hemispheres. Each hemisphere is assessed into four lobes which are
indicated as frontal flap, parietal projection, fleeting projection and occipital flap [14]. In human
brain, a cerebrum is integrated by brainstem to medulla spinalis which is midbrain and medulla
oblongata.

Promotion (Alzheimer’s ailment) is a mind dynamic issue which causes marginally to

lesson thinking pace and memory power. The Alzheimer ailment comes from Dr. Alois
Alzheimer diagnosis in 1906. He intimated that any different changes occur in brain tissue of a
people. Hippocampus is the main key considers being present in the advancement of Alzheimer’s
sickness. Within the Alzheimer’s malady, hippocampus region volume and shape helps to work
out the first detection analysis. An early identification of Alzheimer’s infection, intra-neuronal
45
filamentous are to be deposited and also neuro-fibrillary tangles are to be stored within the
28
neurons [15]. An extension of cubiculum of the parahippocampal gyrus is the gray matter in the
hippocampus. In hippocampus, the two C-shaped structures are interlocked. These two
interlocking C- formed structures are the cornus of Ammonis and the dentate gyrus. One of the
side effects of Dementia is the decay of hippocampus. The MRI allows precise assessment of
structure the high resolution of T1 weighted. In Alzheimer’s disease, compares to healthy tissue
the hippocampus volume is little.

24
 Anjan Gudigar [2019] et. al. proposed an automatic categorization of
abnormalities in brain with the assistance of decomposition techniques with MRI
images. The paper has a creative diagnosis technique for the purpose of automated
processing of cerebrum pictures. There are two modes presented such as bi-dimensional
empirical method of decomposition and variation method of decomposition process. An
ideal classification technique is utilized like bolster vector machine which gives an
accurate result and also peak quality.
P. Muthu Krishnammal [2019] et. al. proposed brain abnormalities with the
assistance of convolution neural network mechanism for the classification and detection
41
process. Convolutional neural network gives the result for classification process as in a
46
curve let domain. This system is employed to supply the best prediction of excellent
sparse and better extraction of features. The performance analysis helps to offer the
knowledge about best accuracy than other traditional mechanisms.

Himar Fabelo [2019] et. al. proposed a brain cancer detection mechanism
through database separation process. During this paper, hyper spectral acquisition
47
system is to be presented which is capable of predicted data from visual and near
infrared within the range of 400 to 1000 nano meter. The paper has four different
classes such as, like normal tissues, background elements, tumor tissues and also blood
vessel. The paper sums up the intra operative acquisition system, human brain database
system, and gold standard database system and repeatability analysis of the acquisition
system.
Goyal et. al. proposed a completely unique technique of detecting brain abnormalities as
automated manner for tumor segmentation analysis in MRI images. In this paper, they provided
the information about automated and clinically-tested mechanism for abnormalities detection
process and segmentation of tumor mechanism using MRI image sequences. This method results
38
in Radiologist’s approach to the brain abnormalities mechanism with the assistance multiple MRI
images instead in training phases [15].
Lashkari et. al. proposed a lesion detection technique automatically in brain, which helps
27
to scale the time of diagnosis and high precision. The main objective of this paper is to classify
the brain tissues to two major classes namely, normal one and abnormal one. In MRI images,
these two classifications are to be applied. The paper elaborates the feature selection mechanism

25
which helps to decrease the space of features and to make this possible a neural network
48
classification algorithm has been applied. Reddy et. al. proposed the enhancement of fuzzy
49
clustering meaning methodology. This paper explores the spatial method to FCM methodology
in the initiated time. This spatial information is to be encoded into mutual influence of
neighboring positions. This paper helps to detect the brain’s abnormalities using MRI images.
The results show the analysis of Fuzzy Cluster Means and k-Means Clustering algorithm.

2.1. RESEARCH ON IMAGE ENHANCEMENT

Sharumathi, K. [2016] et. al. proposed an improvement technique for submerged acoustic
picture model. This paper is to be investigated to change the techniques like wavelet model,
curve let and contour let model [16]. This method is best adaptable for eliminating speckle
noise. This paper gives the knowledge about various sorts of noise removal filtering technique.
It includes data like side output sonar model, picture arrangement procedure, qualities of acoustic
picture, bend let change area, wavelet change, K – L change, contourlet change model, versatile
dot channels, nonlinear sifting techniques, mean channel, vector middle channel, adjusted spatial
middle channel, homomorphic separating, anisotropic sifting, wavelet denoising by normal
channel and execution examination.

Guy Gilboa [2004] et. al. proposed an enhancement and denoising method with the help
of complex diffusion processes [17]. In this paper, straight and nonlinear spaces are produced
with the assistance of genuine esteemed dispersion arrangement procedure. It results in the
imaginary parts, smoothening in second derivative functional analysis [103]. This paper has the
following information like linear complex diffusion, approximate solution for little theta,
analysis of real kernel, analysis of imaginary kernel, normalization process, ramp preserving
denoising process, shock filters and performance analysis.
Randeep Kaur [2016] et. al. proposed an enhancement methodology using different
contrast mechanism in medical image [18]. This method helps to offer better result for top
contrast. Image upgrading assists with expanding the best nature of a picture for human survey.
This system has the following techniques such as, neighborhood operation, average filter
bilateral ratinex and sigmoid function. The present investigation gives a superior outcome, the
examination of various components like mean square mistake, top sign to clamor proportion,

26
root mean square blunder and standardization coefficient [106]. The further development of this
paper is to reinforce this paper into neighborhood operation in clinical x-beams.
Manasvi Kalra [2015] et. al. proposed a picture upgrade technique which is performed in
cancer detection with the assistance of versatile application [19]. The principle outcome of this
paper is to recognize the fundamental harmful impact of carcinoma. In this paper, the picture
assists with improving or abatement differentiate stage and brightness levels in a picture [105].
In this paper, the versatile application assists to attach with EHIS to persistent. This paper has
the following functions such as, spatial and recurrence area, versatile unsharp covering capacity,
versatile histogram leveling, versatile neighborhood differentiate improvement, nearby and
worldwide difference extending examination, and execution investigation utilizing portable
application. The further development of this paper is to enhancing this system into hybrid model
with two or more mechanism.

2.2. RESEARCH ON IMAGE SEGEMENTATION

Liang Shen [1997] et. al. proposed a lossless compression coding process with the
assistance of image segmentation for top resolution process [20]. This paper was developed with
the innovative technique like segmentation based lossless image coding process which provides a
far better result of efficient region growing mechanism. This paper has the following
information like segmentation based compression process, region growing algorithm,
segmentation mechanism depend lossless image, and execution examination. The further
improvement of this paper is to grow high productive strategies with the assistance of region
growing technique.

Yubing Li [2018] et. al. proposed a picture division process supported grab-cut process in
a picture region. In the image segmentation process, grab-cut mechanism is most predicted one
which helps to supply the knowledge about the surface and limit [21]. This system has got to
propose with the blend of snatch-cut and diagram based division. This paper has following
information like grab cut calculation, diagram based picture division, snatch-cut improvement
16
calculation, and assessment process. The further development of this process is to enhance
image color and background because existing method have high complex color and background.
Annegreet van Opbroek [2013] et. al. proposed a picture division process with the
assistance of transfer learning mechanism using imaging protocols [22]. The transfer learning
27
mechanism helps to research differentiation of coaching information data and target information
data. This method gives to peak the performance for segmentation process of various scanning
techniques. This performance analysis provides better result for minimum classification errors
compared with other supervised learning mechanisms.
Robert Manzke [2010] et. al. proposed a picture division procedural process with the help
of anatomic intra-procedural surface generation mechanism [23]. In this paper, this mechanism
helps to segment the LAPV present in AF ablation process. This automatic segmentation
algorithm provides a far better result at high speed and accurate manner. This paper has the
following procedure like intra-procedural imaging and LAPV surface age component, picture
securing process with the intra-procedural rotational imaging, picture remaking system, picture
enlistment procedure, and programmed LAPV age instrument and execution examination.
H. M. Tarek Ullah [2018] et. al. proposed bill board detection process assistance DCNN
(Deep Convolution Neural Network) data from MRI images in three dimensional. This paper
proposed two approaches namely; trick one, less costly one and more reliable one. It offers an
accurate predicted result with the assistance of deep learning mechanism which helps to supply
true bleeding edge. It gives a superior presentation examination than the customary
convolutional neural system which depends on age stanzas exactness and age section misfortune
factor. The further development of this paper is to reinforce this idea for final layer of neural
network.
G. Song [2019] et. al. proposed a non-invasive system purpose of Gliomas based
automatic detection with the assistance hybrid features and PSO – KSVM [24]. In this
paper, image standardization process is to be operated which has normalized size and
removal of background. These processes are to be applied in standard pictures. The
balanced procedure is applied on a picture which assists with improving the low
differentiation pictures in mind and skull expulsion process. Dark level co-event lattice
and pyramid histogram is applied to highlight extraction process. The kernel support
vector machine and molecule swarm advancement technique are applied for grouping
50
process. This method gives a better result than the other current systems.

28
2.3. RESEARCH ON FEATURE EXTRACTION

44
Akshaya Pisal [2017] et. al. proposed a feature extraction process with the assistance of
hierarchical MAX method in face [25]. In pre-processing stage, HSV color space model is to be
used and also Gaussian filter helps to offer accurate image. This paper gives hierarchical max
model which provides an entirely accurate extraction of biologically inspired features in face.
This paper has following information like general age estimation system, feature extraction
process analysis and performance analysis.

Divi Galih Prasetyo Putri [2016] et. al. proposed a feature extraction process with the
help of infrequent extraction method using software [26]. This paper helps to seek out the
opinion data factors using complement collocation process. To scale back the irrelevance
features with the help of feature pruning technique. The paper unfolds following functions like
frequent feature extraction function, POS tagging scheme, mapping process, infrequent feature
extraction process, feature pruning methodology and also performance analysis.
Hongmoing Li [2019] et. al. proposed an Alzheimer’s sickness on early
discovery patients relaying upon the counting on MRI images in hippocampus region
5
and the boundary lines are measures with the help of (DRNN) Deep Recurrent Neural
Networks [27]. In this paper, the classification methods are demonstrated with the
assistance of pattern classifiers which help to succeed in high classification
performance. The information is learning with the assistance of profound learning
model which is counting on intermittent neural system process. Recreation results give
the knowledge of high precision in prognostic model, Alzheimer’s ailment dementia and
also early prediction of Alzheimer’s disease progression.
M. Amin – Naji [2019] et. al. proposed a detection of AD form MRI in structural
images with the help of Siamese Convolutional Neural Network (SCNN) [28]. The
16
Siamese convolutional neural network is to be performed cast off on major three
branches of ResNet-34 which helps to differentiate the Alzheimer disease from normal
control strategies. This method was processed on ODSIS data set and it gives the result
as peak accuracy than other traditional method. This paper has the following
information like Siamese Convolutional Neural Network (SCNN), specification, dataset

29
 24
analysis, diagnosis of Alzheimer’s disease with novel network, evaluation novel
proposed methodology and performance analysis.
B. S. D. Abajo [2019] et. al. proposed a calculation of probability of suffering
Alzheimer’s disease in mobile Alzhe alert [29]. The calculation result has the small
print about the risk of Alzheimer’s disease with the assistance of questionnaire. This
paper covers the following information such as, flujo methodology, risk correlational
analysis, probability analysis and performance analysis. J. Fritsch [2019] et. al.
proposed an automatic diagnosis of Alzheimer’s disease with the assistance of neural
network analysis language models [30]. This paper has statistical approach like N –
grams language models which is developed and enhanced with the assistance of model
like neural network language models. This paper highlights the following information
like neural network language models, language model evaluation, and performance
analysis.

2.4. RESEARCH ON IMAGE CLASSIFICATION

L. Xu [2019] et. al. proposed meager component learning with label information
for Alzheimer’s malady classification depending on reverberation picture [31]. In this
5
paper, the features are to be adjusted with the assistance of the label information. The
support vector machine classification method gives the result in simultaneous manner
which helps to retain the compactness of intra class samples in MRI images. This paper
talks about the following information like material analysis, image pre-preparing and
have extraction process, sparse regression model analysis, manifold regularized sparse
feature learning analysis, optimization algorithm analysis, classification analysis and
performance analysis.
Begüm Demir [2009] et. al. proposed a hierarchical method for support vector machine
classification process in hyper-spectral images. This method gives high accuracy and also less
computational load. In this support vector machine technique, the classification process is
applying at lower level which gives the result as high resolution assessed images. This paper has
36
the following procedure like support vector machine based classification process, hierarchical
hyper spectral image classification process, and performance analysis. The further development
of this paper is to research high accuracy mechanism in spatial correlation region.

30
 Biao Hou [2015] et. al. proposed a picture arrangement with the assistance of profound
conviction organizes in polar metric SAR pictures [32]. The primary stage is to convert normal
matrix data into nine dimensional matrix data. In second stage, each and each nine dimensional
47
data is to be applied for filtration process. In the final stage, the features elements are to be
combined to three layers DBN. The performance analysis gives a result as high efficient and
effective classification process than other traditional mechanisms.
S.Manthira Moorthi [2011] et. al. proposed a satellite image classification mechanism
with the assistance of completely unique approach like kernel based learning approach [33]. The
spectral image data set is to classify with the assistance of bolster vector machine process. This
paper has the following information like principle of support vector machine algorithm, image
format analysis, acquisition analysis, resolution analysis, procedural development analysis,
classification analysis and performance analysis. The further development of this paper is to
reinforce best accuracy in multispectral image classification than proposed technique.
The literature review deals that various researchers have high powerful efforts to enhance
the knowledge on Alzheimer’s disease segmentation, detection and classification. Here the
image processing is planned with the proper MRI images for the segmentation with the
procedure and to realize the proficient division and characterization of MRI cerebrum datasets
for Alzheimer’s illness discovery. The objective is to enhance the precision of division strategy
or to form a completely unique approach. The methods like limit based, edge based, region
based, and histogram-based segmentation are utilized in numerous works. The supervised and
unsupervised based classifiers are discussed to enhance the precision by minimizing the
uncertainties in the image analysis.

31
 Chapter - 3
24
Multi-Model Deep Convolutional Neural Network based
Segmentation and Classification of Hippocampus Region on Alzheimer’s
Disease from MRI

3.1 INTRODUCTION
21
Alzheimer’s disease (AD) is a progressive and irreversible brain degenerative disorder
characterized by memory loss and cognitive impairment. At present, there are around 90 million
people diagnosed with AD, and it is estimated that the number of AD patients will reach 300
million by 2050. To date, no effective drug treatments are available to cure AD, while the
existing AD medicines can only ease symptoms or slow down its progression. Thus, the
34
detection of AD at its early or prodromal stage is important for the prevention and intervention of
21 30
its progression. For example, mild cognitive impairment (MCI) is a prodromal stage of AD. In
21
the past few decades, neuro imaging technologies have been widely used to discover the relevant
10
biomarkers in the human brain for AD and MCI diagnosis. Magnetic resonance imaging (MRI)
is a non-invasive imaging technology that can produce detailed 3D anatomical images of internal
body structures such as the brain, and has been widely used to help us understand anatomical and
functional brain changes related to AD. In particular, structural MRI scans provide detailed
information about the anatomical structures of the brain, which can help detect and measure
brain atrophy patterns in AD [35].
36
In recent years, many machine learning models have been employed to analyze the MRI
data for identifying biomarkers and deciphering the disease etiology. A multi-task method was
51
proposed to select the most discriminative features from 93 ROIs for multimodal classification of
AD/MCI. An approach was proposed to combine the marginal fisher analysis with norm-based
multi-kernel learning to achieve the sparsity of ROIs, which could simultaneously select a subset
53
of relevant brain regions and learn a dimensionality transformation. A deep network with stacked
auto encoder was employed on the ROI features to extract the latent high-level features for
improving the performance of disease classification using neuro imaging data.

32
 Recently, deep learning networks, including convolutional neural networks (CNNs), have
21
been widely used in image classification and computer vision. The deep 3D CNNs were used to
5
extract the features of 3D medical images for classification. A multi-task deep learning (MDL)
24
method was proposed for joint hippocampal segmentation and clinical score regression using
2
MRI scans. Given the very high dimensionality of the brain MRI data, it requires huge
computational resources and a large dataset to train a deeper CNN with robustness. Since the
44
MRI datasets used for AD diagnosis are typically very small compared with the datasets used in
41
computer vision, it remains a major challenge to train a deeper CNN model with a large number
30
of parameters to be learned. Among all brain ROIs, the hippocampus is one of the first affected
regions in AD, and it is an important anatomical region in the AD etiology.
21
Alzheimer’s disease (AD) is a progressive and irreversible brain degenerative disorder.
34
Mild cognitive impairment (MCI) is a clinical precursor of AD. Although some treatments can
delay its progression, no effective cures are available for AD. The early-stage diagnosis
of AD is vital for the prevention and intervention of the disease progression. Hippocampus
is one of the first affected brain regions in AD. To help AD diagnosis, the shape and volume of
the hippocampus are often measured using structural magnetic resonance imaging (MRI).
However, these features encode limited information and may suffer from segmentation errors.
Additionally, the extraction of these features is independent of the classification model, which
could result in sub-optimal performance. Multi-model deep learning framework based on
convolutional neural network (CNN) for joint automatic hippocampal segmentation and AD
classification using structural MRI data is used. Firstly, a multi-task deep CNN model is
constructed for hippocampal segmentation and disease classification. Then, constructing 3D
Densely Connected Convolutional Networks (3D DenseNet) to enhance the features of the 3D
patches extracted based on the hippocampal segmentation results for the
classification task.

Finally, the multi-task CNN and DenseNet models are combined to classify disease
status. This method is evaluated on the baseline T1-weighted structural MRI data collected
from 97 AD, 233 MCI, 119 Normal Control (NC) subjects in the Alzheimer's Disease
Neuroimaging Initiative (ADNI) database. The proposed method achieves a dice similarity
coefficient of 87.0% for hippocampal segmentation. In addition, the proposed method

33
achieves an accuracy of 88.9% and an AUC (area under the ROC curve) of 92.5% for
classifying AD vs. NC subjects, and an accuracy of 76.2% and an AUC of 77.5% for
classifying MCI vs. NC subjects.

3.2 REVIEW OF EARLIER WORKS

28
There are several problems with the existing hippocampal analysis methods. First, both
hippocampal volumetric and shape analyses depend on accurate segmentation of the
hippocampus. It is still a challenging task to segment the hippocampus due to its irregular shape
and blurred boundary in MRI scans. Second, the hand-crafted features of hippocampal volumes
and shapes may not be optimal for the subsequent analysis, which may affect classification
24
performance in disease diagnosis. Third, using the hippocampus alone may not be sufficient for
51
discriminating MCI from NC subjects. Other regions adjacent to the hippocampus such as the
para hippocampus and amygdala are also affected in early stage of AD. Finally, the visual and
18
texture features of MRI scans derived from the hippocampal region can be of great help for AD
diagnosis.
27
To address the above problems in the computer-aided AD diagnosis, a new deep learning
framework with a combination of multiple deep CNN models for simultaneous hippocampal
segmentation and disease classification is using MRI data. In this framework, first construct a
34
multi-task deep CNN model for jointly hippocampal segmentation and disease classification.
30
Then, 3D Densely Connected Convolutional Network (3D DenseNet) model is developed with
34
the inputs of 3D patches extracted based on the hippocampal segmentation results to rich
41 34
features. Finally, the learned features by the multi-task deep CNN model and the DenseNet
41
model are combined with a fully connected layer to yield a final classification of disease status.
The proposed multi-model framework will be shown to outperform each single model method.
30
Our method is evaluated for both hippocampal segmentation and disease classification, where we
34 54
use the baseline T1-weighted structural MRI data from the ADNI database including 97 AD, 233
MCI, and 119 NC subjects. In addition, we also test our method on an additional dataset of 135
25
subjects from the ADNI MRI cohort with the EADC-ADNI Harmonized Protocol (HarP) [36]
for manual hippocampal segmentation.

34
3.3 THEORETICAL BACKGROUND:

3.3.1 Deep Learning

Deep learning is a subfield of machine learning. Deep learning has more hidden layers,
which uses multiple cascaded nonlinear processing units for feature extraction and
transformation. Compared with Machine Learning, deep learning has the following advantages.
First, deep learning has a very strong learning ability. From the results of learning, deep learning
performs better. Second, deep learning has wide coverage and good adaptability. Deep learning
neural networks have many layers and wide widths; can theoretically be mapped to arbitrary
functions, so they can solve very complex problems. Third, deep learning is data-driven and has
great prospects. Deep learning is highly data-dependent, the relevant experiments show that the
6
larger the amount of data, the better it performs. Some tasks such as image recognition, face
recognition, and natural language processing have even surpassed human performance. Fourth,
17
deep learning has good portability. Due to the outstanding performance of deep learning,
currently many frameworks can be used to deploy deep learning, such as MATLAT, TensorFlow
and Pytorch, etc., and these frameworks are compatible with many platforms, such as Windows,
Linux and Mac OS.

3.3.2 Convolutional Neural Network

A convolutional neural network is a network consisting of at least one convolutional

41
layer and a fully connected layer for network output, including association weights and pooling
layers. Convolutional Neural Network can effectively reduce the dimensionality of a large
37
amount of data information in an image to obtain a small amount of data while effectively
retaining the feature of the image, compared with other deep learning models, convolutional
17
neural networks can have better results in image recognition and classification. Convolutional
16
neural networks are comprised of two very simple elements, namely convolutional layers and
pooling layers. Although simple, there are near-infinite ways to arrange these layers for a given
computer vision problem. The elements of a convolutional neural network, such as convolutional
and pooling layers, are relatively straightforward to understand. The challenging part of using
convolutional neural networks in practice is how to design model architectures that best use these
simple elements. The reason why convolutional neural network is hugely popular is because of
their architecture, the best thing is there is no need of feature extraction. The system learns to do
35
feature extraction and the core concept is, it uses convolution of image and filters to generate
invariant features which are passed on to the next layer. The features in next layer are convoluted
with different filters to generate more invariant and abstract features and the process continues
till it gets final feature/output which is invariant to occlusions. The most commonly used
architectures of convolutional neural network are LeNet, AlexNet, ZFNet, GoogLeNet,
55
VGGNet, and ResNet. The figure 3.1 states the deep CNN in healthy control brain and
alzheimers disease in hippocampus on human brain mri images.

Figure 3.1 Deep CNN in MRI view in healthy Brain (Left) and an Alzheimers brain
(Right). (Yellow – Cortex, Blue – Ventricle, Purple – Hippocampus)

3.3.3 Convolutional Layer

The first layer of CNN after the input layer is usually a convolution layer. The image of
52
the input layer can be regarded as an array of pixel values of W × H × C, where W and H are the
width and height of the image, C is the channel of the image. For an RGB image, C =3 usually.
The convolution layer will use an N × N matrix as a filter (usually N is 3, 5 or 7) to perform the
5
convolution operation from the upper left corner of the image, that is, the values in the matrix are
multiplied by the corresponding values in the image covered by the filter, then sum up all the
products to form the convolution value at the location of the filter. After that, the filter will
perform a convolution operation from left to right and from top to bottom according to the stride
19
(the stride is usually 1), finally obtain a new array as the input of the next layer. Among them,
the size of convolution kernel and the number of filters is manually set. During the initialization

36
process, the weight parameters will be randomly generated, they will be continuously optimized
in the subsequent training process to achieve the best classification performance. The
convolution process is to use these weights to continuously multiply the RGB values of these
18
pictures so as to extract visual data information. The convolutional layer not only extracts the
picture information, but also achieves the effect of dimensionality reduction.

3.3.4 Activation Layer

An activation layer is usually applied after each convolutional layer. It is used to

introduce nonlinear features to a system that has just performed a linear computation 15
operation in a convolutional layer. Previously, nonlinear equations such as hyperbolic tangent
and sigmoidal functions were used, but scientists have found that Rectified Linear Units (ReLU)
layer performs much better and can improve computing efficiency because neural networks can
reduce greatly the training time without significant change of accuracy. It can also help alleviate
the problem of gradient disappearance, because the gradient disappears exponentially in the
layer, the training speed of the network is very slow. The ReLU layer applies the function
f (x) = max (0, x) to the inputs. That is, all negative activations become zero. This layer could
28
increase the nonlinear characteristics of the model and the entire neural network, will not affect
the receptive field of the convolutional layer.

3.3.5 Dropout Layer

8
To control the overfitting, the dropout layer is usually applied. It discards a random
activation parameter set, that is, sets these activation parameters to 0 in the forward pass so as to
56
ensure that the neural network will not affect the training samples overmatching, which will
alleviate overfitting issues. The drop out layer is shown in the figure 3.2 respectively.

Fig 3.2 Drop out Layer

37
3.3.6 Fully Connected Layer
37
After feature extraction from the previous layer, the network will use a Fully Connected
Layer to map higher-level activation mappings to the classification of the output layer and
1 54
generate an n-dimensional vector, where n is the number of output layer classifications. This n-
dimensional vector represents the probability of the detected image in N classifications.

3.3.7 Transposed Convolution Layer

6
The convolution operation can be regarded as a down sampling process, so the transposed
convolution on the sampling can be regarded as the inverse process of the convolution operation,
which is generally used for upsampling. It is becoming more common in recently proposed
convolutional neural networks, especially in Generative Adversarial Network (GAN), a
transposed convolutional layer appears in the up-sampling part of the generator network to
restore the reduced dimensionality. The transposed convolution is derived as follows.

Suppose the size of the convolution operation is 4×4 and the element matrix is

[ ]
x1 x2 x3 x4
x5 x6 x7 x8
푋= x x x x (3.1)
9 10 11 12
x13 x14 x15 x16

The size of the convolution kernel is 3×3, the element matrix is

[ ]
푤0,0 푤0,1 푤0,2
푊 = 푤1,0 푤1,1 푤1,2 (3.2)
푤2,0 푤2,1 푤2,2

where strides=1, padding=0, that is i=4, k=3, s=1, p=0. According to the convolution operations,
the size of the output image is 2×2.

푖 + 2푝 ― 푘
푂= 푠
+1 (3.3)

Expand the input element matrix into a column vector X,

푋 = [푥1 푥2 푥3 푥4 푥5 푥6 푥7 푥8 푥9 푥10 푥11 푥12 푥13 푥14 푥15 푥16 ] (3.4)

38
Expand the element matrix of the output image into a column vector Y,

푌 = [ 푦1 푦2 푦3 푦4 ]푇 (3.5)
17
For the input element matrix X and the output element matrix Y, the matrix operation is used to
describe this process:

Y = CX (3.6)
By derivation, the sparse matrix C can be gotten:

[ ]
푤0,0 푤0,1 푤0,2 0 푤1,0 푤1,1 푤1,2 0 푤2,0 푤2,1 푤2,2 0 0 0 0 0
0 푤0,0 푤0,1 푤0,2 0 푤1,0 푤1,1 푤1,2 0 푤2,0 푤2,1 푤2,2 0 0 0 0
퐶= 0000푤 푤 푤 0푤 푤 푤 0푤 푤 푤 0 (3.7)
0,0 0,1 0,2 1,0 1,1 1,2 2,0 2,1 2,2
0 0 0 0 0 푤0,0 푤0,1 푤0,2 0 푤1,0 푤1,1 푤1,2 0 푤2,0 푤2,1 푤2,2

The operation of transposed convolution is to inverse this matrix calculation process, that
is, to obtain X through C and Y. According to the size of each matrix, the calculation process can
be easily obtained, which is the operation of transposed convolution:

Y = CTY (3.8)

However, if it is substituted into the numerical calculation, the operation of transposed

31
convolution only restores the size of the matrix X, but it cannot restore the value of each element
of X.

3.3.8 Hippocampus
43
The hippocampus is an important part of the brain of humans and vertebrates. It is a part
52
of the limbic system of the brain which is located below the cerebral cortex and plays a role in
learning, memory, stress regulation, and spatial navigation. The hippocampus is named because
it looks like a sea horse, which appears in pairs in all animals with the hippocampus and locates
40
in the left and right brain hemisphere [38].
57
In Alzheimer's disease, the hippocampus is the first area to be damaged. With the
development of the disease, the hippocampus gradually shrinks, and its clinical manifestations
25
are memory loss and loss of direction perception. The hippocampus is the only tissue in the brain

39
that can generate neurons. In other words, it has the ability to repair itself, and can make new
31
neurons in the hippocampus. The rate of neuron generation in the hippocampus decreases
40
slightly with age. the figure 3.3 of a and b shows the right and left hippocampus in alzheimers
disease on human brain.

(a) (b)
Fig 3.3 a) Right Hippocampus and b) Left Hippocampus in Alzheimers Disease in Human Brain

3.3.9 ResNet

The network structure continues to deepen the gradient will gradually disappear during
the back propagation process, resulting in the inability to effectively adjust the weights of the
previous network layer, because the accuracy rate is gradually saturated, and decrease rapidly
which is called degradation problem, but this is not caused by overfitting, because the accuracy
is higher based on training set than the validation set. ResNet can simplify network training and
optimize easily which has also achieved higher accuracy while deepening the network structure.

ResNet creatively employs a shortcut connection to provide identity mapping. When back
37
propagating during training, it passes the gradient of the next layer to the previous layer, thereby
41
solving the problem of gradient vanishing in deep networks. The principle of the residual block
28
is shown in Figure 3.4.

40
 Fig 3.4 Residual block
32
It is assumed that the input is x, the output of the convolution layer is F(x), it is added to
x as the mapping input, and the resultant output H (x) = F (x) + x is passed to the next layer. This
is much easier than matching an identity map through a bunch of nonlinear layers, it will not add
extra parameters and calculations to the network. At the same time, it can greatly increase the
training speed and improve the training effect of the model when the number of layers deepens,
this simple structure can well solve the problem of gradient vanishing [54]. ResNet has two types
58
of residual blocks. The first type is suitable for training shallow networks, if the network is
deeper (more than 50 layers), the second type (bottleneck) is recommended. Moreover, the two
types have similar time complexity.

3.3.10 DenseNet

59
The emergence of ResNet allows the network structure to be developed in a deeper
36
direction. The problem is the limitation of memory or video memory, as the network deepens,
the parameters that need to be processed are also increasing. DenseNet can effectively control
the number of parameters while deepening the network[56]. Unlike ResNet, DenseNet is a
special type of convolutional neural network with dense connections. In such a network, there is
41
a direct connection between any two network layers, that is, the input of each layer in the

41
network is the output of all previous layers. At the same time, the feature maps extracted from
this layer will also be passed as input directly to all subsequent layers.

If the transformation function of layer i is H_i, corresponding to one or two batch

normalization, ReLU, and Convolution operations, the output is x_i, then we can use a very
simple equation to describe each layer of DenseNet.

x_i = h_i ([x_0, x_1, ……, x_ {i-1}]) (3.9)

DenseNet differs from other networks in two main ways. First, it allows each layer in the
6
network to be directly connected to its previous layers so as to achieve the reuse of features.
27
Second, it designs each layer of the network to be particularly narrow, that is, only very few
feature maps are learned (the most extreme case is that each layer only learns one feature map) to
35
achieve the purpose of reducing redundancy. The first point is the premise of the second point,
without dense connections, it is impossible to design the network too narrow, otherwise under
fitting will occur during training.
30
DenseNet has the following advantages. First, the parameters of DenseNet are greatly
reduced. Experiments have shown that the same accuracy rate is achieved based on the ImageNet
data set, DenseNet only requires less than half of the parameters. This is of great significance to
the industry because small models can significantly reduce storage overhead and save
bandwidth. Second, DenseNet requires less computation. When the accuracy equivalent to
ResNet is achieved, DenseNet requires only half of the calculation amount of ResNet.
Nowadays, the requirement for computational efficiency in practical applications of deep
learning is a very strong, efficient model will be much meaningful to the industry.

Third, DenseNet has very good anti-overfit performance, which is especially suitable for
54
applications with relatively scarce training data. It is well known that insufficient training data
sets can easily lead to overfitting. Experiments show that DenseNet reduces the error rate of the
best previous results from 28.20% to 19.64% based on the CIFAR data set without data
32
enhancement. This achievement can be achieved because the features extracted from each layer
41
of the convolutional neural network can be equivalent to a nonlinear transformation of the input
data [39]. As the depth increases, the complexity of the transformation also gradually increases
(more complex non-linear functions).

42
 Compared to general neural networks that directly depend on the features of the last layer
(the highest complexity) of the network, DenseNet can comprehensively utilize features with low
complexity in the shallow layer; it is easier to obtain a smoother model with better generalization
performance decision function. The architecture of DenseNet is shown in figure 3.5.

Fig 3.5 Architecture of DenseNet

43
3.3.11 Batch Normalization
5
Batch normalization is a layer that allows every layer of the network to do learning more
independently. It is used to normalize the output of the previous layers. The activations scale the
input layer in normalization. Using batch normalization learning becomes efficient also it can be
used as regularization to avoid over fitting of the model. The layer is added to the sequential
model to standardize the input or the outputs. It can be used at several points in between the
37
layers of the model. Batch normalization is a technique for training very deep neural networks
that standardizes the inputs to a layer for each mini-batch. This has the effect of stabilizing the
learning process and dramatically reducing the number of training epochs required to train deep
networks. Deep neural networks are challenging to train, not least because the input from prior
layers can change after weight updates. Batch normalization is a technique to standardize the
inputs to a network, applied to ether the activations of a prior layer or inputs directly. Batch
normalization accelerates training, in some cases by halving the epochs or better, and provides
some regularization, reducing generalization error. The figure 3.6 shows the architecture of
Batch Normalization and also shows the Before Normalization and After Normalization using
Deep neural networks in batch normalization.

Convolution

ReLu

Batch Normalization

Convolution

Dense Net

Figure 3.6 Architecture of Batch Normalization

44
 (a) Before Normalization (b) After Normalization
Fig 3.6 Batch Normalization – a). Before Normalization b). After Normalization

3.4 METHODOLOGY FOR SEGMENTATION AND CLASSIFICATION OF

HIPPOCAMPUS REGION IN ALZHEIMER’S DISEASE FROM MRI USING DEEP
CONVOLUTIONAL NEURAL NETWORK.

In this process, the proposed deep learning framework consists of two deep learning
models. One model is multi-task deep CNNs for jointly learning hippocampus segmentation and
disease classification, which generates a binary segmentation mask of the hippocampus and for
disease classification. The multi-task model are not sufficient for accurate disease classification.
A 3D patch covering the hippocampus is extracted based on the centroid of the segmentation
mask and input into a 3D DenseNet model that learns more relevant features for disease
41
classification. Finally, a fully connected layer and a softmax layer are appended to combine the
34
learned features from these models for final disease classification. Here, two classes are
60
considered in the classification task. The input of this framework is a large image patch covering
the hippocampus.
The brain images are obtained from ADNI data set of MRI images and from open source
which are easily available in online. The real image of dementia with hippocampus is collected
from Govt. Rajaji Hospital, Madurai and Aahana Hospital, Madurai. Mostly the brain image
cannot be utilized directly to function. The proposed flow diagram of Segmentation and
Classification of hippocampus in AD using Deep CNN is shown in figure 3.7 . The architecture
34
of deep learning framework for hippocampal segmentation and disease classification with the
integration of two deep CNN models is shown in the figure 3.8.

45
Figure 3.7 Proposed flow diagram of Segmentation and Classification of hippocampus in AD
using Deep CNN

46
 Input: 64×48×64 Segmentation: 64×48×64 Input: 62×48×58

ResNet Block1

Convolution Convolution Convolution

Transition Down
ResNet Block1 ResNet Block1 DenseNet Block
Conv(stride=2)+BP+PReLU

Transition Transition Up Transition Up ResNet Block 2

Down

ResNet Block1 ResNet Block1 DenseNet Block

Transition Up
Transition Transition Up
Down Transition Up

TransConv(stride=2)+BP+PReLU

ResNet Block2 ResNet Block2 DenseNet Block

Skip Connection
Transition Transition Up Transition Up
Down
DenseNet Block
ResNet Block2 DenseNet Block

concatenation Transition Layer

BP+PReLU+Conv+Dropout+Conv
Full connection (Stride=2)

Classification: AD/NC

Fig. 3.8 Architecture of deep learning framework for hippocampal segmentation and disease
classification with the integration of two deep CNN models. BN denotes Batch Normalization;
Conv denotes Convolution; TransConv denotes Transpose Convolution; BP denotes Back
Propagation; and PReLU denotes Parametric Rectified Linear Unit activation.

47
 (a)

(b)

(c)
Figure 3.9 Representation of Images of class a). Healthy Control, b). Alzheimers Disease
c). Mild Cognitive Impairment

3.4.1 Data Source and Pre Processing

61 54
The data were obtained from the ADNI database, the ADNI was launched in 2003 by the
National Institute on Aging (NIA), the National Institute of Biomedical Imaging and
14
Bioengineering (NIBIB), the Food and Drug Administration (FDA), private pharmaceutical
30
companies and non-profit organizations, as a $60 million, 5-year public-private partnership. The
54
primary goal of the ADNI was to test whether serial MRI, Positron Emission Tomography(PET),

48
30
other biological markers, and clinical and neuropsychological assessments could be combined to
measure the progression of MCI and early AD [39] [57]. Here randomly, selected 449
participants, including 97 AD, 233 MCI, and 119 NC subjects, from whom the baseline T1-
weighted MR imaging data and the pre-processed images are available in ADNI. The schematic
flowchart for data pre-processing is shown in the figure 3.10.

Figure 3.10 Schematic flowchart for Data Pre-Processing

62
Table 3.1 shows the demographic and clinical information of these subjects, where the
61
CDR denotes the Clinical Dementia Rating and MMSE denotes the Mini-Mental State
10
Examination. The MR images acquired with 1.5T scanners according to the ADNI acquisition
protocol [58]. All MR images were resampled to 256×256×256 voxels of size 1×1×1 mm³. They
were skull-stripped and cerebellum-removed after correction of intensity in homogeneity using
nonparametric non-uniform intensity normalization algorithm. The transformation of slice of one
sample dataset from ADNI for preprocessing is shown in figure 3.11.

49
 Fig 3.11 Transformation of slice from ADNI dataset for Pre-processing

(a)

(b)

(c)
Figure 3.12 Representation of Preprocessing images a) Original image b) Alzheimers Disease
c) Mild Cognitive Impairment

50
 Figure 3. 13 Representation of Skul stripped and Cerebrum removed after normalization

Table 3.1 Demographic and clinical information (mean ± standard deviation) of the studied
ADNI subjects. AD: Alzheimers Disease; MCI: Mild Cognitive Imapairment; NC: Normal
Control; M: Male; F: Female; MMSE: Mini Mental State Examination; CDR: Clinical Dementia
Rating.
Diagnosis Age Gender (M/F) MMSE Education(YEAR) CDR

AD 75.9 ± 6.8 48/49 23.2 ± 1.8 15.0 ± 3.0 0.8 ± 0.3

MCI 75.2 ± 7.3 145/88 26.9 ± 1.8 15.7 ± 2.6 0.5 ± 0

NC 75.9 ± 5.0 59/60 29.2 ± 1.0 15.6 ± 2.7 0±0

Since directly labeling the hippocampus from scratch by radiologists would be time-
consuming, FIRST from FSL (the FMRIB Software Library) was used as a starting tool to obtain
a rough segmentation. Fig. 3.14 shows hippocampal segmentation results before and after
correction together with their hippocampal image patches. Fig. 3.15 shows the outputs of before
and after correction are depicited in cyan and yellow respectively. The outline of ground truth id
depicited in red.
Table 3.2 shows quantitative measures of hippocampal volumes (i.e., mean, standard
18
deviation, and range) before and after manual correction for different groups of subjects. The
51
mean and standard deviation of hippocampal volume are reduced after correction. Fig. 3.16
5
shows the scatterplots of left and right hippocampal volumes for the studied AD, MCI and NC
subjects.

51
Fig. 3.14. Hippocampal image patches: (a) patches without segmentation labels, (b) patches
overlaid with segmentation labels before manual correction, and (c) patches overlaid with
segmentation labels after manual correction. Subj1, Subj2, and Subj3 are three subjects randomly
selected from the AD, MCI and NC groups respectively.

58
Fig 3.15 The outputs of before and after correction are depicited in cyan and yellow and the
outline of groud truth is depicited in red.

52
Table 3.2 Hippocampal volume (mm³) before and after corrections denoted with Mean ±
63
Standard Deviation (Range). AD: Alzheimer's disease; MCI: mild cognitive impairment; NC:
normal control.

AD MCI NC
Groups
Left Right Left Right Left Right
Before 3390±753 3467±836 3670±840 3759±774 4377±677 4445±596
correction (1263~5223) (1103~5553) (1080~5742) (1051~5891) (2558~6024) (2582~6018)

After 3203±592 3310±674 3471±562 3608±574 4139±524 4221±536

correction (2023~5029) (1928~5283) (1963~5480) (1911~5167) (2735~5462) (2971~5648)

Fig. 3.16 The scatterplots of (a) left and (b) right hippocampus volumes from AD, MCI and NC subjects.

The hippocampus is a small region located in the medial temporal lobe of the human
54
brain. The number of voxels in the hippocampus is much smaller than those of background,
which would result in a severe class-imbalance problem. After image preprocessing and affine
registration, crop each MR image into a 3D image patch with a bounding cube for each
hippocampus [40]. The bounding cubes are defined as 3D axes to extract 3D patches from MR
images. The size of the bounding cube has some effects on hippocampus segmentation. Using a
large bounding cube not only increases the computation cost but also may result in the class-
imbalance problem. However, using a small bounding cube may lose important context

53
information for hippocampal segmentation. To balance the tradeoff, empirically has to set the
64
optimal size of the bounding cube to 64×48×64 voxels. Based on the patches, the deep learning
34
model is constructed for hippocampal segmentation and disease classification in the following
sections.
3.4.2 Multi-task deep CNN for joint hippocampal segmentation and disease classification

The proposed multi-task CNN model for jointly learning hippocampal segmentation and
21
disease classification. CNN is a special type of multi-layer neural network widely used in image
classification and object detection[61]. A volumetric and fully CNN named “V-Net” was
41
proposed for the prostate segmentation in MRIs. Motivated by the success of V-Net in prostate
34
segmentation, we exploited a similar network structure to develop a multi-task deep CNN model
for joint hippocampal segmentation and disease classification.
21
A deep CNN has been formulated to learn residual functions at the convolutional stages to
achieve fast convergence [41]. The two residual blocks are as “ResNet Block1” and “ResNet
Block2”, consisting of 3D convolutional, batch normalization (BN), Parametric Rectified Linear
Unit (PReLU) activation and dropout layers as shown in Fig. 3.17 In ResNet Block1, a residual
function is learned by a short connection: the input is added to the output of the second
28
convolutional layer. ResNet Block2 consists of two convolutional layers and the input of each
45
block is added to both the outputs of the second and third convolutional layers to learn the
residual function. The kernels are trained with supervision from the batches of MRI data. Small
kernels have fewer numbers of parameters to train for fast inference. Large kernels learn much
16
more complex patterns and have a stronger expressive power. This effect can be achieved by
stacking more convolutional layers of small kernels. Therefore, the kernel size is set to 3×3×3 for
all convolutions. The learned filters are convolved with the input image followed by a non-linear
PReLU activation and a feature map is generated for each filter.

54
Fig. 3.17 The network architecture of (a) ResNet Block1 and (b) ResNet Block2, consisting of
3D convolution, PReLU, BN, and dropout layers.

As illustrated in Fig. 3.8, this multi-task deep CNN model consists of two parts: shown
12
on the left is a procedure of compressing feature maps, and shown on the right is a procedure of
decompressing feature maps to their original patch size. The left compression part has two
blocks of ResNet Block1 and two blocks of ResNet Block2 followed by down-sampling for each
block, while the right decompression part comprises of two blocks of ResNet Block1 and one
block of ResNet Block2 followed by up-sampling for each block. Table 3.3 lists the details of
21
relevant parameters for the multi-task deep network model. In the compression part, down-
sampling is used to reduce the size of feature maps and increase the receptive field of features in
the subsequent layers. It is implemented by convolution with the kernels of size 2×2×2 and stride
2. In the decompression part, the spatial support of the lower resolution feature maps is expanded

55
to extract features and assemble the necessary information to generate a volumetric segmentation
mask [42]. The up-sampling by de-convolution is performed with the kernels of 2×2×2 and stride
30
2. Then, a convolutional layer with 1×1×1 kernel and stride 1 is performed to generate the
13
outputs with the same size as the input patch, which are converted to probabilistic segmentation
of the hippocampal regions by applying voxel-wise softmax. Finally, the threshold is set to 0.5 to
convert the probabilistic output to a binary mask.
Table 3.3 The network architecture of the multi-task deep model (the layers with * are used for
classification and the unit of input patch is 1×1×1 mm³).
41
Layers Output Size Filter size, stride, number
Input Layer 64×48×64 -
Convolution 16, 64×48×64 3 × 3 × 3, 1, 16

ResNet Block1 16, 64×48×64 [33 ×× 33 ×× 3,3, 1,1, 16

16]
Down-sampling 32, 32×24×32 2 × 2 × 2, 2, 32

ResNet Block1 32, 32×24×32 [33 ×× 33 ×× 3,3, 1,1, 32

32]
Down-sampling 64, 16×12×16 2 × 2 × 2, 2, 64

ResNet Block2 64, 16×12×16 [3 3× 3× 3× 3,× 3,1, 1,6464× 2]

Down-sampling 128, 8×6×8 2 × 2 × 2, 2, 128

ResNet Block2 (*) 128, 8×6×8 [3 3× 3× 3× 3,× 3,1, 1,128128× 2]

Up-sampling 64, 16×12×16 2 × 2 × 2, 2, 64

ResNet Block2 (*) 32, 32×24×32 [3 3× 3× 3× 3,× 3,1, 1,6464× 2]

Up-sampling 32, 32×24×32 2 × 2 × 2, 2, 32

ResNet Block1 (*) 32, 32×24×32 [33 ×× 33 ×× 3,3, 1,1, 32

32]
Up-sampling 16, 64×48×64 2 × 2 × 2, 2, 16

ResNet Block1 (*) 16, 64×48×64 [33 ×× 33 ×× 3,3, 1,1, 16

16]
Convolution 64×48×64 1 × 1 × 1, 1, 1

56
 As shown in Fig. 3.8, the deep network outputs the hippocampal mask and disease status
prediction. For hippocampal segmentation of subject m, the optimization objective is to
minimize the Dice loss function which evaluates the capability of our model to segment
hippocampal voxels from the background:
푁
2 ∑푖=1 푝푖푞푖 +∈
퐿푚
푠 =1― (3.10)
∑푁 2
푖=1 푝푖 +
푁
∑푖=1 푞2푖 + ∈
1
where N is the number of total voxels on the segmentation output; and are the predicted
46
segmentation result and the ground truth label of voxel i, respectively; - is a small value to
prevent denominator from being zero. The Dice loss function can deal with situations where a
strong imbalance exists between the numbers of foreground and background voxels (Sled et al.,
1998). As to classification, a fully connected layer is stacked to concatenate the outputs of the
compression part and each layer of the decompression part, which integrates information from
44
two sources to enhance classification accuracy. The loss function for classification of subject m
is the categorical cross-entropy loss to evaluate the difference between the predicted label and
the ground truth label as follows:

푐 = ― (푦푚log 푦푚 + ( 1 ― 푦푚 )log(1 ― 푦푚))

퐿푚 (3.11)
55
The loss function for the multi-task dep CNN model is a weighted sum of the
segmentation loss and classification loss, calculated as follows:
1 푀
퐿푚 = 훼 . 퐿푠 + ( 1 ― 훼 ). 퐿푐 = 푠 ― (1 ― 훼).(푦푚 log 푦푚 + (1 ― 푦푚) log(1 ― 푦푚 ))]
[훼 . 퐿푚
푀 푚=1

(3.12)
26
where M is the total number of subjects; "# and " are the ground truth label and the
predicted label for subject m. The parameter * ∈ .0, 1/ is a weight to adjust for the losses in
34
training the hippocampal segmentation and disease classification [42]. During the training
procedure of the multi-task deep CNN model, the segmentation task is more important than
classification at the early stage while the classification task takes effect at the late stage when the
segmentation objective function value converges [43]. In our implementation, the value of *
evolves throughout the training process. During the initial warm-up phase, * is set to 1 to
emphasize on the segmentation task. Then it changes to 0.5 for multi-task training. In the final
37
phase, * is set to 0 to focus on the classification task. The joint optimization of the multi-task

57
network model is performed with the Adam method, and a back propagation algorithm is used to
calculate the network gradients.

3.4.3 The deep 3D DenseNet model

30
Traditionally, the shape and volume features are calculated from the hippocampal mask
41
for AD diagnosis. Different from that, we propose to construct a deep 3D DenseNet model based
on the hippocampal region to learn the discriminative features for AD diagnosis. Although the
features learned by the multi-task deep CNN model can capture some important information for
disease classification, they are still not sufficient for accurate diagnosis of the disease. Therefore,
a 3D patch of a fixed size centered on the centroid of a segmentation mask is extracted to build
the proposed CNN model. To enhance the representation power, CNN becomes increasingly
deep to learn features [59].

The information may loss when the input passes through many layers to reach the end of
network. DenseNet is proposed to connect each layer to every other layer in a feed-forward
fashion, which increases direct connections between the low and high level. Compared with
traditional CNN, DenseNets have several advantages. First, they can alleviate the vanishing-
gradient problem since there is a direct connection from the low to high-level layers. Second,
feature propagation is strengthened to reuse the low-level features. Third, they can reduce the
number of parameters.
30
Table 3.4 shows the structure and parameters of proposed 3D DenseNet model,
consisting of a convolutional layer, 4 dense blocks, 3 transition layers, a global average pooling
layer, and a softmax layer. First, a convolutional layer was added to the input layer with stride 2,
followed by dense blocks. Then, the dense block uses dense connectivity through which the l-th
layer receives the feature maps of all preceding layers as follows:

푥푙 = 퐻푙 ([푥0, 푥1, …….., 푥푙―1]) (3.13)

Where [푥0, 푥1, …….., 푥푙―1] is the concatenation of feature maps from all previous layers
into a single tensor, and 퐻푙 denotes a composite nonlinear transformation function of four
consecutive operations: BN, PReLU, 3×3×3 convolution, and voxel-wise dropout. Dense block
includes three dense layers with each layer consisting of one 1×1×1 and one 3×3×3
convolutional layers, two BN layers and two activation layers. Every dense layer receives the
58
feature maps of all previous dense layers by shortcut connections. A transition layer, which
consists of five consecutive operations: BN, PReLU, a 1×1×1 convolution, voxel-wise dropout
and a 3×3×3 convolution with a 2×2×2 stride, is set between two dense blocks for feature map
reduction. Following the last dense block, an average pooling and a softmax classifier are
appended to reduce feature dimension and classify disease status. The subject labels are used
through back-propagation for updating the weights of DenseNet. All layers of DenseNet receive
direct supervision from loss function through shortcut connections [43].

Table 3.4 The network architecture of our deep DenseNet Model (input patch unit is voxel of
1×1×1 mm³).

Layers Output Size Filter size, stride, number

Input Layer 62×48×58 -
Convolution 64, 31×24×29 3×3×3, 2, 64, conv

Dense Block (1) 96, 31×24×29 [13 ×× 13 ×× 1,3, 1,1, 64,

16, 푐표푛푣] × 2
푐표푛푣

Transition Layer 48, 16×13×15 [12 ×× 12 ×× 1,2, 1,2, 48,

48, 푐표푛푣]
푐표푛푣

Dense Block (2) 80, 16×13×15 [13 ×× 13 ×× 1,3, 1,1, 64,

16, 푐표푛푣] × 2
푐표푛푣

Transition Layer 40, 9×7×8 [12 ×× 12 ×× 1,2, 1,2, 40,

40, 푐표푛푣]
푐표푛푣

Dense Block (3) 72, 9×7×8 [13 ×× 13 ×× 1,3, 1,1, 64,

16, 푐표푛푣] × 2
푐표푛푣

Transition Layer 36, 6×4×6 [13 ×× 13 ×× 1,3, 1,1, 36,

36, 푐표푛푣]
푐표푛푣

Dense Block (4) 68, 6×4×6 [13 ×× 13 ×× 1,3, 1,1, 64,

16, 푐표푛푣] × 2
푐표푛푣

Global Average Pooling 68, 1×1×1 -

Softmax Layer 2 -

For training DenseNets, the initial weights for the whole network are uniform and Adam
optimizer is adopted with a learning rate of 10-4, The network is stable after iteration of 120
epochs. The batch size is set to 64 and PReLU is used for each neuron of DenseNet. Dropout
27
layers are used to alleviate the over fitting problem. The DenseNet models for the left and right

59
21
hippocampi have the same structure but are trained individually with different patches for the
classification task.

3.4.4 Final multi-model ensemble classification

The multi-task deep CNN model captures the multi-level features for joint hippocampal
segmentation and disease classification, while the deep 3D DenseNet model features from the
image patches of the hippocampus for disease classification [44]. To integrate these deep
41
models, an extra fully connected layer above the concatenation features from deep models for
disease classification. The DenseNet models and multi-task model are individually trained, and a
21 30
fullyconnected layer followed by a softmax layer is finely tuned to make the final classification.
They are implemented with Keras library in the framework of Tensorflow. The proposed multi-
model deep network framework out performs the single-model approaches [61].

3.5 EXPERIMENTAL OUTCOMES

15
The proposed deep learning framework was tested on the structural MRI data from the
baseline visits of 449 ADNI participants consisting of 97 AD, 233 MCI, and 119 NC subjects.
34
Both hippocampal segmentation and disease classification tasks were conducted for method
41
evaluation. The classification task was tested to distinguish AD vs. NC and MCI vs. NC. Here
randomly divided the whole MRI data set into 5 folds and 5-fold cross-validation was used to
28
train and test the proposed method. Each time, one fold of dataset was used for testing, while the
other 4 folds were for training. For tuning the parameter of the iteration number in the deep
learning model, randomly select 10% of training data as the validation set while the remaining
training data were used to train model. The iteration number was increased gradually one by one
2
for parameter tuning and the optimal values were 110, 94, 96, 125 and 124 for the 5 different
28
testing folds, showing relatively robust results. The testing set was not used for model training
and parameter tuning but for general performance evaluation.
6
To evaluate the performance of hippocampal segmentation are calculated in four measures: Dice
similarity coefficient (DSC), sensitivity (SEN_S), positive predicted value (PPV) and volume
error (VE):

2푇푃
퐷푆퐶 = (3.14)
2푇푃 + 퐹푃 + 퐹푁

60
 푇푃
푃푃푉 = (3.15)
푇푃 + 퐹푃

푇푃
푆퐸푁 = (3.16)
푇푃 + 퐹푁

2 |푣훼 ― 푣푚 |
푉퐸 = 푣훼 + 푣푚 (3.17)

where TP denotes the true positives, i.e., the predicted hippocampal voxels inside the
positive regions of ground-truth; FP denotes the false positives, i.e., the predicted hippocampal
voxels outside the positive regions of ground-truth; FN denotes the false negatives as the
predicted background voxels inside the positive regions of ground-truth; 푉훼 and 푉푚 denote the
hippocampal volumes by automatic and manual segmentation, respectively. For classification
tasks, four performance measures were computed for evaluation, including classification
accuracy (ACC), sensitivity (SEN_C), specificity (SPE), receiver operating characteristic (ROC)
42
curve and area under ROC curve (AUC). ACC is the proportion of correctly classified subjects
among the whole population. SEN_C is the proportion of AD/MCI patients correctly classified.
SPE is the proportion of correctly classified NC subjects. The ROC curve is generated by
plotting the true positive rate (TPR) against the false positive rate (FPR) at various threshold
settings on the prediction scores.

The implementation in the trade-off parameter α is in the loss function of multi-task deep
CNN framework evolved from 1 to 0 throughout the training process [45]. The segmentation
and classification tasks are tested by setting α to 1 or 0 for single task learning and setting α to
0.5 for multi-task learning. Shown in Table 3.5 and 3.6 is the comparison of the segmentation
and classification results by setting different α’s for single and multi-task learning as well as the
proposed adaptive method. From the results, the multi-task learning with fixed α performs better
21
than single-task learning for classification. The proposed adaptive method performs better than
the single-task and multi-task learnings with fixed α for both segmentation and classification.
From Table 3.5, the better segmentation results on NC subjects than AD subjects even though
28
the training set includes AD subjects. This may be caused by the hippocampal atrophy of AD
and its variation among different subjects, which introduces difficulties in both automatic and
manual segmentations. The comparison of segmentation results by setting different α for single-

61
task and multi-task learning are shown in figure 3.18 and comparison of classification results by
setting different α for single-task and multi-task learning are shown in figure 3.19 respectively.

Table 3.5. Comparison of segmentation results by setting different α for single-task and multi-
task learning. ALL: all studied subjects; AD: Alzheimer's disease; MCI: mild cognitive
impairment; NC: normal control; DSC: Dice similarity coefficient; PPV: positive predicted
value; SEN_S: segmentation sensitivity; VE: volume error.

Performance α=1 (Single-task) α=0.5 (Multi-task) Adaptive α (Multi-task)

(%) ALL AD MCI NC ALL AD MCI NC ALL AD MCI NC

DSC 86.7 86.3 86.7 87.2 81.8 81.5 81.7 82.3 87.0 86.4 86.9 87.5

PPV 84.7 84.4 84.9 86.5 80.4 79.6 80.5 81.0 84.6 83.6 84.0 85.1

SEN 88.9 88.5 88.9 88.4 84.5 84.2 84.8 84.6 89.7 89.5 90.0 89.7

VE 6.1 6.6 6.3 5.7 10.5 10.7 10.5 10.3 6.0 6.6 6.1 5.6

Comparison of segmentation results by setting different α

for single-task and multi-task learning
100
90
80
70
60
50 DSC
40
30
PPV
20 SEN
10
0
VE
ALL AD MCI NC ALL AD MCI NC ALL AD MCI NC
α=1 α=0. Ada
(Sin 5 ptiv
gle- (Mu eα
task) lti- (Mu
task) lti
62
Fig 3.18 Comparison of segmentation results by setting different α for single-task and multi-task learning

Table 3.6 Comparison of classification results by setting different α for single-task and multi-
63
task learning. AD: Alzheimer's disease; MCI: mild cognitive impairment; NC: normal control;
ACC: classification accuracy; SEN_C: classification sensitivity; SPE: specificity; AUC: area
10
under receiver operating characteristic curve

Performance α=0 (Single-task) α=0.5 (Multi-task) Adaptive α (Multi-task)

(%) AD vs NC MCI vs NC AD vs NC MCI vs NC AD vs NC MCI vs NC

ACC 75.0 67.8 78.0 69.5 80.1 71.5

SEN 74.2 84.9 76.3 77.6 79.9 73.2

SPE 75.6 50.0 79.4 61.6 80.3 69.8

AUC 82.5 71.6 84.9 72.5 86.6 74.4

Comparison of classification results by setting different α for

single-task and multi-task learning
100
90
80
70
60
50 ACC
40 SEN
30
SPE
20
AUC
10
0
AD vs NC MCI vs AD vs NC MCI vs AD vs NC MCI vs
α=0 NC α=0.5 NC Adaptive NC
(Single- (Multi- α (Multi-
task) task) task)

Fig 3.19 Comparison of classification results by setting different α for single-task and multi-task
learning

63
3.5.1 Results on hippocampal segmentation
6
In this experiment, the performance of the proposed framework on hippocampus
26
segmentation in terms of DSC, PPV, SEN_S, and VE and compared the results with different
65
parameter settings and also performed with empirical comparation with a few other competing
methods. Fig. 3.20 (a), (b), (c) and (d) show the results of DSC, PPV, SEN_S, and VE,
respectively, by the proposed segmentation method in 5 cross-validation trials using box-plots.

Fig. 3.16. The segmentation performances of (a) DSC, (b) PPV, (c) SEN and (d) VE in 5 folders.

The inputs of our multi-task deep CNN framework were 3D image patches obtained by
cropping MR images with a bounding cube for each hippocampus. The first experiment was to
6
test the effects of various bounding cube sizes on the segmentation results. the bounding box of
cube size increased gradually from 48×40×48, 64×48×64 to 88×80×88 to test hippocampal
segmentation [46]. The results in Table 3.7 demonstrate that the segmentation performance is

64
37
improved by increasing the cube size from 48×40×48 to 64×48×64 and is degraded by further
increasing the cube size to 88×80×88. While a small boundary bounding cube might have limited
context information, a large bounding cube could include unnecessary background voxels
increasing both overfitting risk and computational burden. With this observation, the bounding
52
cube to 64×48×64 is set in the following experiments. Comparison of segmentation results by
34
different bounding cube sizes for multi-task deep CNN model is shown in figure 3.21.

Table 3.7 Comparison of segmentation results by different bounding cube sizes for multi-task
deep CNN model (the patch unit is 1×1×1 mm³). DSC: Dice similarity coefficient; PPV: positive
predicted value; SEN_S: segmentation sensitivity; VE: volume error.

Performance Training time Testing time

DSC PPV SEN VE
(%) (hour) (minute)

48×40×48 86.2 84.8 87.5 6.4 1.1 1.8

64×48×64 86.7 84.7 88.9 6.1 1.4 2.2

88×80×88 85.7 83.8 88.2 6.8 1.8 2.6

Comparison of segmentation results by different bounding

cube sizes for multi-task deep CNN model
100
90
80
70
60
50 48×40×48
40
64×48×64
30
20
88×80×88
10
0
DSC PPV SEN VE Training Testing
time time
(hour) (minute)

65
Fig 3.21 Comparison of segmentation results by different bounding cube sizes for multi-task
deep CNN model

Second, to compare the segmentation results of our method with other three methods
including FSL and two deep learning-based methods, denoted as “FSL”, “Thyreau’s method”
and “Cao’s method”, respectively.

FSL is a comprehensive library of analysis tools for fMRI, MRI and DTI brain images.
41
The FIRST algorithm in the FSL library was used for hippocampal segmentation on our dataset.
However, 27 subjects have large segmentation errors due to internal registration failure. For fair
comparison, these subjects were removed for performance evaluation.

Thyreau’s method was proposed to segment the bilateral hippocampi using a deep-
learning appearance model by transferring algorithmic knowledge for large cohort processing. A
wide and variable training set from multiple cohorts was used to train the segmentation model.
37
Since the trained model is available online to test our data set for comparison. It is worth noting
38
that the differences in the definitions of manual segmentations for training and testing data may
bias the evaluation results.
41
Cao’s method was proposed to construct a multi-task deep learning method for joint
24
hippocampal segmentation and clinical score regression [47]. For fair comparison, the multi-task
42
deep network was rebuilt as in, which were trained and tested on our data set. Fig. 3.23 shows
41
the comparison of the segmentation results using different methods for three example subjects
65
selected from the AD, MCI and NC groups of the test data, respectively. The original images, the
segmented results of FSL, Thyreau’s method, Cao’s method, and our method and ground truth
are demonstrated in columns from left to right in Fig. 3.24. From the original images, it is not
48
easy to distinguish the hippocampal regions from the adjacent tissues due to the small difference
between their intensity values. The segmented hippocampal regions obtained by our method
23
appear to be smoother and more accurate than those by other methods. In addition, Table 3.8
shows the comparison of segmentation results in terms of DSC, PPV, SEN_S and VE on the
28
same training and testing data sets. Deep CNN method achieves the best performance compared
to other methods. In particular, proposed method achieves the best PPV, which indicates that it
can effectively detect hippocampal regions from the background. Incorporating this accurately

66
segmented hippocampal region into our subsequent DenseNet can provide meaningful
anatomical information to help improve disease classification.

Figure 3.22 Representation of Segmentation results of the proposed method. A: Original images
18
of the hippocampus area, B: The hippocampus mask added to the original images, C: Deep
learning segmentation of the hippocampus area, D: The hippocampus border in the transvers

67
plane obtained from CNN, E: The hippocampus border in the coronal plane obtained from CNN,
F: The hippocampus border in the sagittal plane obtained from CNN.

Fig. 3.23. Comparison of segmented hippocampal regions by different methods for three subjects
from the test data. Shown from top left, top right, bottom left to bottom right are coronal view,
sagittal view, axial view and 3D rendering, respectively.
28
Table 3.8 Comparison of the segmentation results of the proposed method to other methods.
DSC: Dice similarity coefficient; PPV: positive predicted value; SEN_S: segmentation
sensitivity; VE: volume error.

Performance (%) FSL Thyreau’s method Cao’s method Proposed Method

DSC 72.2 73.8 84.6 87.0
PPV 70.4 68.8 83.7 84.6
SEN 74.6 80.2 85.7 89.7

68
 VE 20.3 18.4 7.2 6.0

Comparison of the segmentation results of the proposed

method to other methods
100
90
80
70
60
50 DSC
40 PPV
30
20 SEN
10 VE
0
FSL Thyreau’s Cao’s Proposed
method method Method
41
Proposed Method with Other Methods

28
Fig 3.24 Comparison of the segmentation results of the proposed method to other methods

52
The hippocampal label definition was used for training and testing in this work, there
may be a bias in the comparison of segmentation results because some competitive methods were
not trained using this particular label definition [48]. To address this issue, furthermore, we
1
tested our segmentation method on a new dataset with the EADC-ADNI Harmonized Protocol
(HarP) for manual hippocampus labeling. This dataset consists of 135 subjects from the ADNI
MRI cohort balanced by magnet field strength, age ranges, diagnosis, medial temporal atrophy,
and scanner manufacturers. First, the trained segmentation models are directly tested on new
dataset, denoted as “Our model on the new dataset”. Second, segmentation model on this new
dataset are retrained and tested by 10-fold cross-validation strategy, denoted as “Retrained model
66
on the new dataset”. The comparison of results in Table 3.9 demonstrates that the performance is
poor when our model is directly tested on the new dataset due to potentially different strategies
for labeling the hippocampus. But the retrained model can work well on this new HarP dataset.
The results are presented where an indicative in saggital, coronal and axial slicw which is shown
28
as best, median and worst case. The qualitative results for the HarP dataset on the outline of
hippocampal segmentation and ground truth are depicited in yellow and red respectively are

69
shown in figure 3.25. The segmentation results in the deep model trained and tested on different
datasets are represented in the figure 3.26.

Figure 3.25 Results for the HarP dataset on the outline of hippocampal segmentation and ground
truth are depicited in yellow and red.

70
Table 3.9 Segmentation results in the deep model trained and tested on different datasets. DSC:
Dice similarity coefficient; PPV: positive predicted value; SEN_S: segmentation sensitivity; VE:
volume error.
Performance (%) DSC PPV SEN VE

Proposed model on proposed dataset 87.0 84.6 89.7 6.0

Proposed model on the new dataset 72.2 81.0 65.7 21.6

Retrained model on the new dataset 86.7 85.3 88.2 6.6

Segmentation results in the deep model trained and tested on

different datasets
100

80
Proposed model on
proposed dataset
60
Proposed model on the
40 new dataset
Retrained model on the
20
new dataset
0
DSC PPV SEN VE

Fig 3.26 Segmentation results in the deep model trained and tested on different datasets

3.5.2 Results on disease classification

The proposed strategy, incorporating the multiple tasks deep CNN and DenseNet models
was tried on sickness characterization. The principal test was to test the arrangement exhibitions
of DenseNet when contrasted with other deep organizations, for example, the famous LeNet and
VGGNet [49]. The LeNet network comprises of 2 convolutional layers followed by 2 completely

71
associated layers, while VGGNet consists of 13 convolutional layers and 3 completely associated
layers. Subsequently, VGGNet has a more profound design with essentially a bigger number of
layers than LeNet. They were executed with the delivered codes by supplanting the 2D
convolutions with 3D ones. Table 3.10 denotes the order consequences of AD versus NC and
MCI versus NC by various profound organizations. Here the DenseNet beat LeNet and VGGNet.
41
Table 3.10 Comparison of different deep networks for classification of AD vs. NC and MCI vs.
63
NC. AD: Alzheimer's disease; MCI: mild cognitive impairment; NC: normal control; ACC:
classification accuracy; SEN_C: classification sensitivity; SPE: specificity; AUC: area under
receiver operating characteristic curve.

AD vs. NC (%) MCI vs. NC (%)

Methods
ACC SEN SPE AUC ACC SEN SPE AUC

LeNet 83.8 73.2 92.4 84.3 68.8 79.3 63.4 70.8

VGGNet 84.7 77.3 90.8 85.8 70.9 81.9 65.2 72.6

DenseNet 86.6 79.4 92.4 87.2 74.1 77.2 68.1 72.8

41
The second experiment was to test the effects of different patch sizes used to train
30
DenseNet on the classification performance. The large patch covers more information near the
hippocampus for classification. In this work, the patch size was gradually increased from
50×40×48 to 68×50×64. Table 3.11 shows the results of different patch sizes using DenseNet for
41 56
classifying AD vs. NC and MCI vs. NC. The classification results are compared with the method
combining multi-task deep CNN model and DenseNet is obtained, as shown in Table 3.12.

The comparison demonstrates the classification results are improved by increasing the
patch size from 50×40×48 to 62×48×58. The large patch requires more memory and
computational time for both training and testing. The patch of 62×48×58 was used to train
DenseNets in our following experiments. It should be mentioned that this patch size is optimal
23
based on the results of our dataset but it may not be optimal for other datasets. The results show
26
that there is no great difference in the AUC with the patch size over 62×48×58. For simplicity,
the same patch size can be used for both hippocampus segmentation and disease classification.
28
72
Table 3.11 Classification results of different patch sizes for DenseNets (the patch unit is 1×1×1
63
mm³). AD: Alzheimer's disease; MCI: mild cognitive impairment; NC: normal control; ACC:
classification accuracy; SEN_C: classification sensitivity; SPE: specificity; AUC: area under
receiver operating characteristic curve.

Patch AD vs. NC (%) MCI vs. NC (%)

size ACC SEN SPE AUC ACC SEN SPE AUC
50×40×48 85.2 76.2 92.4 86.3 72.3 74.5 68.1 70.8
56×42×52 85.6 82.5 88.2 86.8 72.9 77.2 64.7 71.9
62×48×58 86.6 79.4 92.4 87.2 74.1 77.2 68.1 72.8

68×50×64 86.1 84.5 87.4 87.3 73.2 74.6 70.7 73.2

36
Table 3.12 Classification results of different patch sizes for the combination method (the patch
63
unit is 1×1×1 mm³). AD: Alzheimer's disease; MCI: mild cognitive impairment; NC: normal
control; ACC: classification accuracy; SEN_C: classification sensitivity; SPE: specificity; AUC:
10
area under receiver operating characteristic curve.

AD vs. NC (%) MCI vs. NC (%)

Combination
ACC SEN SPE AUC ACC SEN SPE AUC
50×40×48 87.0 85.6 88.2 90.3 74.1 72.8 76.7 76.4

56×42×52 87.9 84.5 90.8 90.9 74.4 77.6 68.1 77.1

62×48×58 88.9 86.6 90.8 92.5 76.2 79.5 69.8 77.5

68×50×64 88.0 87.6 88.2 91.4 74.7 76.8 70.9 76.9

The third experiment was to test the classification performances of the multi-task deep
41
CNN model, the DenseNet model and the combination model. Table 3.13 shows the results of
34
three models for classifying AD vs. NC and MCI vs. NC. Fig. 3.27 (a) and (b) show the ROC
41
curves of three models for classifying AD vs. NC and MCI vs. NC, respectively. The DenseNet
model greatly outperformed the multi-task deep CNN model for disease classification,
suggesting that the features learned by DenseNet is more important for disease classification.
The combination model outperformed both the multi-task deep CNN and DenseNet models,

73
indicating that the features of individual models could contain complementary information for
disease classification.

Table 3.13 Comparison of the multi-task deep CNN, DenseNet and their combination methods
for classifications of AD vs. NC, and MCI vs. NC. AD: Alzheimer's disease; MCI: mild
cognitive impairment; NC: normal control; ACC: classification accuracy; SEN_C: classification
67
sensitivity; SPE: specificity; AUC: area under receiver operating characteristic curve.
41
AD vs. NC (%) MCI vs. NC (%)
Methods
ACC SEN SPE AUC ACC SEN SPE AUC
Multi task deep CNN 80.1 79.9 80.3 86.6 71.5 73.2 69.8 74.4
DenseNet 86.6 79.4 92.4 87.2 74.1 77.2 68.1 72.8
Combination 88.9 86.6 90.8 92.5 76.2 79.5 69.8 77.5

68
Fig. 3.27 Comparison of the ROC curves with the multi-task deep CNN, DenseNet and their
41
combination models for classifying (a) AD vs. NC and (b) MCI vs. NC.

3.5.3 Comparison with other methods

6
The proposed method was compared to other existing methods based on structural MRI
data. First, the proposed deep learning-based method was compared to other methods based on
hand-crafted features such as the hippocampal volume, the GM volumes of ROIs and voxel-wise
features. An automatic method was proposed for hippocampal segmentation and hippocampal
volumes were extracted for AD diagnosis. Instead of just focusing on hippocampal volumes, 93

74
ROIs were parcellated in the brain and their volumetric features were calculated to train SVM
41
classifiers for AD classification [50]. To capture fine-level features, the voxel-wise tissue
25
densities were calculated for ensemble sparse classification of AD. These features are hand-
crafted from one-region, multi-region to voxel levels and are widely used for AD diagnosis.
21
To extract the hippocampal volumes, the segmentation mask was used to obtain the
40
multi-task deep model. The volumes from the left and right hippocampi were normalized by the
total intracranial volume for classification. The hippocampus segmentation and nonlinear
registration using the FAST model in the FSL package and the registration tool of HAMMER.
41
FAST was used to segment the MRIs into three different tissues: GM, WM, and CSF, while
25
HAMMER was used for the nonlinear image registration and mapping the image onto 93
manually labeled ROIs. For each labeled image, the normalized GM volumes from 93 ROIs were
calculated as features for classification. After the image warping through HAMMER, the warped
mass-preserving tissue volumes reflected the spatial distribution of tissues in an original brain
and were used as the voxel-wise features for classification [51, 52]. Furthermore, Lasso was used
51
to select the most discriminative ROI and voxel-wise features for classification. To evaluate the
impact of features, thesame preprocessing procedures were performed for each image, and the
21
Multi-Layer Perceptron (MLP) model with two fully connected layers followed by a softmax
layer was used for classification with different features.
37
In addition, deep CNN method is to build a deep DenseNet model on the whole brain
image, which is denoted as “Whole MRI DesnseNet”. The whole MR image of 256×256×256
41
voxels is too large to train the DenseNet model with our GPU GTX 1080 Ti because of the
memory limit. Therefore, the voxels of zeros values are removed and down-sampled the images
by 2, and finally obtained images of 98×78×76 voxels for training a DenseNet model. The data
augmentation are performed by shifting the down-sampled image to train a robust model. Table
3.14 shows the comparison of classification performances with the above features as well as our
proposed method. Fig. 3.29 (a) and (b) illustrate theROC curves of our proposed method and
34 41 21
other methods for classifying AD vs. NC and MCI vs. NC, respectively. It is worth noting that
the results in Table 3.14 and Fig. 3.28 were obtained based on different feature extraction
41
methods, not the design of classifier, so the results may be different from those reported in the
literature.

75
Table 3.14 Comparison of different features for classification of AD vs. NC, and MCI vs. NC.
63
AD: Alzheimer's disease; MCI: mild cognitive impairment; NC: normal control; ACC:
classification accuracy; SEN_C: classification sensitivity; SPE: specificity; AUC: area under
receiver operating characteristic curve.

AD vs. NC (%) MCI vs. NC (%)

Method 41
ACC SEN SPE AUC ACC SEN SPE AUC
Hippocampal Volumes 80.3 73.5 86.2 86.3 68.7 84.4 38.5 70.0

ROI features 84.7 77.4 90.5 88.1 73.5 83.9 53.4 73.6
Whole MRI DenseNet 84.3 83.8 84.5 86.2 70.0 83.4 48.1 74.7

Voxel-wise features 86.1 84.0 87.9 90.1 74.4 87.1 50.0 75.3

Proposed Method 88.9 86.6 90.8 92.5 76.2 79.5 69.8 77.5

Comparison of different features for classification of AD vs. NC, and MCI

vs. NC
100
90
80 Methods
70
60
50 Hippocampal Volumes
40
30 ROI features
20
10 Whole MRI DenseNet
0

Voxel-wise features

Proposed Method

Fig 3.28 Comparison of different features for classification of AD vs. NC, and MCI vs. NC

76
Fig. 3.29. Comparison of ROC curves with different features for classifying (a) AD vs. NC and
(b) MCI vs. NC, where AD: Alzheimer's disease; MCI: mild cognitive impairment; NC: normal
control.
54
Furthermore, comparison of classification results are those reported in the literature using
the baseline sMRI data of ADNI, as shown in Table 3.15. The compared methods include the
conventional learning-based methods and deep learning-based methods. In the conventional
21
learning-based methods, the SVM classifier was used for classification with the engineered
features from ROIs or landmarks. A multi-kernel learning was proposed to simultaneously
conduct feature selection, manifold learning and over-sampling with the ROI features for AD
41
diagnosis. The deep learning models such as Residual and plain 3D CNNs and stacked auto-
21
encoders were investigated for AD and MCI diagnosis [51, 52].

A fully hierarchical convolutional network was proposed to identify discriminative

patches and regions from the whole brain sMRI, upon which multi-scale features were jointly
learned and fused to construct hierarchical classification models for AD diagnosis. The results
10
show that the proposed method based on hippocampal region obtained competitive performances
6
when compared with the methods that use whole-brain sMRI. These empirical comparison
21
studies validated the efficacy of proposed method. It is worth noting that the different results
may be caused not only by the methods and but also by different ADNI subjects and partitions of
28
training and testing sets.

77
 41
Table 3.15 Comparison of our results with published results for the classification of AD vs. NC
and MCI vs. NC. AD: Alzheimer's disease; MCI: mild cognitive impairment; NC: normal
control; ACC: classification accuracy; SEN_C: classification sensitivity; SPE: specificity; AUC:
10
area under receiver operating characteristic curve.

AD vs. NC (%) MCI vs. NC (5)

Method Subjects AC SE AU
SEN SPE AUC ACC SPE
C N C

Ye et al., 2016 51AD+99MCI+52NC 87.3 88.4 86.2 93.0 68.2 76.9 51.1 71.0

Zhang et al., 2016 51AD+99MCI+52NC 83.1 80.5 85.1 - 73.6 75.3 69.7 -

Cao et al., 2017 192AD+397MCI+229N 88.6 85.7 90.4 89.8 71.9 79.0 60.7 75.3
C

Lian et al., 2018 35AD+429NC 90.3 82.4 96.5 95.1 - - - -

Liu et al., 2014 65AD+169MCI+77NC 87.8 88.6 87.2 - 76.9 74.2 78.1 -

Korolev et al.,
50AD+43pMCI+61NC 80.0 - - 87.0 61.0 - - 65.0
2017

Proposed method 97AD+233MCI+119NC 88.9 86.6 90.8 92.5 76.2 79.5 69.8 77.5

3.5.4 Discussion
34
The proposed multi-model learning-based method out performs the single model methods
35
for both segmentation and classification tasks. As for the segmentation, the proposed method
achieved the DSC of 87.0% on the ADNI dataset and DSC of 86.7% on the HarP dataset. The
result comparison in Table 3.8 shows our method performs a few competing methods including
the FSL, Thyreau’s and Cao’s methods in terms of DSC, PPV, SEN_S, and VE.

As for the classification, Table 3.10 shows that our proposed DenseNet model performs
28
better than the other popular deep networks such as the LeNet and VGGNet. The effects of patch
47
sizes on the classification performance were also analyzed for DenseNet as shown in Tables 3.11
and 3.12. Furthermore, the multi-model deep network by integrating the multi-task deep model
and DenseNet improved the performance of disease classification as shown in Table 3.13. This
78
indicates that the two individual deep models learned complementary features for disease
classification. When compared with other methods in Table 3.14, deep CNN method achieved a
41
very competitive classification accuracy of 88.9% for AD vs. NC classification, which is higher
than the ROI features (84.7%) and voxel-wise features (86.1%). Our method also achieved an
21 41
accuracy of 76.2% for MCI vs. NC classification, which is higher than the ROI features (73.5%)
and voxel-wise features (74.4%). The features from multiple ROIs were more informative than
hippocampal volumes and yielded better performance.

The voxel-wise features also demonstrated better discriminant ability than the ROI
27
features because of the fine-level features. Compared to the DenseNet on the whole image, our
method built the DenseNet on the hippocampus, which was easier to be trained than on the
whole image with several hundreds of training subjects. The proposed classification method was
1
tested on the additional HarP dataset consisting of 45 AD, 46 MCI, and 44 NC subjects. The
34 41
AUCs are 89.8% and 72.2% for classifying AD vs. NC and MCI vs. NC, respectively, which
6
shows the robustness when compared with the other results on the same dataset.
30
Training a deep model is not an easy task as the current datasets for hippocampal
segmentation and disease classification are relatively small compared to the computer vision
tasks [62]. To alleviate this challenge, the data augmentation by shifting three coordinates was
used to improve the robustness of the model. Fig. 3.25 (a) and (b) show the loss curves of the
69
multi-task deep model for both training and validation on hippocampus segmentation and disease
classification, respectively, while Fig. 3.30 (c) shows the loss curves of DenseNet model for both
30 41
training and validation on the classifications of AD vs. NC and MCI vs. NC. The results show
that the loss curves of hippocampal segmentation converges after 40 epochs, while the loss
curves of all disease classifications converge when the training epochgrows to 80. The loss
convergence of AD vs. NC classification is faster than that of MCI vs. NC because the
classification of MCI vs. NC is more challenging than that of AD vs. NC.

79
Fig. 3.30. The loss curves of the multi-task deep model for (a) hippocampus segmentation and
34
(b) disease classification, and (c) the loss curves of DenseNet model for classifying AD vs. NC
41
and MCI vs. NC on both training and validation, denoted as “AD vs. NC train”, “MCI vs. NC
train”, “AD vs. NC val”, “MCI vs. NC val”, respectively.
30
As for the computational complexity, the proposed deep learning combination method
includes both the offline training and online testing stages. In the offline training stage, the
computational cost includes training the multi-task deep CNN model and the DenseNet model,
6
which take 0.93 hours and 1.40 hours in our experiments, respectively. Thus, it takes about 2.33
hours to train the whole combination model. In the online testing stage, it takes 0.29s and 0.85s
41
on average to test the proposed algorithm for segmentation and classification of a given image,
30
respectively, which demonstrates the usefulness of the proposed method in a real application.
The over-fitting problem was alleviated by using dropout techniques and data augmentation. The
inverted dropout was used on the CNN layers, which performed the scaling at training time,
leaving the forward pass at test time untouched.

Although the proposed method can jointly learn the feature extraction and classification
model to achieve optimal diagnosis performance, it has limitations in medical interpretation and
41
characterization of the learned features relevant to disease (i.e., AD or MCI) for clinical
application [63]. The learned features are limited in providing sucient clinical information for
understanding brain abnormalities. To facilitate the interpretability, a visualization technique
1
proposed on computing the gradient of the class score for the input image. An image-specific
58
class saliency map was generated to highlight the discriminative areas of a given image to

80
disease classification.The saliency map are generated by the proposed method of deep CNN
individual for Alzheimers Disease, Mild Cognitative Impairment and Normal Control are shown
in the figure 3.31, 3.32and 3.33 respectively.

Fig 3.31 Saliency map generated by the proposed method of deep CNN for AD

Fig 3.32 Saliency map generated by the proposed method of deep CNN for MCI

Fig 3. 33 Saliency map generated by the proposed method of deep CNN for Normal

The saliency maps of the hippocampus patches are generated on all test images and then
calculated the average of them for illustration as shown in Fig. 3.34. The saliency maps indicate
how the network learns the importance of corresponding areas in the prediction of disease status.
The deeper the highlighted color is, the more relevant the covered region is to disease diagnosis

81
[64, 65]. The highlighted areas were observed covering the regions that are important for AD
diagnosis, such as hippocampus, amygdala, and parahippocampal gyrus, etc..

41
Fig. 3.34 The saliency maps of hippocampal regions by the proposed method for (a) AD and (b)
MCI patients, where the red color highlighted regions are more relevant to AD/MCI diagnosis.

82
3.6 CONCLUSION
14
In this, a new classification is proposed in framework based on multi-model deep CNNs
34
for jointly learning hippocampal segmentation and disease classification. First, a multi-task deep
CNN model was constructed to jointly learn the features for hippocampal segmentation and
disease classification. Based on the segmented hippocampal region, an additional 3D DenseNet
41
was built to learn the rich and detailed image features for disease classification. Finally, the
34
learned features from the multi-task CNN and DenseNet models are combined to classify
10
disease. The proposed framework can not only output the disease status, but also provide the
30
hippocampal segmentation result. No tissue segmentation and nonlinear registration are required
for MR image processing. The experimental results based on the ADNI dataset have
demonstrated that deep CNN approach has achieved promising performance for AD and MCI
diagnosis.

83
 Chapter - 4
Fuzzy Neural Network based Classification of Hippocampus on
Alzheimer’s Disease and Feature Extraction with Discrete Wavelet
Transform in MRI Images

4.1 INTRODUCTION
A neurodegenerative malady is Alzheimer ailment. The recognizable proof of those
infections is amnesia and other mental capacities during which it declines, as neurons kick the
bucket and distinctive region of mind. After determination, it presents span of 10 years on the
seriousness of the illness in changing [66]. Inside the world the chief regular clarification for
dementia is Alzheimer. The reason and movement of those infections is not surely known. Its
related plaques and tangles inside the cerebrum. This prompts to the gross decay of the
influenced regions are degeneration inside the flap, parietal projection, portions of the frontal
territory and cingulated gyrus. It is a dynamic issue. It makes the synapses die and bite the dust.
Overlooking ongoing occasions or discussions are the early indications of the illness [67]. The
incidentally improve side effects or moderate the speed of decrease is drugs. No treatment fixes
Alzheimer’s or adjusts the ailment procedure in the cerebrum. The Alzheimer‘s illness influences
dispositions and practices. The tissues are Depression, Apathy, Social withdrawal, Mood swing,
Distrust in others, Irritability and forcefulness, Changes in resting propensities, Wandering, Loss
of restraints, Delusions [68].
A mix of hereditary, way of life and natural factors that influence the cerebrum after
25
some time may be a reason for Alzheimer’s. The way of life and in this way the heart soundness
of the sickness are Lack of activity, Obesity, Smoking or presentation to secondhand smoke,
High essential sign, High cholesterol, and Poorly controlled diabetes. It is anything but a
preventable condition. Exercise normally, Eat an eating regimen of new produce, solid oils and
nourishments low in soaked fat, deal with high indispensable sign, diabetes and elevated
cholesterol are utilized to diminish the danger of Alzheimer ‘s malady [69]. Taking an interest in
get-togethers, perusing, moving, playing table games, making craftsmanship, playing an
instrument and different exercise are additionally used to decrease the layer of those sicknesses.

84
 70
Alzheimer's infection is the commonest neurodegenerative dementia and a developing
sick wellbeing. In this part, the researcher use Magnetic Resonance Image pictures. To make
itemized pictures of the organs and tissues inside the body it utilizes a strong attractive motion
33
and radio waves. It does not include x-beam radiation. It is a simple sweep. In MRI filtering the
claustrophobic sensation is going on. X-ray filtering could be a non-obtrusive methodology. The
30
essential MRI full-body scanner is made by Raymond Damadian. It is a radiology method. A
cylinder encompassed by a goliath round magnet in MRI filter. The magnet makes a vigorous
attractive motion inside the output. An especially precise strategy for ailment discovery all
through the body is that the employments of MRI. It assesses the structure of the guts and aorta,
where it can recognize aneurysms or tears. It is vary from Computerized Tomography outputs
and X-beams. X-ray does not create possibly destructive ionizing radiation. Abnormalities of the
mind and medulla spinalis, tumors, pimples and different inconsistencies in different pieces of
the body, certain sorts of heart issues, ailments of the liver , other stomach organs the MRI
scanner is utilized in these infection. The radiologist will look at the photos to see whether any
more drawn out required after the MRI check.

4.2 RELATED WORKS

In the brain MRI segmentation the foremost popular conventional methods are commonly
used. Utilizing the principal well known traditional techniques the multifaceted nature and
subsequent difficulties of the cerebrum MRI division, time interim and precision are tended to.
To get rid of the unrelated tissues improve the time interval MRI preprocessing steps are
18
involved. The necessity for classification of images is an automatic intelligent classification
careers. The advancement of precision rate and diminish the mistake pace of MRI cerebrum
tumor is that the consequences of robotized smart framework [70].

Golland et. al. proposed a principal component analysis for image size reduction process
with the assistance of feature pool [71]. After finishing of this procedure bolster vector machine
assists with grouping the component separated image. PCA doesn’t assist with anticipating
highlights which are requiring appropriate for division, however just picks highlight with a huge
fluctuation. Thus, these highlights are to be picked by head part examination to give best
45
classification results.

85
 Xing et. al. proposed an aircraft ground icing based classification process with the great
technique of support vector machine. This method helps to predict the aircraft ground icing which
is to be compared with the novel classification technique of BP neural network model [72]. This
strategy gives a superior outcome with following parameters like conjecture precision, high solid
speculation model with little examples and wide application model.
The paper has following capacity like airplane icing type and climate factor, bolster
machine based airplane icing type arrangement model and calculation, strategy for developing
multi class bolster vector machine, fabricating the help vector machine based icing type
characterization model, part type impact on the exhibition of help vector machine
acknowledgment, examination of icing type information test and execution investigation. A
51
human brain is an organ which is located in central portion of human systema nervosum. A
brain has three major parts like cerebrum, brainstem and cerebellum parts. In human brain,
cerebrum is the largest part which is assessed to two cerebral hemispheres. Lashkari et. al.
proposed a tissue detection technique with automatically in brain, which helps to reduce time of
diagnosis and high accuracy. The main finding of this paper is to classify the brain tissues to two
major classes like normal one and abnormal one. In MRI images, these two classifications are to
1
be applied. In this paper, feature selection mechanism helps to decrease the space of features and
neural network classification algorithm is to be applied [73].

4.3 THEORETICAL BACKGROUND

This section explains the theoretical background of the concepts employed in this chapter.

4.3.1 DISCRETE WAVELET TRANSFORM

Disintegration forms are regularly iterated deteriorating progressively the approximations

thus, with the goal that the sign is separated into different degrees of goals. If there should be an
occurrence of pictures, the DWT is applied to each measurement independently breaking down
a picture into four sub-groups which are low-low, low-high, high-high and high- low [74].
14
where the low-low sub-band can be viewed as the estimation part and it is utilized for the
following degree of the 2D-DWT, in the interim the other sub-groups would be viewed as
39
the nitty gritty segment of the picture are appeared in the figure 4.1.

86
 Figure 4.1 DWT Level decomposition in single brain slice.
4.4 METHODOLOGY FOR CLASSIFICATION OF ALZHEIMERS DISEASE IN
HIPPOCAMPUS USING FUZZY NEURAL NETWORK AND FEATURE
EXTRACTION BY DISCRETE WAVELET TRANSFORM
In this segment, the procedure of the exploration paper chiefly represent considerable
authority in Feature Extraction on Discrete Wavelet Transform and consequently the grouping
Alzheimer sickness is finished by utilizing fuzzy neural network [75]. The pre-processing is
71
predicated on ant colony optimization to enhance the detection of suspicious region from MRI
images. The dataset are obtained from MRI data set. The important MRI images are utilized in
this data set are taken from KG scan centre, Madurai. The aim of automated multiclass diagnosis
of Dementia the Fuzzy Neural Network (FNN) method is proposed. Also this proposed in
accordance with the classification of MRI of human brain. The pre-processing is an initial stage
of the work.
To remove noises the pre-processing work is completed. Get rid of the film artifacts like
labels and x-ray marks the preprocessing method is cast off [76]. Using the Ant Colony based
Optimization (ACO) technique the high frequency components are removed. From the pre-
processed MR images of brain the 2D histogram signal is obtained. Utilizing Discrete Wavelet
Transform (DWT) the further component extraction is finished. The primary few DWT
coefficients as features could be a proposed method use. Using FNN for classification might be
a uses of proposed method. To characterize highlights into three classes like Alzheimer’s,
28
Mellow Alzheimer‘s and Huntington’s sickness is that the purpose of training a Fuzzy Neural
Network (FNN). The figure 4.2 shows the proposed flow analysis on this chapter.

87
 MRI Images MRI Image
Database

Pre-Processing

X- ray marks removal using Tracking

Pre-Processed Data

2D Histogram Signal

MRI Pre-processed Image

Representation of MRI images in Histogram

Segmentation
Segmentation of hippocampus using Region based
Active Contour and ACO for noise removal

Feature Extraction
MRI Wavelet Co- Results of
efficient Feature
Images Extraction

Classification using Fuzzy Neural Network (FNN)

Alzheimer’s, Mild Alzheimer’s and Huntington’s

Figure 4.2 Proposed work of feature extraction using DWT and classification by FNN

88
4.4.1 Pre-Processing based on Ant Colony Optimization
12
Pre-processing techniques is used to improve the detection of the suspicious region from
Magnetic Resonance Image (MRI). The pre-processing method consists of two steps; first the
removal of film artifacts such as labels and X-ray marks are removed from the MRI using
tracking algorithm. Second, the removal of high frequency components used Ant Colony
Optimization (ACO) technique. It gives high resolution MRI compare than median filter,
6
Adaptive filter and spatial filter. The performance of the proposed method is also evaluated by
means of peak single-to noise-ratio (PSNR), Average Signal-to-Noise Ratio (ASNR). The
schematic flow chart for pre-processing is shown in figure 4.3.

Fig 4.3 Flow chart for Pre-Processing based on ACO

The first step of pre-processing is tracking functions, which involve the operations that
are generally essential before the major data examination and, extraction are commonly grouped
as radiometric or geometric improvements. MRI includes film artifacts or label for instance,
patient name, age and marks. Film artifacts are eliminated by means of tracking algorithm. At
12
this point, beginning from the primary row and initial column, the intensity value of the pixels is
18
analyzed, consequently the threshold value of the film artifacts is found. It is to be noted that the

89
threshold value is extremely more than that of the threshold value is eliminated from MRI. The
high intensity value of film artifacts are eliminated from MRI. The pre-processed image are
shown in figure 4.4 respectively.

Fig 4.4 Preprocessing image – First row (Original image) Second row (Alzheimers Disease)

4.4.2 Tracking algorithm for removal of artifacts

Step 1: It is during a two dimensional matrix to rework and examine the MRI.
Step 2: The aim of eliminating white labels is completed by choosing the height threshold
value.
Step 3: Keep flag value as 255.
Step 4: Intensity value of pixel is adequate to 255.
Step 5: As a result from MRI the labels are eliminated, the flag value is fixed to zero,
when the intensity value is 255 subsequently.
Step 6: If not skip to the next pixels.
To expel the high recurrence segments of commotion, the Ant Colony Optimization
(ACO) strategy is utilizing inside the preprocessing method. The nearby variety of the pictures
power esteems the development of ants is coordinated. After the ACO, when it further runs for N

6
90
number of emphases is pheromone lattice is made. Development procedure and update process
are the iterative procedure. The decision procedure is last during which recurrence commotion is
chosen [77].
10
The majority of complications through guide search of the answer space optimized by
ACO. Fundamental activities of ants are supported on ACO algorithm. Through a graph ACO
met heuristic engages solution is construction [78]. Until they attain a comprehensive solution
tons of ants move through the answer space adding together solution constituents to partial
65
solution. It depends on the pheromone substance of the paths and a heuristics assessment. Means
of a probabilistic action selection rule a specific ant k chooses the next node in each stage of the
development. The probability with which ant k will plan to move from current node i to next
node j:
| τij| α| ηij| β
pkij= If jɛNik (4.1)
∑ lɛNki | τij| α| ηij| β

Here hub i to hub j the pheromone substance of the curve showed. On hub i represent the
neighborhood for a selected ant k provided. The precise ant k the neighborhood only comprises
nodes that have not been visited. After the whole neighbors o the present node become available
45
for visit when all possible nodes are visited. The influence of pheromone content and heuristic is
indicating the constants α and ß. At last, going from node i and node j the ηij indicates the
heuristic details. The measure of the value of extending the present partial solution is heuristic
44
information [79]. Proportionate to the standard of the answer the high frequency of pheromone is
deposited. On the arcs the ants deposit pheromone as given:

τnew
ij
m Δ k
= τcijurrent+ Σk=1 τij (4.2)

Where from hub i to hub j show the quantity of pheromone and k will put into the arc
travel. The general number of ants is indicated by m. On all arc the whole arcs within the
equivalent path also are executed given:

τij = (1 ― p)τij 0 < p ≤ 1 (4.3)

The high frequency of noise is taken by the pheromone values. From the Resonance Imaging the
noises are eliminated.

91
4.4.3 2D Histogram
The 2D Histogram signals are represent the pre-processed Magnetic Resonance Image.
18
The relative frequency is amount of gray levels inside an image indicated by the histogram of
image. Modifying a picture, the histogram modeling schemes assist. Histogram has a desired
shape [80]. With a narrow histogram, the low contrast levels of an image in stretching, there’s
12
way the contrast enhancement is realized. The target is to urge hold of a uniform histogram for
the yield picture is that the instance of Histogram Equalization (HE). Overall contrast is
perceived the intention than an “optimal”.
4.4.4 Multi Channel Images
The expansion of the idea of picture histogram to multi-channel pictures is
multidimensional histogram. During the specific among the two pictures the picture handling a
2D histogram delineates the relationship of powers. In multi channel images the 2D histogram is
typically employed for the purpose of comparing two channels. The powers of first channel are
x-hub. The powers of second channel are y-hub. Nothing above tallying the quantity of voxels
with a particular force happens in the picture is 1D histogram. The bins partitioned the intensity
limit of the image. When the intensity is incorporated inside the range the bin represents a voxels
subsequently belongs to the bin. The 1D and 2D histogram it calculates the rate of the
5
combinations of intensities. The image require to be identical in size is the working out of a 2D
histogram [81].
4.4.5 Segmentation of Hippocampus Using a Region-Based Active Contour
8
A region-based active contour model is based on intensity information in local regions that must
be initialized with the approximate location of the hippocampus, using an initial bounding box
26
based on the approximate size of the hippocampus. In the next step, the local intensity fitting
energy in terms of a contour and two fitting functions that locally approximate the image
intensities on the two sides of the contour. For a given point x ϵ Ω

εFit(C, f1 (x), f2 (x)) = ∑2i=1 λi ∫ K (x - y)| I (y) - fi (x)|2 dy (4.4)

64
Where I: Ω→R is a vector valued image, C is a closed contour in the image domain Ω which
separates Ω into two regions: Ω1 outside (C), Ω2 = inside (C), 휆1 and 휆2 are positive constants,
f1 (x) and f2 (x) are two values that approximate image intensities in Ω1 and Ω2 , the intensities

92
2
I(y) that are effectively involved in the above fitting energy are in a local region centered at the
37
point x and K is a Gaussian kernel function with a scale parameter 휎 > 0.
1 2 /2σ2
Kσ (x - y) = (2 π)π /2 ση
e-| x-y| ) (4.5)

To obtain the object boundary, find a contour C that minimizes the gy 휀퐹푖푡 for all x in the
domain. In addition it is necessary to smooth the contour by penalizing its length. Thus, the
following energy functional:

휀 (퐶, 푓1 (푥), 푓2 (푥)) = ∫ 휀퐹푖푡 (퐶, 푓1 (푥), 푓2 (푥)) 푑푥 + 휐 |퐶| (4.6)

23
This energy is then incorporated into a formulation to derive curve evolution equation for
27
energy minimization. Due to a kernel function in the data fitting term, intensity information in
local regions is extracted to guide the motion of the contour, which thereby enables our model to
cope with intensity in homogeneity. Thus, the energy can be written as

휀휖 (휙 , 푓1, 푓2) = ∑2푖=1 휆푖 ∫( ∫ 퐾휎 (푥 - 푦)| 퐼 (푦) - 푓푖 (푥)|2 푀푖휖 (휙 (푦)) 푑푦 )푑푥

+ 휐 ∫ |∇ 퐻∈ ( 휙 (푥))| 푑푥
(4.7)

Where
1
퐻휖 (푥) = 2 [1 + 2
휋
tan-1 ( )
푥
휖
] (4.8)

For accurate computation and stable level evolution, the regularization term as;

풫 (휙 ) = ∫ 1 ( |∇휙 (푥)| - 1)2 푑푥 (4.9)

2

Therefore, by keeping f1and f2 to minimize the energy functional respect to 휙 as follows;

F (휙 , 푓1, 푓2 + 휇 풫 ( 휙 ) (4.10)

4.4.6 Feature Extraction

28
The element extraction is that the change of a picture into its arrangement of highlights.
For grouping reason from the picture the valuable highlights of the pictures are extricated [82].
To remove great list of capabilities for characterization could be a difficult assignment. During
this paper, include extraction utilized discrete wavelet change strategy. At each degree of

93
wavelet disintegration to require on the highlights of X-ray cerebrum pictures inside the kind of
vector guess is the point of the component extraction process.
It comprises of three phases:
1. Using wavelet transformation in image disintegration.
2. Wavelet sorts will be utilized are Haar to level. The wavelet disintegration
process is controlled.
3. Utilizing the discrete wavelet change the yield coefficient vector is determined.
Various sub-groups are created by applying discrete wavelet change. X (t) is the nonstop
wavelet change of a sign, ψ(t) is that the square integrable capacity and near to a genuine
esteemed wavelet is given as:
∞
wΨ(a,b)=∫―∞ x(t) ∗ Ψa,b(t)d x (4.11)

Where, the numerous discrete wavelets are ψ by methods of interpretation and expansion
10
process. The enlargement factor is a and the translation parameter is b. Restraining of mean a, b
to discrete cross section (a=2b, a ϵ R+, b ϵ R) to flexibly the DWT. The discrete wavelet ψ meets
4
the limitation of getting zero mean. DWT is a linear transformation. The identical length of the
DWT is converting into a numerically a unique vector. With the Haar wavelet function, wavelet
coefficients are figured for the LL sub band. The multistage portrayal of capacity could be a
significant successful of DWT [83, 84]. Feature extraction using DWT and Wavelet
decomposition is shown in figure 4.5 and 4.6.

50
Figure 4.5 Representation of Feature Extraction using discrete Wavelet

94
 Fig 4.6 Wavelet decomposition

4.4.7 Fuzzy Neural Network Classification

The classification method is utilized to the extraction of highlight sets. MRI mind picture
classification the Fuzzy Neural Network is pushed off [85]. In view of three classifications, the
MR mind pictures give classification result. The three classes are Alzheimer’s illness, Gentle
44
Alzheimer‘s sickness and Huntington’s ailment. Accuracy, sensitivity and specificity are the
performance of the classification method evaluated. The FNN is executed between two classes.
FNN works every class grouping of the similar manner [86]. During the ultimate output node, it
is to settle on a maximum value winner.

Prototype feature vectors are the features within the input. Presently inside the
preparation model information{x (q), t (q): q=1… Q} there are two classes. The two distinctive
have labels having t (q). During which every such node an exemplar feature vector that features a
connected label indicates a Gaussian function centered. The category 1 took the initial group of
5
hidden nodes. In class р (here it is taken as p=1, 2), there could be a enormous number of Kp of
feature vectors. With the identical label is eradicated the feature vectors are almost a special
feature vector. The results of Gaussians (nodes) and therefore the amount of centers indicate

95
each class p. The Gaussian Fuzzy set Membership Function (FSMF) centered on x (q) [87]. The
FSMF capacity is:

X →g (x; x (q)) = exp {− || x−x (q) ||2 / (2σ2)} (4.12)

Among the whole exemplar pairs is one half of the typical is taken into account as σ.

4.4.7.1 Fuzzy Neural Network Algorithm

Provision of the step below the high level algorithm is nearly clear cut. With a view to
proceed with the program it is uncomplicated. In the exemplar feature vector the training and
learning reside [88]. The aim of coaching their labels does not require expanding computation
time.

The stages are:

Stage 1: Get the information document (N the measure of highlights, Q the quantity of
highlight vectors, J the measurement of the names, K the quantity of classes, the
entire Q highlights vector and the whole Q marks).
Stage 2: over the entire element vector sets, find negligible separation Dmin.
Put σ =Dmin/2
Beginning no. of Gaussian focuses - insert G=Q
Stage 3: Find two model vectors of min.
k1 and k2 are showing the space.
On the off chance that d< (1/2) Dmin - if vector are close
In the event that mark [k1] = name [k2] - same name
Kill Gaussian k2
Decrease the number of Gaussians - G=G-1
Go to stage 2
Stage 4: Contribution to next obscure x to SFNN to be surveyed
For k=1 to G do
Process g[k] =exp {-||x-x (k) ||2/ (2σ2)}
Discover most extreme g [k*], ober k=1… G
Yield x, mark [k*] - name [k*] is class of x
Stage 5: if all sources of info ordering are done, stop else, go to stage 3.

96
4.5 EXPERIMENTAL OUTCOMES

In this segment, the performance of proposed Discrete Wavelet Transform is directed and
4
assessed with the Fuzzy Neural Network classifier is evaluated. On the mat lab the experiments
25
are regulated. T2-weighted Magnetic Resonance Image dataset are in incorporates inside the
axial plane where the resolution is 256 × 256 in plane resolution. From Kg scan center and
outsource of OASIS dataset and ADNI dataset is obtained. In comparison against the T1 and
PET modalities the T2 pictures are of higher complexity and better vision. Each classification of
1
cerebrum, a total of 20 pictures was arbitrarily picked. Inside the dataset there are one
classification of typical mind and seven classification of unusual cerebrum. Comprising of 20
typical the 160 images are picked. 7category of illness 20 pictures / malady unusual cerebrum
pictures =140. The strange cerebrum MRI dataset have following ailments like Alzheimer's
ailment, Alzheimer's illness in addition to agnosia. The sample image of each and every ailment
is appeared in figure 4.7.

Figure 4.7 Sample picture of each and every ailment from dataset [A. Normal barian MR images;
B . Alzheimers Disease; C. Alzheimer’s disease plus visual agnosia]

56
The result of pre-processing method based on film artifacts removal. While acquiring the
MRI scan, to identify the patient and scanned images, basic information about the patients such
as name, age, gender, date of MRI shot, hospital name and similar other things are printed on
MRI images. By removing the labels from the MRI images, the images were anonymized
23
visually. Also, the text labels were not considered in the research work. To remove the text
artifacts from MRI, the tracking algorithm was used. The initial step of tracking algorithm, the
brain MRI images were read and stored into a two-dimensional matrix.

97
 After checking the pixel intensity of different image pixels in different parts of image a
threshold value was chosen. The artifacts appear in four corners of the MRI images. The pixel
images were scanned from the top left corner to right using raster scan. Pixels which were having
an intensity value above 255 were considered as artifacts, the figure 4.8 shows MRI with film
artifacts with a resultant MRI.

b
Fig 4.8 Pre-processing of film artifacts a) MRI images with Film Artifacts b) MRI images with
removed Film Artifacts using Tracking

8
4.5.2 Result of hippocampus segmentation in Region based Active Contour

The proposed method for hippocampus segmentation has been applied after
preprocessing. The initial conditions for the Region based Active Contour algorithm are as
follows:
Iteration time step Δ t = 0.1, iteration step is 300 times and the parameters of initial level
function are: C0 = 2 , σ = 3 , μ = 0.004 * 255 * 255 and λ1 = λ2 = 1. Figure 4.9 in second rows
shows the detected boundary of the hippocampus in sagittal images (red curve).

98
Fig. 4.9 First row: three slices from one MRI scans, second row: segmentation of the
hippocampus result

These results were obtained after pre-processing and this method is evaluated in terms of
35
objective results which are obvious based on the hippocampus shape and ground truth. Figure
4.10 illustrates the evaluation of these results.

25
Fig 4.10 First row: segmentation of the hippocampus results for three slices of one MRI scans,
second row: ground truth

99
26
In order to apply this method for different AD patients, five slices from 5 MRI scans have been
28
chosen which are preprocessed to show the object better and then the region is manually
determined as bounding box to implement the methodology. Figure 4.11 shows these results and
each image are compared with its ground truth.

Original MRI
images

Bounding
Box

Images
without Pre-
Processing

Applying
Region based
Active
Contour

Ground Truth

Fig 4.11 Results of each images compared with Ground Truth

4.5.3 Result of Ant Colony Optimization for Noise Removal

71
The experimental outcome of Ant Colony Optimization shows the removal of high
frequency. This gives a high resolution MRI, compare with median filter, adaptive filter and
spatial filter. To assess the performance of ACO procedure, the implementation is resulted with
an original image of T1 and T2 weighted image [89]. The images are 256 × 256, 8 bit gray scale
51
with intensity values scaled in the range from 0 to 255. The figure 4.12 shows the original MR
23
brain images with T1 and T2 weighted images of skull stripped in hippocampus.

100
 a b
Fig 4.12 The original MR brain images. (a) skull-stripped T1- weighted image (b) skull-stripped
T2-weighted image.

26
The process of ACO were shown in different iterations, according to the iteration times
the white matter is segmented after two hundred iterations for T1 – weighted image and the gray
matter is segmented after two hundred iterations for T2 – weighted image. The figure 4.13 shows
45
the segmentation results using the ACO algorithm for different iterations.

10 iterations

101
 50 iterations

100 iterations

200 iterations
Figure 4.13 The segmentation results using the ACO algorithm for different iterations. In each
37
case, the first column shows the results of the T1 image and the second column shows the results
of the T2 image.

102
 6
The result for the comparison was tested using Meta – heuristic segmentation methods such as
genetic algorithm and old ant colony optimization algorithm. The proposed method is more
26
exactly than the other method. The figure 4.14 shows comparison of proposed method with other
methods.

a b
41
Figure 4.14 Comparing the proposed method with other methods. (a) T1-weighted image, (b) T2-
weighted image. In each case, the first row shows the genetic algorithm, the second row shows
the old ant colony algorithm, and the third row shows the proposed method.

103
4.5.4 Performance analysis for resultant images
The Image Quality Measurements (IQM) is important to research the performance of
image processing techniques like contrast enhancement, median filter as de-noising etc. IQM is
finished to assess the exhibition bolstered the provisions of value.
4.5.4.1 Quality Measures supported enhanced images
The resultant enhanced images are compared with different quality parameter are
pinnacle signal clamor proportion or peak signal noise ratio and mean square blunder or mean
square error.
4.5.4.2 Peak Signal Noise Ratio (PSNR)
65
Pinnacle signal clamor proportion is between the most extreme potential to the facility of
ruining commotion. An impact on the fidelity present on its portrayal. It is the logarithmic
capacity. The logarithmic capacity of pinnacle estimation of picture and mean square mistake are
referenced.
PSNR = 10log10(MAXi2) / MSE (4.13)

4.5.4.3Mean Square Error (MSE)

The genuine estimation of t amount being assessed is MSE. To measure the distinction
between estimators is MSE.
1
MSE = mn
∑m―1
i=0
∑n―1
j=0 [ I (i,j) ― K (i,j)]
2 (4.14)

4.5.5 Quality measures of enhance image error compare with other classifier

4.5.5.1 Peak Signal Noise Ratio

To assess the nature of preprocessing in MR mind picture is contrasted with unique ones.
To line, the standard of picture has the stature signal clamor proportion is determined. The
performance of the proposed DWT with FNN classifier is compared with an existing proposed
work in DCT with ANN classifier and SVM classifier. The peak signal noise ratio results
compared with DWT with FNN is high when compare to the prevailing methods of DCT with
ANN and SVM classifier. The comparison data with other classifier with an enhanced image
with error is shown in table 4.1 and therefore the experimental result of PSNR is shown in
figure 4.15.

104
 Table 4.1 Enhanced image with error in PSNR with other classifier
S. No Comparison with other classifier Enhanced image with error in PSNR

1 DWT with FNN classifier 45

2 DCT with ANN classifier 40
3 SVM classifier 37

50

40
DWT with FNN
30 classifier
DCT with ANN
20 classifier
SVM classifier
10

Classification Methods
Fig 4.15 Experimental results of PSNR comparison

4.5.5.2 Mean Square Error

The performance of mean square blunder is proposed on DWT with FNN classifier which
is compared with an existing proposed work of DCT with ANN classifier and SVM classifier.
Therefore, the MSE results of proposed work of DWT with FNN classifier have less MSE error.
The comparison data with other classifier with an enhanced image with error is shown in table
4.2 and therefore the experimental result of MSE is shown in figure 4.16.

Table 4.2 Enhanced Image with an Error in MSE with other classifier
S. No Comparison with other classifier Enhanced image with error in MSE

1 DWT with FNN classifier 0.65

2 DCT with ANN classifier 0.70
3 SVM classifier 0.74

105
 0.76
0.74
0.72
0.7
DWT with FNN classifier
0.68
DCT with ANN classifier
0.66
0.64 SVM classifier
0.62
0.6

Classification Methods
Fig 4.16 Experimental result of MSE comparison

4.5.6 Quality measures on comparison of volume with classifier

4.5.6.1 Accuracy Comparison
Better accuracy rate is produced by proposed Discrete Wavelet Transform with Fuzzy
Neural Network classifier. The existing Discrete Cosine Transform with Artificial Neural
Network classifier and Support Vector Machine classifier gives much greater accuracy results.
The accuracy result is increased when the number of images increases. In comparison to existing
system this produces high accuracy. The numbers of images are administered for 100 images to
compare the accuracy. The number of images versus Accuracy of their comparison is shown in
table 4.3 and the result analysis of Accuracy comparison is shown in figure 4.17

Table 4.3 Number of Image versus Accuracy

No of Images SVM Classifier DCT with ANN Classifier DWT with FNN Classifier
20 88 90 95
40 88.3 90.5 95.6
60 88.5 91.5 95.7
80 88.8 92 95.8
100 90 93 95.9

106
 100

95
SVM Classifier

90
DCT with ANN
85 Classifier
80
20 40 60 80 100
DWT with FNN
Classifier
No of Images

Figure 4.17 Accuracy versus No of Images with different classifier

4.5.6.2 Sensitivity Comparison:
It produces high sensitivity. Existing DCT with ANN classifier and SVM classifier this
is often much greater accuracy results. The sensitivity of the result increases when the number
of images increases. In comparison to existing system it produce high sensitivity rate. The
numbers of images are carried out for 100 images to match the sensitivity. The number of images
versus sensitivity of their comparison is appeared in table 4.4 therefore the result of research
sensitivity comparison is appeared in the figure 4.18.
Table 4.4 Number of images versus Sensitivity
No of images SVM classifier DCT with ANN classifier DWT with FNN classifier
20 78 80 86.5
40 78.5 80.5 87
60 79 81 87.5
80 79.5 81.5 87.8
100 80 82 88

90

85 SVM classifier

80
DCT with ANN
75 classifier

70 DWT with FNN

20 40 60 80 100 classifier
No of Images

Figure 4.18 Sensitivity versus No of Images with different classifier

107
4.5.6.3 Specificity Comparison:
The FNN classifier produces high specificity. For existing DCT with ANN classifier
and SVM classifier is often higher. The specificity of the results increases when the number of
feature increases. In comparison to existing system it produce effective specificity rate. The
numbers of images are carried out for 100 images to match the specificity. The number of
images versus specificity of their comparison is shown in table 4.5 and therefore the result
analysis of specificity comparison is shown in figure 4.19
Table 4.5 Number of Image versus Specificity
No of Images SVM classifier DCT with ANN classifier DWT with FNN classifier
20 75 79 85
40 75.5 79.4 85.6
60 75.7 79.6 85.7
80 75.9 79.8 86.8
100 76 80 88

90

85
SVM classifier
80

75 DCT with ANN

classifier
70
DWT with FNN
65 classifier
20 40 60 80 100
No of Images

Figure 4.19 Specificity versus No of Images with different classifier

4.5.7 Comparison of Enhanced Image with Overall Classification Method

The current Discrete Cosine Transform (DCT) with Artificial Neural Network (ANN)
and Support Vector Machine (SVM) classifier are contrast with the proposed Discrete Wavelet
Transform with Fuzzy Neural Network classifier [90]. Alzheimer’s infection, Mellow
Alzheimer’s malady, Huntington’s illness and ordinary are four classes of the grouping results.
Positives the category and negative the category is recognized.

108
 1) True Positives (TP): predicts class as class.
2) True Negatives (TN): predicts non-class as class.
3) False Positives (FP): predicts non-class as non-class.
4) False Negatives (FN): predicts class as non-class.
The following standard equations are obtained:
TN
Specificity = TN + FP
,
(4.15)
TP
Sensitivity = TP + FN
, (4.16)
TP + TN
Accuracy = TP + TN + FN + FP (4.17)
The performance analysis gives better result supported following parameters like
accuracy specificity, specificity and sensitivity which is representing on table 4.6 and graphical
analysis are shown on the figure 4.20.
Table 4.6 Performance parameters analysis
Parameters SVM classifier DCT with ANN classifier DWT with FNN classifier
Accuracy 88.5 91.5 94.5
Specificity 74 79 84
Sensitivity 80 81 87

Analysis of Accuracy, Specificity and Sensitivity

100

80

60

40

20

0
DCT with ANN DWT with FNN
SVM classifier
classifier classifier
Accuracy 88.5 91.5 94.5
Specificity 74 79 84
Sensitivity 80 81 87

109
 Figure 4.20 Parameter Analysis of Accuracy, Specificity and Sensitivity
4.6 SUMMARY
In this research the MRI for highlight extraction is utilizing discrete wavelet change in
grouping of dementia. For Alzheimer‘s sickness diagnosis an easy and robust grouping approach
of MRI scans is proposed. Automated multiclass diagnosis of Dementia the FNN method is cast
off. The FNN method utilized in classification supported human brain MR images. Within the
initial stage, to get rid of noise, the pre-processing work is completed. Remove the high
frequency components using the Ant Colony based Optimization ACO technique. From the pre-
processed MR images of brain the 2D Histogram signal is obtained. Using DWT the further
feature extraction is completed. The Alzheimer’s, Mellow Alzheimer’s and Huntington’s ailment
are the goals of training a FNN, so as to classify the features into classes. Compared to other
volumetric method the proposed method provides better classification accuracy.
In this technique the MRI picture is inserting into the MRI picture database. The MRI
Image database collects the info of this MRI image. The proposed Fuzzy Neural Network
method is used for multiclass diagnosis of Dementia and the classification is based on human
brain images. The primary step is pre-processing; it is done to remove the noises. The Ant
Colony Optimization technique is used to remove the frequency components. 2D Histogram
signal is used to pre-processed the Magnetic Resonance Image to enhance the detection of the
suspicious region.. A definitive advance arrangement used in Fuzzy Neural System is utilized for
MR cerebrum picture order and the characterization result bolstered Alzheimer’s ailment,
Mellow Alzheimer’s, and Huntington’s malady. The order of execution is assessed by exactness,
affectability and explicitness.

110
 Chapter - 5
Feature Ranking based Classification of Hippocampus on Alzheimers
Disease From Structural MRI

5.1 INTRODUCTION

In human brain, cerebrum is the largest part which is assessed into two cerebral
hemispheres. Each and every hemisphere is assessed into four lobes which are denoted as frontal
flap, parietal flap, fleeting projection and occipital flaps. In human brain, a cerebrum is
integrated by brainstem to medulla spinalis which is midbrain and medulla oblongata. This
research helps to introduce an excellent objective of graphical interface application which helps
to process on brain and also morphological disorders. Alzheimer’s malady is a mind dynamic
issue which causes to lessen thinking velocity and memory power [91].
A human brain is an organ which is found in central portion of human systema nervosum.
A brain has three major parts like cerebrum, brainstem and cerebellum parts. Within the
Alzheimer’s disease, hippocampus region volume and shape helps to work out the first detection
analysis [92]. An early location of Alzheimer’s sickness, intra neuronal filamentous are to be
deposited and also neuro-fibrillary tangles are to be stored within the neurons. The adulthood
people and children whom they were affected to Alzheimer’s disease faces most critical
abnormality problem in cerebral mantle region. In the early stage, Alzheimer’s sickness doesn’t
refine the hippocampus region or cortex area [93].
Medical imaging has new horizons in analyzing and subtracts signifying the brain
structure and its functions. Utilizing x-ray gigantic advancement in getting to cerebrum wounds
and investigating mind life systems has been made [94]. By utilizing the scanner on tissue related
information namely gray level or intensity the first features of MRI cerebrum pictures are
translated. In image processing technique one among the main considerations within the field of
visual perception to spot the presence of objects or tumor in a picture is image segmentation.
Divide the image into meaningful segments which correspond to significant objects present
within the image is image segmentation main motivation. Assigning similar pixel of a picture

111
supported the similarity criteria, such the pixels with an equivalent properties combined to make
a significant area of the picture is the image segmentation process.
5.2 RELATED WORKS
A. Javed et. al. proposed a micro Doppler signature classification from ground target
using support vector machine algorithm. The NR-V3 ground surveillance radar is to be
performed in micro Doppler signature classifier processing unit. During this method, the features
are to be extracted with the assistance of Fast Fourier Transform spectrum [95, 96]. The featured
parameters are to be assessed supported on support vector machine, polynomial kernel operation
and radial basis functional mechanism. This system gives high accuracy than other mechanisms.
The paper has following functions like classification scheme, analysis of signature databases like
pedestrian class, vehicle class and no target class, pre-processing operation, feature extraction
analysis and classification process. It breaks down the size and shape of hippocampus,
utilizations of hippocampus volume and LDDMM separation for location [97, 98]. It doesn’t
demonstrate the right arrangement of dementia. It shows that hippocampus volume alone.
To register the pair-wise enlistment between x-rays a Symmetric Log Domain
Diffeomorphic Demos Algorithm (SLDDDA) is first employments. To figure the Riemannian
separation between them it utilizes the twisting field from diffemorphism mapping. Upheld on
Riemannian separation grid the ghastly implanting calculation is performed. In the cerebrum x-
ray division the most well known ordinary strategies are normally utilized. Utilizing the most
famous customary techniques the multifaceted nature and ensuing difficulties of the cerebrum
x-ray division, time interim and precision are tended to. To get rid of the unrelated tissues
improve the processing time MRI pre-processing steps are involved. Calculate within the
processing pixels / voxels, elements, respectively, and their 2D neighbourhoods over which
image features are the most difference between 2D image segmentation. To 3D space and 2D
space image segmentation methods are often extended [99, 100].
Lashkari et. al. proposed a tumour detection technique with automatically in brain, which
helps to reduce time of diagnosis and high accuracy. The most initiative of this paper is to group
the brain tissues in to two major classes like normal one and abnormal one. In MRI images,
these two classifications are to be applied. During this paper, feature selection mechanism helps
to decrease the space of features and neural network classification algorithm is to be
applied [101].

112
5.3 THEORETICAL BACKGROUND

MR images are to be processed, supported all mechanisms with perfect subjects with the
assistance of 1.5 – T system placed in GE Medical Systems present in United States of America.
If a high resolution scans are to be needed with the assistance of coronal 3 dimensional speed
spoiled echo signals based on echo time, angle in flip, matrix format like 256 x 192 using a fixed
standard coil head [102, 103, 104, 105]. The threshold-based technique based on a certain
threshold value most simple to segment the objects from background. The threshold technique
pays no attention to the spatial assistance of an image. Hinged on the sting information, objects
and directions pixel is based on gradient operator in a foot hold detection approach technique.

The regions of an image are determined directly; by comparing one pixel to its
neighboring pixel is within the region based division. The region based division achieved either
by region growing and merging, or by splitting and merging the similar region. It utilizes
attraction, radio waves and a pc to gracefully pictures of body structures. It is easy and does not
include x-beam radiation. In x-ray patients with heart pacemakers, metal chips or chips in or
around the eyes cannot be scanned [106, 107, 108]. In MRI scanning claustrophobic sensation
can occur. It uses strong magnetic flux. It is a non-invasive procedure. It varies from CT sweep
and X-beams. It does not utilize possibly hurtful ionizing radiation. It is done on different parts
of your body. Watching at soft tissues and therefore the systema nervosum this scan is
particularly useful. It helps a doctor diagnose a disease or injury.
Image classification is a crucial process which is in a position to separate an image set up
on visual content [109]. It is a process of mixing different pixels into predicted image with high
visuality. During this classification process, a gaggle of knowledge is to be separated into
different classes or different categories. This grouping examination is to be isolated into two
distinct classifications like regulated learning and solo learning strategy. The administered
learning category has two major types like classification model and regression model. Within the
classification model, support vector machine and random forest algorithm is to be presented. In
regression model, logistic model and rectilinear regression model is to be presented.

113
5.3.1 FILTERING PROCESS

In the filtering process, an excellent filtering technique like median filter is to be used
which could be a nonlinear filter which provides a most powerful filtered image. This strategy
assists with taking out the salt and pepper commotion present in a preparing picture. In this
filtering process, there are major two intrinsic properties are to be proposed like preservation of
edge and attenuation coefficient process [110, 111]. This filtering process is to be applied after
completion of pre-processing stage in a picture. The middle channel works in pixel by pixel in a
picture by supplanting each an incentive with the middle of neighboring pixels. The middle
channel precedes the improved image with size of mask. In this, the pattern of neighbor pixel in
seeded growing segmentation is rightly mean as window. It is determined by arranging all pixel
esteems from window into a numerical request supplanting the pixel is being considered as
Middle Median Pixel Values.

5.3.2 PRE-HANDLING

Pre-handling is that the significant activity of picture handling of both info and yield are
force pictures. The most goal of the preprocessing assists with strengthening the picture
information which are stifled from undesirable bends and improve the picture highlights for
additional procedures philosophy. At the absolute bottom degree of deliberation pre-handling
might be a typical name for activities with pictures. During this technique both information and
yield are power pictures. Improvement of picture information that stifles undesirable bends or
upgrades some picture highlights significant for additional preparing is that the point of pre-
handling [112].

As indicated by the components of the pixel neighborhood that is utilized for the count of
most recent pixel brilliance has four classifications of picture pre-processing techniques are Pixel
splendor changes, Geometric changes, Use a zone neighborhood of the prepared pixel and
picture rebuilding that needs information about the entire picture. It is utilized in light of the fact
that impressive excess in image [113].

114
5.3.3 FISHER CRITERION

In this criterion, it helps to work out predicted number of features which is high
discriminative features [114]. In this method, there is different atrophy methodologies are to be
applied which helps to collect the voxels clusters. Voxel groups are related to volume of intrigue
present in the dim scale matter in MRI pictures. The procedure in this section is the information
combination which assists with upgrading the grouping execution drew on volume parameter.
This performance is to be estimated for the most score combination approach and an element
vector link depends on score combination approach [115]. Inside the examination, MRI image
helps to joined different information which is to be recorded underneath.
 Positron Emission Tomography.
 Cognitive Scores.
 Mini Mental State Examination.
These above strategies are utilized to improve the classifier execution. During this
procedure, the most significant number of highlights is to be chosen which is caught with the
assistance of Fisher Criterion Mechanism [116]. This mechanism helps to extend the category
separation analysis between the diseases affected and control subjects. The figure 5.1 shows
Amyloid deposition is usually quite sparse within the hippocampus, compared with neuro-
fibrillary degeneration [117].

Figure 5.1 Amyloid deposition is usually quite sparse within the hippocampus, compared with
neurofibrillary degeneration

115
5.3.4 HIPPOCAMPAL SHAPE

A hippocampus is major part of a brain and humans have two sort of hippocampus
region. The hippocampus is to be presented one in all side of a brain. This is often intended to
visceral brain and it performs vital roles in consolidation process as a result information
transmitted from Short Term Memory (STM) to Long Term Memory (LTM). It is placed below
the cerebral mantle and it has two major interlocking components like Ammon’s horn and
dentate gyrus. In brain, there are two sorts of matter is to be presented like grey matter and
substantia albe. A grey matter is to be formed outside of a brain and substantia albe is to be
formed into inner part of a brain. MR images scans are to be performed with accurate, clearly
manner and also it provides more information on few processing image. It is more powerful for
analyzing blood vessel changes present in white matter formation. Form the loss of brain vessels
or cells present in hippocampus region, the result are going to be provided from 80 percentages
to 90 percentages of special cases of individually affected with Alzheimer's disease [118]. The
figure 5.2 shows MRI techniques uncover changes identified with typical maturing, which
happen in frontal regions and those conspicuous in Alzheimer's ailment, which influence the
hippocampus.

Figure 5.2 MRI techniques uncover changes identified with typical maturing, which happen in
frontal regions and those conspicuous in Alzheimer's ailment.

116
5.4 METHODOLOGY FOR CLASSIFICATION OF HIPPOCAMPUS ON
ALZHEIMERS DISEASE USING SVM CLASSIFIER

In this exploration, the proposed methodology has diverse drawn closer with the
assistance of efficient utilization of different thoughts in the levels. The positioning based
arrangement of Alzheimer’s malady in hippocampus area from auxiliary reverberation imaging
utilizing fisher standard is to be proposed. During this chapter, the statistical features are to be
ranked developed on the ranked approach, which is feature selection process.

5.4.1 PROPOSED WORK FLOW

During this segment, VBM strategy is to be utilized for decide the various gathering
astute separations which are broken down between cross segment basic MRI examines. The MRI
voxels are to be resolved between the segregated Alzheimer’s sickness gathering and HC
gathering. So this approach has elicited from on DARTEL (Diffeomorphic Anatomic
Registration through Exponentiated Lie Algebra) calculation which helps to provide the perfect
accuracy of localization, precise analysis and exact separation of AD regions [119, 120]. In the
above approach and algorithm, the results are going to be analyzed centered on input file of 58
AD patients. During this result, 58 AD patients are compared to be 58 control subjects. The
quantity of interests is to be defined with five alternative ways which are listed below:
 Volume of interest1 has right hippocampus region
 Volume of interest2 has left hippocampus region
 Volume of interest3 has the proper inferior parietal lobule region
 Volume of interest4 has the proper anterior cingulated region
 Volume of interestall has collection of all decay bunch regions.
In this level, the anticipated volume of intrigue is to be examined which assists with
exploring voxels values for creating the crude element vectors. This crude element vector is to be
utilized for information determination process system. The separated highlights are positioned on
premise of t-test estimations of the component acquired from preparing set. t-Test can be
resolved as factual marker for the degree of partition and separation between two gatherings of
Alzheimer’s ailment and Healthy Control in preparing set. Fisher Criterion is proposed to decide
the quantity of highlights and to distinguish the discriminative subset of elite highlights [121,
122, 123].

117
5.4.2 Level of Performance
The proposed new Alzheimers Disease classification system approach is based on a
combination of t-test feature ranking and the fisher criterion for the optimal selection of feature
vectors for high performance of MRI classification of Alzheimers Disease. During this
methodology, the system has five different levels of processing performance. The extent level of
proposed system is represented on figure 5.3.
. In the main stage, the voxel based transform metric in addition to DARTEL
approach is to be utilized for preprocessing 3 dimensional MRI pictures.
. In the subsequent stage, the component extraction process is to be performed
which is relies upon VBM in addition to DARTEL investigation.
. In the third stage, the positioning procedure is performed on separated highlights
which depend on the t-test include extraction preparing instrument in the
preparation information. At that point, Fisher procedure is to apply for examining
and deciding top positioned highlights.
. In the fourth stage, the proposed strategy is to be applied and performances are
assessed with the assistance the bolster vector machine classifier.
. At last stage, the information combination method is to be performed on groups
which are connecting with to extend the classification result.

5.4.3 PREPROCESSING AND STATISTICAL ANALYSIS

The reverberation picture are preprocessing with the assistance of the factual parameter
mapping investigation and along these lines voxel based transform tool compartment which is
handled inside the instrument of MATLAB. The voxel based morph metric is that the fully
automated analysis for assessment of the entire brain structure with the different voxel structure
analysis. This is often developed to the different tissues concentrations which are to be calculated
the volumes between motif groups for various conditions with the dementia [124, 125]. The
VBM toolbox is the benefited for the predicted unique segmentation processing model with the
utmost posterior methodology and the partial volume estimation process which helps to account
for the partial volume effects as a result of more segmentation process of the sub cortical areas.
This helps to improve the signal to noise ration with the assistance spatial constrained on the
linear affine the transformation process and the non linear deformation with the assistance the

118
high dimensional normalization process. The limit veiling assists with expelling conceivable
edge impacts around the outskirt happens between the dark issue, substantia albe and thusly the
spinal fluid [126].

1) Pre-processing 2) Feature Extraction

Original DCM VBM + 3D Mask Feature Extraction Voxel values as

volumes DARTEL based on VBM raw feature
analysis analysis vector

GM Volumes

5) Classification 4) Data Fusion 3) Feature Selection

Technique

Classification & Validation Data fusion among Voxel t-test feature ranking
Clusters + Fisher Criterion

Figure 5.3 Flow chart of proposed systems for classification of Alzheimer’s disease

5.4.4 FEATURE EXTRACTION PROCESS

The component extraction handling methodology is to be applied on volume of interests

which is relies upon VBM in addition to DARTEL examination. This analysis gives a good
performance like reducing gray matter volume in brain regions. Then the anticipated region is to
be segmented based on Mars Bar ROI approach which precise to generating predicted binary
masks. Each and every mask is applied to process of VBM plus DARTEL analysis. This three
dimensional mask is processed on smoothed gray matter which is to extract the values of voxels.

119
5.4.5 FEATURE SELECTION

In the high dimensional vector space, the features vectors are span is extremely small
region. The feature selection mechanism is that the most desired post processing process. This is
often to be considered within the sort of the quality dimensional reduction process with the
assistance of the quality method like principal component analysis. This is often to be performed
which helps to predict the high discriminative subset present in the raw feature vector. This is the
blend of the t-test highlight positioning procedure and the Fisher Rule which helps to scale back
the dimensionality and also increase the discriminability.

5.4.6 PRINCIPAL COMPONENT ANALYSIS OF DIMENSIONALITY REDUCTION

The central part examination is the most factual dimensionality decrease process which
separates the arrangement of the symmetrical head segments from the first informational
indexes. During this process, there are 10 fold cross validation fields are to be used for
estimating the performance classifiers with 136 samples which helps to imply with 119 PCs in
principal component analysis process. This PC is to be used to generate of the training and
testing data are going to be chosen as h=119 [127, 128].

5.4.7 SYSTEM OF FEATURE RANKING

The component positioning technique assists with investigating the highlights

significance and furthermost class factors. This technique performs the majority of the decision
of discriminant highlights which help to advancing execution of classifier models. Consider D =
[X1, X2………, XN] T be a dataset have N tests, here Xi = (xi1, xi2…….,xiM) might be a vector of
M esteems. The vector fj = (x1j, x2j……xNj) T could be a vector of estimation of highlight fj.
Presently, D represents as N × M framework, here line I is the subject Xi and section j is that the
element of fj [129].

The positioning system calculation dependent on highlights is applied to dataset D

produces an arranged rundown of highlights Ψ = [f1, f2… fj] where the superscript indicates the
situation in the positioned rundown of highlight f and this rundown is requested by decrease
significance. Inferred on highlight positioning select the top k positioned highlights [f1,

120
f2,……,fk], k ≥ M where k dictated by client. Here the t-test is hand down for highlight
positioning methodology in condition 1, as follows [130]:
μC1 ― μC2
T= σ2
C1 σ2
+ n c2
(5.1)
n c1 c2

2
Where T is that the t-test esteem, 휇퐶1,휎2퐶1, 푛푐1 and 휇퐶2,휎푐2 , 푛푐2 are the mean, difference and
number of tests of two classes c1 and c2 separately. The most noteworthy enlightening highlights
are chosen by positioning all highlights steady with T esteems.

5.4.8 FISHER CRITERION

The ideal number of highlights is applied for fisher basis process, which is depends on
discriminative performance of features. This is often wont to automatic formation approach
[131]. Fisher Criterion,

wTSBw
J(w) = (5.2)
wTSBw

Here SB and SW indicate determinants among class and inside class dissipate networks. The
number of top discriminative features,
SB = (μC1 ― μC2)(μC1 ― μC2)T (5.3)
Consider two classes’ c1 and c2, the class disperses and inside class dissipate networks are
analyzed in condition 6.4 as follows:
w = s―1
w (μc1 ― μc2) (5.4)
μci is the mean information in each class. This strategy is utilized in adaptively finding the k
top discriminative highlights which is relies upon positioned t-test esteems with the help of
preparing information in each overlay as opposed of utilizing a fixxed k. when the highlights
are positioned, the quantity of top positioned includes iteratively increments from 1 to M
(number of highlights) by computing the separate Fisher Criterion. The quantity of top
positioned highlights is augmenting the Fisher Criterion, which is decided to be the ideal
number of top positioned highlights k.

121
5.4.9 DATA CLASIFICATION ON SVM CLASSIFIER

A bolster vector machine classifier assists with ordering the element extraction preparing
pictures which is relies upon factual learning specialized procedure. In light on the factual
learning hypothesis the Support Vector Machine (SVM) is the best classifier [132, 133, 134,
135]. The rule of auxiliary hazard minimization has incredible speculation capacity. In this
chapter, this classifier helps to predict the accuracy analysis with the assistance Cross Validiation
mechanism. This mechanism has two nested loops like outer loop and inner loop. At each and
every step, the data are divided to K1folds in outer loops. At each and every step, the data
are further more divided into (K1−1 folds) in an inner loops.

In the framework search, the estimations of C and γ are differed among the competitor
sets {2−5, 2−4… 0, 219, 220} and {2−15, 2−14… 0… 214, 215}, individually. The inward circle is
rehashed K2times, estimating the exactness of the classier over the K2folds for bend of of C
and γ. This strategy stage process is to be rehashed at multiple times in 10 – fold cross
overlap procedure in figure 5.4. In this time, there is no overlap is occurred [136].

The classification performance analysis parameters are determined which are

represented below.
(TP + TN)
ACC = (TP + FP + FN + TN) (5.5)
TP
SEN = TP + FN
(5.6)
TN
SPE = TN + FP
(5.7)

Where,
TP (True Positive): The number of AD is correctly identified as AD.
TN (True Negative): The number of HC patients correctly identified as HC.
FN (False Negative): The number of AD patients is incorrectly identified as HC.
FP (False Positive): The number of HC patients incorrectly identified as AD.

122
 Figure 5.4 Ten fold cross approved technique utilized for parameter tuning examination

5.4.10 ATROPHY CLUSTERS IN DATA FUSION

In this research, the atrophy clusters data fusion method assists with expanding
the general execution of the Alzheimer’s sickness classification analysis. The primary
objective of this technique is to integrate the data information from multiple sources of
classifiers to increase the high level of performance. In this research, the data fusion has
two different divisions like source fusion and score fusion methodology.

123
 5.4.10.1 SOURCE DATA FUSION
In this method, the high ranking features are to be predicted which depends on the
feature selection methodology from various volume of interests. Consider fv1, fv2 … fvn are
include vectors which are to be produced with the assistance proposed highlight determination
technique for every single decay bunch. The Feature Vector Fusion (FVF) is,
n
fvf = [fv1, fv2,……,fvn]1 × ∑i=1 mi (5.8)
Here mi is the vector length for fvi. At that point the connected component vector is
performed for classification process. The source information combination depends on techniques
for highlight withdrawal [137].

5.4.10.2 SCORE DATA FUSION

This method involves various classifiers methodologies. In this stage, the total number
of classifiers are determined depending on the clusters with the assistance VBM plus
DARTEL approach. In this research, score data fusion techniques are performed like larger
part casting a ballot method shown in figure 5.5.

Classifier
VOI1 R1
D1

Grey VBM + Atrophy VOI2 R2 D2 Fusion Results

Matter DARTEL clusters

VOIn R3 Dn

Figure 5.5 Larger parts casting a ballot based on Score Data Fusion

124
5.5 EXPERIMENTAL RESULTS
In this exploratory result, the data are obtained from MRI dataset in which it is acquired
for pre-handling the stage utilizing VBM with DARTEL examination on T1 weighted picture.
This is a marker unveiling, the importance of diminished dark issue volumes in Alzheimer’s
ailment. The performance depend on classification of t-Test based of Feature Selection that is
proposed dependent on Fisher Criterion is measured. The outcome of the result was obtained
with original image of quality of gray matter images using VBM with DARTEL. All MR images
are correlated for bias field in homogeneities are normalized in gray matter and white matter. For
the resultant work, gray matter images are only used. The absolute threshold masking is of
around 0.1 which is used to avoid possible edge that affects around the border between gray
matter and white matter. The result of gray matter volumes is investigated with VBM towards
advancing age in healthy control is shown in figure 5.6.

Figure 5.6 Age-related atrophy in gray and white matter in healthy controls. VBM analysis for
159 healthy volunteers aged from 17 to 68 years showed significant negative correlation of gray
matter and white matter volume with advancing age.
The result for the feature extraction based on atrophy of raw feature vectors is based on
VBM plus DARTEL is isolated the VOIs. The result shows the decreased gray matter volume
which is obtained using VBM plus DARTEL analysis in AD patients relative to HC are
segmented using 3D masks. The mask is applied to all smoothed gray matter density volumes

125
resulting with VBM plus DARTEL analysis, to extract voxel values as raw feature vectors. The
result of pictorized image of decreased gray matter in AD is shown in figure 5.7.

Figure 5.7 Gray Matter atrophy associated with the progress of AD.
The result of group comparison by VBM plus DARTEL reveals a significant decline in GM
volume in right hippocampus, left hippocampus, right inferior parietal lobule and right anterior
lobule in AD patients compared HC patients. The result of the brain region of gray matter
reduction atrophy in 68 patients and 68 age matched HC subjects are shown in figure5.8.

Figure 5.8 Brain region of gray matter reduction atrophy in 68 patients and 68 age matched HC
subjects.
The result of three dimensional view of group comparison is representing relative gray matter
atrophy patients with AD compared to HCs. VOI masks are applied to gray matter to extract the
voxel values into raw feature vectors in feature selection and classification. Figure 5.9 shows the
three dimensional view of brain showing gray matter atrophy using VBM technique plus
DARTEL.

126
 Figure 5.9 Three dimensional reconstruction of the brain showing gray matter atrophy using
VBM technique plus DARTEL. The regions of gray matter loss are shown from a) anterior,
b) posterior, c) right lateral, d) left lateral, e) inferior and f) superior view. The red region
represents the region of gray matter loss.

5.5.1 IMAGE QUALITY MEASUREMENT WITH ENHANCED IMAGE WITH

ERRORS
The image is measured by two binary classification classes which are performed in
classification error as re-substitution error and cross validation error.

5.5.1.1 Re-substitution Error (eresub)

The re-substitution error rate often underestimates the true error. The gap always depends
on the functionality of MRI dataset and classifier. The optimum will be high, when classifier is
with high complexity and sample size is small where n is low. In this eresub design a classifer as
sn, sn is the estimation of the respective error. It is calculated as:
1
eresub = n
‖( Ti ― φn (Xi))‖ (5.9)
It is exceptionally quick to process in contrast with other blunder estimator.

5.5.1.2 Cross Validation Error

The cross validation error is a randomized blunder. It haphazardly chooses the K folds. In
K - overlap cross validation the information are part in K fold as K = 10. One overlap is utilized
127
as test whereas (Sq = {(x1,c1), ….., (xq, cq)}) and the remaining folds are used for training (Sr =
{(x1,c1),……(xr,cr)}) where q and r are the training set. It is calculated as:
1
ek = q
‖(Ti ― φr ( Xi))‖ (5.10)

Different feature ranking methods:

The different feature ranking methods were obtained are as follows;

Statistical Dependency:
SD measures the level of dependency between the values of a feature and the associated
class labels. The SD between feature value X and class label C can be obtained as follows
P (Xi , Cj)
SD = ∑ i ∑j P (XI , Cj) P (Xi)P (Cj) (5.11)

Where P (Xi , Cj)is the frequency count of data X with value xi in the class cj, P(xi) is the
frequency count of data X with value xi, and P(Cj) is the frequency count of class C with value cj.
SD is nonnegative in the range of [0, 1], with SD = 0 indicating no correlation and SD = 1
denoting that C can be inferred once X is known. A larger SD means higher dependency between
the feature value and class labels.
Mutual Information:

MI measures the relevance of the feature value X and class label C by:

P (Xi , Cj)
MI = ∑i ∑j P (XI , Cj)log2 P (Xi)P (Cj) (5.12)

MI is similar to SD. P(xi, cj) is the frequency count of data X with value xi in the class cj, P(xi) is
the frequency count of data X with value xi, and P(cj) is the frequency count of class C with value
cj. MI is nonnegative in the range of [0, 1], with MI = 0 indicating no correlation, and MI = 1
meaning that C can be inferred once X is known.
Pearson’s Correlation:

PC is a measure of the relevance between the features and class label. PC of the feature
value X and class label C is calculated as follows:

cov (X , C)
PC = var (X)var (C)
(5.13)

128
which in binary classification becomes;
∑N
i=1 (Xi ― μx)(Ci ― μc)
PC = ∑N 2 2
i=1(Xi ― μx) (Ci ― μc )

(5.14)
Where PC is Pearson’s correlation value, and μx and μc are the mean of all samples of X and C,
respectively. PC has a value in the range of [−1, 1]. PC = 0 indicates independency of X and C,
PC = 1 denotes the highest positive correlation of them, and PC = −1denotes the highest negative
correlation. To select the top informative features, all the features were ranked According to their
absolute PC values.

t-Test Score:

The TS measures the statistical significance of the value differences between the two
classes. The t-test is performed by:

μc1 ― μc2

TS = σ2
c1 σ2
c2
n c1
+ n c2

(5.15)

Where TS is the t-test value and μc1, σc21, nc1 and μc2, σc22, nc2 are the mean, variance values, and
number of samples of two classes, c1 and c2. To select the top informative features, all the
features were ranked according to their absolute TS values.

Fishers Criterion:

FC measures between-class and within-class scatter matrices between two classes, as

shown below:

wTSBw
FC = (5.16)
wTSWw

129
where SB and SW represent the determinant of the between class and within-class scatter matrices,
respectively. The proposed feature selection method based on error for the training set is shown
in table 5.1.

Table 5.1 Feature selection method based on error

Measure / Methods Resubstitution Error Cross Validation Error
MI 854 470
SD 355 05
PC 1140 1271
TS 1140 1271
FC 1140 1271

The proposed feature selection method with error in Accuracy, Sensitivity and Specificity for re-
substitution error is shown in table 5.2 and the figure 5.14 is shown below the table.
Table 5.2 Feature Selection method with error in ACC, SEN and SPE in Re-substitution Error
with different measures
Measures Accuracy Sensitivity Specificity
SD 84.92 81.07 88.76
MI 86.84 85.92 81.75
PC 84.15 84.92 83.37
TS 84.15 84.92 83.37
FC 84.15 84.92 83.37

130
 Resubstitution Error
92

90

88
86
Accuracy
84
Sensitivity
82
Specificity
80

78
SD MI PC TS FC
Measures

Figure 5.14 Feature Selection methods with error in ACC, SEN and SPE in Resubstitution Error
with different measures.
The proposed feature selection methods with error in ACC, SEN and SPE in Cross
Validation Error with different measures are shown in table 5.3 and the figure 6.15 is shown
below the table.
Table 5.3 Feature selection methods with error in ACC, SEN and SPE in cross validation error
with different measures
Measures Accuracy Sensitivity Specificity
SD 87.61 86.46 88.76
MI 89.53 88.01 91.07
PC 87.23 89.53 84.92
TS 87.23 89.53 84.92
FC 87.23 89.53 84.92

131
 Cross Validation Error
92
91
90
89
88
87
Accuracy
86
Sensitivity
85
84 Specificity
83
82
81
SD MI PC TS FC
Measures

Figure 5.15 Feature selection methods with error in ACC, SEN, and SPE in cross validation error
with different measures.

5.5.2 EXECUTION ANALYSIS

5.5.2.1 EXECUTION ANALYSIS ON RAW FEATURE VECTORS OF ATROPHY

CLUSTERS

In this chapter, ranking based classification of Alzheimer’s disease in hippocampus

region from structural magnetic resonance imaging using support vector machine is to be
proposed. There are different magnetic reverberation picture datasets are to be collected which
have of 68 AD and also 68 control subjects. There are different parameters are to be analyzed
like accuracy, specificity and sensitivity. These are to be analyzed with the assistance 10 fold
cross support vector machine classifier. There are five different VOIs obtained in table 5.4 and
which is representing in figure 5.16 and figure 5.17.
Table 5.4 Raw feature vectors consummation of atrophy clusters using 10 fold cross validation
different SVM

132
 Linear SVM RBF SVM

ACC SEN SPE AUC ACC SEN SPE AUC

(%) (%) (%) (%) (%) (%) (%) (%)

VOI1 80.14 79.41 80.77 84.37 81.35 80.77 82.81 87.71

VOI2 77.20 77.94 75.47 83.93 78.41 75.47 81.35 86.69

VOI3 71.32 70.38 71.05 74.63 74.25 71.05 77.34 80.25

VOI4 69.85 69.07 70.49 77.82 71.52 72.52 66.54 77.33

VOIall 77.20 79.21 74.58 83.39 82.35 82.82 82.79 85.00

Average 75.14 75.20 74.47 80.82 77.57 76.52 78.16 83.39

Figure5.16 Linear SVM parameters analysis

133
 Figure 5.17 RBF SVM parameters analysis

5.5.2.2 EXECUTION ANALYSIS OF STAGING OF PCA METHOD

There are different raw feature vectors extracted with the assistance PCA which is
depends on the reduction methodology process. In the PCA analysis, the dataset and the features
are to be detected which are decreased to lower dimensional features with 122 PCs.ACC –
Accuracy, SEN – Sensitivity, SPE – Specificity, AUC – Area Under Curve, SVM – Support
Vector Machine, RBF – Radial Basis Function. In this analysis, the PCA will be analyzed based
on different parameters like accuracy, sensitivity, specificity and area which are to be
represented in table 5.5 and figure 5.18 and 5.19. The raw feature vectors average accuracy will
be analyzed without dimensionally reduction.

Table 5.5 PCA staging of atrophy clusters using 10-fold cross validation with 122 PCs.

Linear SVM RBF SVM

134
 ACC SEN SPE AUC ACC SEN SPE AUC
(%) (%) (%) (%) (%) (%) (%) (%)

VOI1 79.41 81.35 75.47 85.80 81.61 85.76 75.47 87.27

VOI2 74.26 75.47 71.50 82.06 82.35 81.35 81.35 86.59

VOI3 70.58 72.52 67.54 71.48 69.85 68.85 70.11 77.33

VOI4 69.58 69.11 70.58 80.54 71.32 70.32 75.47 78.35

VOIall 77.20 75.17 71.34 86.47 85.29 84.29 82.52 87.74

Average 74.20 74.72 71.28 81.27 78.08 78.11 76.98 83.45

Figure 5.18 Linear SVM parameters analysis – PCA

135
 Figure 5.19 RBF SVM parameters analysis – PCA
5.5.2.3. PERFORMANCE OF THE PROPOSED FEATUURE SELECTION USING t-
TEST RANKING AND FISHER CRITERION
The result of the proposed feature selection, using t-Test ranking and Fisher Criterion is
done for ranking the features. It is determined to find the optimal number of top ranked features
in Irving Fisher. The Fisher Score for the dataset range from 1 to VOIall. The result shown in the
figure 5.10 mention the maximum Fisher Score located at 111, which means the top ranked
features are to be used in classification process. The typical Fisher Scores are observed between
30 and 150 for all folds of validation of five different VOIs.

136
 Fig. 5.10 . Fisher scores for the respective ranked features in fold 1 dataset of VOIall.

The result for the same data is represented using t-Test values are shown in figure 5.11.

Fig. 5.11 t-test (T) values for the respective ranked features in fold 1 dataset of VOIall.
The result for the accuracy is contribution of features is plotted with Linear SVM. The t-
Test value is highly correlated to the contribution features. The accuracy of the result is shown
in figure 5.12.

137
 Fig. 5.12 Classification accuracies of linear SVM with respect to different numbers of features
selected in fold 1 dataset of VOIall.
The performance in improvement of accuracy is obtained by using progressive inclusion
of the ranked features in feature vector with Linear SVM. The 111is top ranked features in this
fold validation is top, which does not increase further. The result shows the increased number of
ranked features in classification is shown in figure 5.13.

Fig. 5.13. Classification accuracies of linear SVM with respect to different numbers of top
ranked features selected in fold 1 dataset of VOIall.
The SVM based classifiers are used to observe the classification performance of
the selected feature vectors from five different VOIs. The result of classifiers are presented in
table 5.6. The table 5.4 and 5.6 confirms that the proposed feature section method significantly
improves the prediction capability of AD subjects when compared to prediction using raw
138
features. The average accuracy for raw data for linear and RBF SVM classifiers is 75.14% and
78.23%, while the average accuracy for the proposed feature selection method is 86.76% and
86.76%. The improvement is around 10% for all performance indicators: ACC, SEN, SPE and
AUC. the result analysis of feature selection method is presented on table 5.6 and graphical
representation is shown in figure 5.20 and 5.21.

Table 5.6 Performance results of the Feature Selection method

Linear SVM RBF SVM

ACC SEN SPE (%) AUC ACC SEN SPE (%) AUC
(%) (%) (%) (%) (%) (%)

VOI1 91.17 92.64 88.70 95.90 90.44 88.70 90.17 94.07

VOI2 92.64 91.17 93.11 96.93 94.11 91.64 94.58 97.74

VOI3 76.47 73.52 78.41 83.93 76.47 75.00 76.84 83.66

VOI4 79.41 75.00 81.82 85.67 80.14 72.52 85.76 88.29

VOIall 94.11 95.58 91.64 96.33 92.64 93.11 90.17 97.13

Average 86.76 85.58 86.73 91.75 86.76 84.19 87.50 92.17

139
Figure 5.20 Linear SVM parameters analysis – Feature Selection

Figure 5.21 RBF SVM parameters analysis – Feature Selection

140
5.5.2.3 PERFORMANCE OF DATA FUSION IN ATROPHY CLUSTERS
In this section, the performance will be analyzed like data fusion support which is 10
percentage more than the predicted average performance. The performance can be obtained
with the assistance predicted unique clusters. The score fusion (major voting) approach is
more than or equal to source concatenation approach shown in the table 5.7.
Table 5.7 Performance of data fusion technique among atrophy clusters of Gray Matter.
Linear SVM RBF SVM

ACC SEN SPE AUC ACC SEN SPE AUC

(%) (%) (%) (%) (%) (%) (%) (%)
Source 95.35 93.11 96.05 96.52 95.35 93.09 96.05 96.31
Concatenation
Majority 96.12 93.11 97.52 98.93 95.39 93.11 96.05 98.72
Voting

5.5.3 CORRELATION ANALYSIS ON CLASSIFICATION

5.5.3.1 CORRELATION OF CLASSIFICATION PERFORMANCE WITH OTHER
EXISTING METHODS
The correlation of the outcome in characterization depends on highlight positioning
with other existing techniques and fisher rule is viewed as halting standard. The exhibition of
MI highlight in exactness is higher than other positioning techniques. The correlation of
arrangement execution with other existing techniques or proportion of various component
positioning strategies is shown in table 5.8.
Table 5.8 Correlation of characterization execution with other existing techniques for various
positioning strategies.
Method Stopping criterion Accuracy (%) Sensitivity (%) Specificity (%)

SD FC 85.30 84.51 85.67

MI FC 86.05 84.81 87.21
PC FC 84.81 83.21 86.41
TS FC 84.81 83.21 86.41
FC FC 84.81 83.21 86.41

141
5.5.3.2 CORRELATION ON PERFORMANCE WITH OTHER METHODS ON MRI
DATA
The performance of feature selection and data fusion technique are out performed in
alternative techniques. The parameter of classification performance with different methods on
MRI data shows the proposed system in highly performance terms for accuracy, sensitivity and
specificity. Several studies have reported the classification results to distinguish on AD patients
and HCs based on dataset of MRI. Zhang et al. used a multimodal classification of AD based on
a combination of MRI, CSF, and PET data. They reported an ACC of 86.2% in the classification
of AD/HC using the MRI data. By combining the MRI, CSF, and PET results, they achieved a
high ACC of 93.2%. Westman et al. reported an ACC of 87% using MRI data and increased the
ACC to 91.8% by combining the MRI data with CSF measures. Aguilar et al. employed the
FreeSurfer software to compute cortical thickness and volumetric measures. Based on an
artificial neural network classifier and MRI data, they achieved an ACC of 84.9% and an ACC of
88.8% using an SVM classifier and a combination of MRI data with educational and
demographic data. Koppel et al. achieved an ACC of 81.1% by combining GM and white matter
modalities in MRI data. The result shows the performance of proposed feature selection method
using only in MRI data is higher or comparable to other methods reported in literature. The
supervised classification results of Alzheimer’s ailment and a healthy control subject on MRI
data is shown in table 5.9.
Table 5.9 Supervised classification results of AD and HCS on MRI Data.
Author Imaging AD / HC Validation ACC SEN SPE

Modality Method % % %

Zhang et. al. MRI 51/52 10 Fold 86.2 86.0 86.3

West man et. al. MRI 96/111 10 Fold 91.8 88.5 94.6

Aguilar et. al. MRI 116/110 10 Fold 84.9 80.2 90.0

Koppel et. al. MRI 53/57 10 Fold 81.1 80.6 93.0

Proposed Method MRI 68/68 10 Fold 96.3 94.1 98.5

142
5.6 SUMMARY
In image processing technique one among the main considerations within the field of
visual perception to spot the presence of objects or tumor in a picture is image segmentation.
The division of the image into meaningful segments which correspond to significant objects
present within the image is image segmentation main motivation. Assigning similar pixel of a
picture supported the similarity criteria, the pixels with an equivalent properties combined to
make a meaningful region of the image is that the image segmentation process.
In human brain, cerebrum is that the largest part which is assessed to two cerebral
hemispheres. Each and each hemisphere is assessed into four lobes which are denoted as frontal
flap, parietal flap, fleeting projection and occipital flaps. In human brain, a cerebrum is
integrated by brainstem to medulla spinalis which is midbrain and medulla oblongata. The
mixture of the two different classification analyses gives an honest result in MRI images. These
methods are very effective, efficient and vigorous technique for expectation of Alzheimer’s
malady. These dataset are collected with the assistance of correspond doctor who have already
been diagnosed. An excellent expansion method of knowledge mining technique is understood as
image mining mechanism. The implementation method of Bayesian networks with high
extensive processing stage with image forensic work will present as next stage. Inside the
Alzheimer's illness on reverberation picture examine in cerebrum is broken down that the
ultimate result of utilizing region developing calculation. Using shape analysis technique the
disease area of AD is analyzed. To characterize and foresee the hippocampus shape area to
process the picture of AD in human mind applies the more drawn out term arrangement
procedures.

143
 Chapter - 6
Comparison of all Proposed Segmentation and Classification Strategy
6.1 INTRODUCTION

Alzheimer’s is a type of dementia which is common in older ages. This disease is a

progressive form of neurological disease that causes the destruction of brain cells. In this deep
learning architecture were used. Deep learning is a type of learning in which algorithms process
data and performs the learning process inspired by the human brain. Alzheimer’s disease in
hippocampus region is segmented and classified successfully based on hippocampal
segmentation and classification using jointly deep convolutional neural network, fuzzy neural
network and support vector machines. The accuracy is compared with various segmentation
methods. Subsequently, the hippocampus image in brain were classified based on the outcome of
segmented results by using suitable classifier on the baseline of T1 – weighted and flair images
of structural MRI data. In this chapter, all the methods regarding segmentation and classification
for Alzheimer’s disease in hippocampus on human brain images dataset are investigated. Then
the performances of various methods are compared among themselves.
In this part, executive investigation is organized as follows: Section 6.2 presents the
correlation fluctuated of the varied sifting, region examination and improved strategies, Section
6.3 shows the examination of the various segmentation methods analysis with different stages;
Section 6.4 speaks to the correlation of shifted characterization examination for Alzheimer’s
illness cerebrum pictures and Section 6.5 synopsis of the perception.

6.2 Correlation of different upgrade investigations

An image enhancement defines the image features increasing process like edge analytics,
contrast sketching analysis and developing images than the first image. This image enhancing
process has the following analysis like preprocessing stage of cerebrum pictures that are the
detection of MRI images, segmentation process, recognition and classification analysis. One can
find in this section, there are different Alzheimer’s disease affected patients dataset which are
collected from different scanning areas and also different hospitals. The brain datasets are to be
performed for image enhancement techniques. All proposed methods are to be crazy normal and

144
affected brain analysis. There are various parameters are to be analysed like sensitivity,
specificity, precision and recall which are to be compared with different enhancement process.
The ADNI data set is to be used. This dataset developed by Dr. Randy Buckner at the spot of
Hughes Medical Institute. During this dataset, it contains 416 images presented which cover
infants up to 18 years old. The sample scan slice of ADNI dataset of Alzheimers Disease, Mild
Coginitive Impairment and Normal Control in hippocampus of human brain is shown in the
figure 6.1 respectively.

a) Sample of AD b) Sample of MCI c) Sample of NC

Figure 6.1 Sample of scan slice of ADNI dataset
The final grain field registered image helps for transverse cross sectional operation. On
this research, the Alzheimer disease detection is more accurate process and also it gives the result
as level of severity. There are different matrics are used to predict the simulations of the results
of various denoising methodologies. They are various conventional median, wiener filter,
Gaussian filter and similar things are to be analyzed. The typical error of the improved denoised
images with their corresponding original images is analyzed. The ANE values lie in the range of
0 to 1. The minimum value is taken as the minimum error rate. According to the NAE, minimum
error value is appreciated. Table 6.1 presents ANE evaluation of different filtering techniques.
The absolute error value is low in the Poisson filter. The deconvolution filter has the minimum
error value with the Poisson distribution. If the mean of the image pixels is increased, the error
rate also increases. It works good at minimum error level. The median filter also performs well in
reducing noises. Here, wiener filter is not suitable as the normal error rate ANE is high. The
figure 6.2 illustrates the error encountered in different filtering techniques. The default values of
the mean and variance are tested and the average results are presented in Table 6.1. In this table,
wiener filter giving the low error rate and high signal to noise ratio is presented. The absolute

145
error value is low in the wiener filter. The wiener filter has the minimum error value. This
approach is suitable, if the mean of the image pixels is increased. Median filter also performs
well in reducing noises. Here, wiener filter is not suitable as the normal error rate NAE is high.
Table 6.1 Statistical Analysis with Poisson removal

ABSOLUTE NORMALIZED ERROR

Shot Salt & Uniform

Metrics / Images Gaussian Film Grain Poisson Speckle
noise pepper noise

Wiener filter 0.42 0.39 0.54 0.72 0.79 0.85 0.61

Gaussian filter 0.56 0.67 0.65 0.53 0.65 0.79 0.85

Median filter 0.62 0.81 0.89 0.85 0.8 0.81 0.88

De-convolution
0.59 0.75 0.79 0.81 0.59 0.80 0.71
filter

Error evaluation of different filtering techniques

1

0.8

0.6

0.4

0.2

0
Wiener filter Gaussian filter Median filter De-convolution filter
Filtering Techniques

Absolute Normalized Error Absolute Normalized Error Absolute Normalized Error

Absolute Normalized Error Absolute Normalized Error Absolute Normalized Error
Absolute Normalized Error

Figure 6.2 Statistical analysis of different filtering techniques

146
6.3 Correlation of segmentation strategies

Segmentation of image is the important process of differentiating or partitioning the

digital source images to varied sorts of segments. The processing of images is to be separated in
digital source images like group of pixels or individual pixels. The main aim or objective of
segmentation process is extremely simple for formation of digital images which is represented
into something which is extremely easy to research and also to analyze. Digital image
segmentation process helps to place an object which has suitable boundaries and also assigning
each and every pixels into label predicted characteristics format.

The performance of the proposed framework on hippocampus segmentation in terms of

DSC, PPV, SEN, and VE are compared the results with different parameter settings and also
performed with empirical comparation with a few other competing methods. The comparison of
segmentation results of proposed method with other three methods including FSL and two deep
learning-based methods were denoted as “FSL”, “Thyreau’s method” and “Cao’s method”,
respectively. FSL is a comprehensive library of analysis tools for fMRI, MRI and DTI brain
images.

Thyreau’s method was proposed to segment the bilateral hippocampi using a deep-
learning appearance model by transferring algorithmic knowledge for large cohort processing. A
wide and variable training set from multiple cohorts was used to train the segmentation model.
Since the trained model is available online to test our data set for comparison. It is worth noting
that the differences in the definitions of manual segmentations for training and testing data may
bias the evaluation results. Cao’s method was proposed to construct a multi-task deep learning
method for joint hippocampal segmentation and clinical score regression. Here figure. 6.3 shows
the comparison of the segmentation results using different methods for three example subjects
selected from the AD, MCI and NC groups of the MRI data, respectively. The original images,
the segmented results of FSL, Thyreau’s method, Cao’s method, and proposed method and
ground truth are demonstrated in columns from left to right in Fig. 6.3. From the original
images, it is not easy to distinguish the hippocampal regions from the adjacent tissues due to the
small difference between their intensity values. The segmented hippocampal regions obtained by
proposed method appear to be smoother and more accurate than those by other methods. In
Table 6.2 shows the comparison of segmentation results in terms of DSC, PPV, SEN and VE on

147
the same training and testing data sets. Deep CNN method achieves the best performance
compared to other methods.As for the segmentation, the proposed method achieved the DSC of
87.0% on the ADNI dataset and DSC of 86.7% on the HarP dataset. The comparison of
segmentation results in terms of DSC, PPV, SEN and VE is shown in figure 6.4 respectively.

Figure 6.3 Comparison of segmentation results using different methods

Table 6.2 Comparison of segmentation results in terms of DSC, PPV, SEN and VE

Performance (%) FSL Thyreau’s method Cao’s method Proposed Method

DSC 72.2 73.8 84.6 87.0
PPV 70.4 68.8 83.7 84.6
SEN 74.6 80.2 85.7 89.7
VE 20.3 18.4 7.2 6.0

148
 Comparison of Segmentation results using different Methods
100
90
80
70
60
50 DSC
40 PPV
30
SEN
20
10
VE
0
FSL Thyreau’s Cao’s method Proposed
method Method
Methods

Figure 6.4 Comparison of segmentation results in terms of DSC, PPV, SEN and VE

6.4 Correlation of classification techniques

Image classification is a crucial process which is in a position to separate a picture drawn
on visual content. It is a process of mixing different pixels into predicted image with high
visuality. During this classification process, a gaggle of knowledge is to be separated into
different classes or different categories. This arrangement investigation is to be isolated into two
distinct classifications like administered learning and unailed learning technique.
The analysis of different classifier for the brain in alzheimers disease on hippocampus
image classification is discussed in this section. In general, the MRI brain alzheimers disease
brain images are in different modalities and sequences. The assessment of the performance for
the deep CNN classifier with the different distribution functions are illustrated in table 6.3, here
the classification is performed and compared with other deep networks method such as LeNet,
VGGNet and DenseNet. The VGGNet has a deeper architecture with significantly more layers
than LeNet. The process of DenseNet greatly outperformed the deep CNN model for disease
classification, suggesting the features learned by DenseNet are more important for disease
classification. Therefore the performance for the deep CNN classifier with the different
distribution functions is shown in figure 6.5.

149
 Table 6.3 Performance for the deep CNN classifier with the different distribution functions
Methods Accuracy (%) Sensitivity (%) Specificity (%)
LeNet 83.8 73.2 92.4
VGGNet 84.7 77.3 90.8
DenseNet 86.6 79.4 92.4
Deep CNN 80.1 79.9 80.3

Performance of deep CNN classifier with different

distribution functions
100
90
80
70
60
50 Accuracy (%)
40 Sensitivity (%)
30
Specificity (%)
20
10
0
LeNet VGGNet DenseNet Deep CNN
Methods

Figure 6.5 Performance for the deep CNN classifier with the different functions
The comparison of different features for classification of alzheimers disease in hippocampus
with normal control classification is shown in table 6.4 and its figure is shown in 6.5.
Table 6.4 Comparison of different features for classification
Methods Accuracy (%) Sensitivity (%) Specificity (%)
Hippocampal Volumes 80.3 73.5 86.2

ROI features 84.7 77.4 90.5

Voxel-wise features 86.1 84.1 87.5

Proposed Method 88.7 86.5 90.7

150
 Comparison of different features for classification of Alzheimers
Disease
100

80

60
Accuracy (%)
40
Sensitivity (%)
20
Specificity (%)
0
Hippocampal ROI features Voxel-wise Proposed Method
Volumes features
Methods

Figure 6.6 Comparison of different features for Classification of AD

The proposed method achieved a very competitive classification accuracy of 88.7 percentages
for alzheimers disease classification, which is higher than ROI features 84.7 percentages and
voxel wise features 86.1 percentages. The features from ROIs were more informative than
hippocampal volumes and yielded better performance.

6.5 RECAPITULATION

In this research study, there are various performance parameters that are present within
the research proposed mechanism. The developed images are gathered by the proposed approach,
which are regularly processed for further processing segmentation methodologies. In this
chapter, the proposed performance is to be analyzed with different stages like comparative image
enhancement process, image segmentation process and image classification process. Over this
research this proposed approach reaches 98.7percentage accuracy and requires less time to
process. The bolster vector machine classifier is applied to the mind picture order process. In
future, the proposed approach will prolong to accomplish the most extreme exactness and it is
frequently reached out for weighted pictures.

151
 Chapter - 7
Conclusion and Scope for Future Work
7.1 CONCLUSION
In recent years, many machine learning models have been employed to analyze the MRI
data for identifying biomarkers and deciphering the disease etiology. A multi-task method was
proposed to select the most discriminative features from region of interests for multimodal
classification of Alzheimer’s disease / Mild Cognitive Impairment. An approach was proposed to
combine the marginal fisher analysis with norm-based multi-kernel learning to achieve the
sparsity of region of interests, which could simultaneously select a subset of relevant brain
regions and learn a dimensionality transformation. A deep network with stacked auto encoder
was employed on the region of interest features to extract the latent high-level features for
improving the performance of disease classification using neuro imaging data.
Recently, deep learning convolutional neural networks (CNNs) have been widely used in
image segmentation, classification and computer vision. The deep 3D CNNs were used to extract
the features of 3D medical images for classification. A multi-task deep learning method was
proposed for joint hippocampal segmentation and clinical score regression using MRI scans.
Given the very high dimensionality of the brain MRI data, it requires huge computational
resources and a large dataset to train a deeper CNN with robustness. Since the MRI datasets used
for AD diagnosis are typically very small compared with the datasets used in computer vision, it
remains a major challenge to train a deeper CNN model with a large number of parameters to be
learned. Among all brain ROIs, the hippocampus is one of the first affected regions in AD, and it
is an important anatomical region in the AD etiology.
The proposed new classification framework based on multi-model deep CNNs for jointly
hippocampal segmentation and disease classification. First, a multi-task deep CNN model was
constructed to learn the features for hippocampal segmentation and disease classification. Based
on the segmented hippocampal region, an additional 3D DenseNet was built to learn the rich and
detailed image features for disease classification. Finally, the learned features from the multi-task
CNN and DenseNet models are combined to classify disease status. The proposed framework
can not only output the disease status, but also provide the hippocampal segmentation result. No
tissue segmentation and nonlinear registration are required for MR image processing. The

152
experimental results based on the ADNI dataset have demonstrated that our proposed approach
has achieved promising performance for AD and MCI diagnosis.
The proposed fuzzy neural network method is used for automated multiclass diagnosis of
Alzheimer’s disease and classification based on human brain MR images. It is simple and robust
for classification approach in MRI scans. In the initial stage of the work, the pre-processing is
done to remove noises. Ant Colony based Optimization technique used for remove the high
frequency components. 2D histogram signal is obtained from pre-processed MR images of brain
and then further feature extraction is done by using Discrete Wavelet Transform. The features
are utilized for the purpose of training a fuzzy neural network, in order to classify the features
into three classes like Alzheimer’s, Mild Alzheimer's and Huntington's disease. The proposed
method provides better classification accuracy compared to other volumetric methods.
The main propound is a feature selection method used in t-test based feature ranking,
which is mainly for classification of Alzheimer’s disease. The optimal size of the selected
features is determined by using the Irving Fisher, which maximizes the class separation between
Alzheimer’s disease and Healthy Control. The feature selection is applied to all voxels that
passes through masks modeled by overall atrophy clusters are determined by using VBM
analysis plus DARTEL analysis. The talaiarch coordinates is a 3 dimensional co-ordinate system
of human brain, which is used to map the location of Alzheimer’s disease independently from
individual difference in the size and overall shape of hippocampus region in human brain.
Talaiarch is for functional brain imaging on Alzheimer’s disease and to target transactional
simulation of brain regions. It is also rightly called as Tournoux system. Linear and RBF kernel
based Support Vector Machine classifiers are used for the classification of extracted features
vectors after the propounded feature selection method. A consummation improvement is also
initiate by applying data fusion among the individual atrophy clusters as well as overall atrophy
clusters. The data fusion technique such as source and score fusion are used to obtain improved
staging in the classification of Alzheimer’s disease. The performance is measured with 136
subjects, where as 68 Alzheimer’s disease and 68 Healthy Controls Subjects, the complete data
set is obtained from open source technologies of ADNI data set and from Advance KG scan
research center using 10-fold cross validation. The results reveal that the propound approach for
accuracy, specificity and area under the curve is highly competitive with the state of art
techniques using Magnetic Resonance Images data.

153
7.2 SCOPE FOR FUTURE WORK

The combination of the two different classification analyses gives an honest end in MRI
images. These techniques are extremely successful, proficient and heart strategy for forecast of
Alzheimer’s sickness. The dataset are gathered with the assistance relate specialist who were at
that point analyzed. In these techniques, Alzheimer’s infection classification results are more
precise one and top quality.

This classifier is taken as information parameters which are to be extracted in auxiliary

attractive reverberation pictures. This system will be assessed from MCI patients and for other
neurodegenerative maladies and its power will be evaluated on patients with pictures got from
various MR scanners with different obtaining parameters.
The present research specializes in Alzheimer’s infection identification, division and
characterization investigation, and this is a continued work for improving detection and
classification of various groups of affected cerebrum pictures. In future, the strategies will be
expounded to the accommodation of further anticipated procedures with various multi modular
advancements.

154
 REFERENCES

1. Anjan Gudigar, U. Ragavendra, Edward J, Ciaccio, N. Arunkumar, Enas

Abdulhay, and U. Rajendra Acharya, “ Automated categorization of multi class brain
abnormalities using decomposition techniques with MRI images: A comparative
study”, IEEE, 2169 – 3536, Vol.7, 2019.

2. Gonzalez, R.C. and R.E. Woods, Digital Image Processing, 2nd ed., Beijing: Publishing
House of Electronics Industry,Prentice-Hall, 2002.

3. Kumar,A. and F.Shaik, Image Processing in Diabetic related causes, Forensic and
Medical Bioinformatics – Springer publications, ISBN:978-981-287-623-2, 2016.

4. Magnin, L. Mesrob, S. Kinkingnéhun et al., “Support vector machine-based classification

of Alzheimer's disease from whole-brain anatomical MRI,” Neuroradiology, vol. 51, no. 2, pp.
73–83, 2009.

5. Chen, B. D. Ward, C. Xie et al., “Classification of Alzheimer disease, mild cognitive

impairment, and normal cognitive status with large-scale network analysis based on resting-
state functional MR imaging,” Radiology, vol. 259, no. 1, pp. 213–221, 2011.

6. Goyal, Soniya and Biswas. “Automatic Detection of Brain Abnormalities and

Hippocampal Segmentation in MRI Sequences”. Image and Vision Computing New Zealand
Conference,IVCNZ 2011:NewZealand.

7. Gonzalez, R.C. and Woods, R.E. and Eddins, S.L., Digital Image Processing using
MATLAB, Gatesmark Publishing, A Divison of Gatesmark, LLC.
8. Himar Fabelo, S. Ortega, A. Szolna, D. Bulters, J. F. Pineiro, “In – Vivo
hyperspectral human brain image database for brain cancer detection” IEEE, 2169 –
3536, 2019.
9. P. Muthu Krishnammal and S. Selvakumar Raja, “Convolutional Neural Network
based image classification and detection of abnormalities in MRI brain images” IEEE,
978- 1 – 5386 – 7595 – 319, 2019.

10. Gonzalez R.C. and R.E. Woods, Digital Image Processing, Third ed., Prentice Hall, Upper
saddle River, NJ.

155
11. Makandar,A. and B.Halalli, Breast Cancer Image Enhancement using Median Filter and
CLAHE, International Journal of Scientific & Engineering Research, 2015, 6(4): 462-465.
12. M. R. Daliri, “Automated diagnosis of Alzheimer disease using the scale-invariant feature
transforms in magnetic resonance images”, MSJ, vol. 36, no. 2, 2012.
13. G. Song, Z. Huang, Y. Zhao, Z. Zhao, Y. Liu, M. Bao, J. Han, and P. Li, “A
noninvasive system for the automatic detection of gliomas based on hybrid features
and PSO – KSVM”, IEEE, 2169 – 3536, 2019.
14. Convit A, de Asis J, de Leon MJ, et al. Atrophy of the medial occipito temporal, inferior,
and middle temporal gyri in non-demented elderly predict decline to Alzheimer's
disease.Neurobiol Aging2000;21:19–26.
15. R. Wolz, V. Julkunen, J. Koikkalainen, et al., “Multi-method analysis of MRI images in
early diagnostics of Alzheimer’s disease,” vol. 6, no. 10, Article ID e25446, 2011.
16. Sharumathi.K, R. Priyadharsini, “A SURVEY ON VARIOUS IMAGE
ENHANCEMENT TECHNIQUES FOR UNDERWATER ACOUSTIC IMAGES”, IEEE, 978-
1-4673-9939-5, 2016
17. Guy Gilboa, Nir Sochen, Yehoshua Y. Zeevi, “Image Enhancement and Denoising by
Complex Diffusion Processes”, IEEE, 0162-8828, 2004.
18. Randeep Kaur, Sandeep Kaur, “Comparison of Contrast Enhancement Techniques for
Medical Image”, IEEE, 2016.
19. Manasvi Kalra, Sujeet Kumar, “Various Image Enhancement Techniques for Skin Cancer
Detection Using Mobile App”, IEEE, 2015.
20. Liang Shen, and Rangaraj M. Rangayyan, “A Segmentation-Based Lossless Image
Coding Method for High-Resolution Medical Image Compression”, IEEE, 0278–0062, 1997.
21. Yubing Li, Jinbo Zhang, Peng Gao, Liangcheng Jiang, Ming Chen, “Grab Cut Image
Segmentation Based on Image Region”, IEEE, 978-1-5386-4991-6, 2018.
22. Annegreet van Opbroek, M. Arfan Ikram, Meike W. Vernooij, Marleen de Bruijne,
“Transfer Learning Improves Supervised Image Segmentation Across Imaging Protocols”,
IEEE, 0278-0062, 2013.
23. G. Song, Z. Huang, “A noninvasive system for the automatic detection of gliomas
based on hybrid features and PSO – KSVM”, IEEE, 2169 – 3536, 2019.

156
24. Akshaya Pisal, Ravindra Sor, Dr.K.S.Kinage, “Facial Feature Extraction Using
Hierarchical MAX(HMAX) Method”, IEEE, 978-1-5386-4008-1, 2017.
25. Divi Galih Prasetyo Putri, Daniel Oranova Siahaan, “Software Feature Extraction using
Infrequent Feature Extraction”, IEEE, 978-1-5090-0741-7, 2016.
26. Hongmoing Li, Y. Fan, “Early prediction of Alzhemer’s disease dementia based
on baseline hippocampal MRI and 1 – year follow up cognitive measures using deep
recurrent neural networks”, IEEE, 978 – 1 – 5386 – 3640 – 4, 2019.
27. M. Amin – Naji, H. Mahdavinataj, A. Aghagolzadesh, “Alzheimer’s disease
diagnosis form structural MRI using Siamese convolutional neural network”, IEEE,
978 – 1 – 7281 – 1621 – 1, 2019.
28. B. S. D. Abajo, I. D. L. T. Diez, S. H. Bueno, R. M. Sanz, M. L. Coronado,
“Mobile Alzhes_alert: app for calculating the probability of suffering Alzheimer’s
disease”, IEEE, 978 – 989 – 98434 – 9 – 3, 2019.
29. J. Fritsch, S. Wankerl, E. Noth, “Automatic diagnosis of Alzheimer’s disease
using neural network language models”, IEEE, 978 – 1 – 5386 – 4658 - , 2019.
30. L. Xu, Z. Yao, J. Li, chen lv, H. Zhang and B. Hu, “Sparse feature learning with
label information for Alzheimer’s disease classification based on magnetic resonance
imaging”, IEEE, 2169 – 3536, 2019.
31. Biao Hou, Xiaohuan Luo, Shuang Wang, Licheng Jiao, Xiangrong Zhang,
“POLARIMETRIC SAR IMAGES CLASSIFICATION USING DEEP BELIEF NETWORKS
WITH LEARNING FEATURES”, IEEE, 978-1-4799-7929-5, 2015.
32. S.Manthira Moorthi, Indranil Misra, Rajdeep Kaur, Nikunj P Darji and R. Ramakrishnan,
“Kernel based learning approach for satellite image classification using support vector
machine”, IEEE, 978-1-4244-9477-4, 2011.
33. Rajkumar, "A Multi- Stage Hybrid , CAD Approach for MRI Brain Alzheimer
Recognition and Classification", The IIOAB Journal School of Computing Science and
Engineering VIT University, Januar 2016, India.
34. Ahmed, O.B., Benois-Pineau, J., Allard, M., Amar, C.B., Catheline, G., 2015.
Classification of Alzheimer’s disease subjects from MRI using hippocampal visual features.
Multimedia Tools & Applications 74, 1249-1266.

35. Amoroso, N., Rocca, M., Bellotti, R., Fanizzi, A., Monaco, A., Tangaro, S., 2018.

157
Alzheimer’s disease diagnosis based on the Hippocampal Unified Multi-Atlas Network
(HUMAN) algorithm. Biomedical Engineering Online 17, 6.

36. Azar, Z., S., F.V., C., P.J., D., L.C., The EADC-ADNI harmonized protocol for
hippocampal segmentation: A validation study. Neuroimage, S1053811918305846.

37. Beg, M.F., Raamana, P.R., Barbieri, S., Wang, L., 2013. Comparison of four shape
features for detecting hippocampal shape changes in early Alzheimer's. Statistical Methods in
Medical Research 22, 439-462.

38. Boccardi, M., Bocchetta, M., Morency, F.C., Collins, D.L., Nishikawa, M., Ganzola, R.,
Grothe, M.J., Wolf, D., Redolfi, A., Pievani, M., 2015. Training labels for hippocampal
segmentation based on the EADC-ADNI harmonized hippocampal protocol. Alzheimers &
Dementia 11, 175-183.

39. Cao, L., Li, L., Zheng, J., Fan, X., Yin, F., Shen, H., Zhang, J., 2018. Multi-task neural
networks for joint hippocampus segmentation and clinical score regression. Multimedia Tools
& Applications, 1-18.

40. Cao, P., Liu, X., Yang, J., Zhao, D., Huang, M., Zhang, J., Zaiane, O., 2017. Nonlinearity-
aware based dimensionality reduction and over-sampling for AD/MCI classification from MRI
measures. Computers in Biology & Medicine 91.

41. Chen, J., Yang, L., Zhang, Y., Alber, M., Chen, D.Z., 2016. Combining Fully
Convolutional and Recurrent Neural Networks for 3D Biomedical Image Segmentation.
Advances in Neural Information Processing Systems, 3036-3044.

42. Cheng, D., Wang, Y., 2017. Classification of MR brain images by combination of multi-
CNNs for AD diagnosis. International Conference on Digital Image Processing, p. 1042042.

43. Chupin, M., Gérardin, E., Cuingnet, R., Boutet, C., Lemieux, L., Lehéricy, S., Benali, H.,
Garnero, L., Colliot, O., 2009. Fully automatic hippocampus segmentation and classification in
Alzheimer's disease and mild cognitive impairment applied on data from ADNI. Hippocampus
19, 579–587.

44. Gerardin, E., Chételat, G., Chupin, M., Cuingnet, R., Desgranges, B., Kim, H.S.,

158
Niethammer, M., Dubois, B., Lehéricy, S., Garnero, L., 2009. Multidimensional classification
of hippocampal shape features discriminates Alzheimer's disease and mild cognitive impairment
from normal aging. Neuroimage 47, 1476.

45. Gray, K.R., Wolz, R., Keihaninejad, S., Heckemann, R.A., Aljabar, P., Hammers, A.,
Rueckert, D., 2011. Regional analysis of FDG-PET for use in the classification of Alzheimer'S
Disease. IEEE International Symposium on Biomedical Imaging: From Nano To Macro, pp.
1082-1085.

46. Herrup, K., 2011. Commentary on "Recommendations from the National Institute on
Aging-Alzheimer's Association workgroups on diagnostic guidelines for Alzheimer's disease."
Addressing the challenge of Alzheimer's disease in the 21st century. Alzheimers & Dementia
the Journal of the Alzheimers Association 7, 335.

47. Ho, A.J., Raji, C.A., Saharan, P., Degiorgio, A., Madsen, S.K., Hibar, D.P., Stein, J.L.,
Becker, J.T., Lopez, O.L., Toga, A.W., 2011. Hippocampal volume is related to body mass
index in Alzheimer’s disease. Neuroreport 22, 10-14.

48. Hosseini-Asl, E., Keynton, R., El-Baz, A., 2016. Alzheimer's Disease Diagnostics by
Adaptation of 3d Convolutional Network. 2016 Ieee International Conference on Image
Processing (Icip), 126-130.

49. Huang, G., Liu, Z., Weinberger, K.Q., 2016. Densely Connected Convolutional
Networks. CVPR.

50. Jack, C.R., Jr., Bernstein, M.A., Fox, N.C., Thompson, P., Alexander, G., Harvey, D.,
Borowski, B., Britson, P.J., J, L.W., Ward, C., Dale, A.M., Felmlee, J.P., Gunter, J.L., Hill,
D.L., Killiany, R., Schuff, N., Fox-Bosetti, S., Lin, C., Studholme, C., DeCarli, C.S., Krueger,
G., Ward, H.A., Metzger, G.J., Scott, K.T., Mallozzi, R., Blezek, D., Levy, J., Debbins, J.P.,
Fleisher, A.S., Albert, M., Green, R., Bartzokis, G., Glover, G., Mugler, J., Weiner, M.W.,
2008. The Alzheimer's Disease Neuroimaging Initiative (ADNI): MRI methods. Journal of
magnetic resonance imaging : JMRI 27, 685-691.
51. Jr, J.C., Albert, M.S., Knopman, D.S., Mckhann, G.M., Sperling, R.A., Carrillo, M.C.,
Thies, B., Phelps, C.H., 2011. Introduction to the recommendations from the National Institute

159
on Aging-Alzheimer's Association workgroups on diagnostic guidelines for Alzheimer's
disease. Alzheimers & Dementia the Journal of the Alzheimers Association 7, 257.
52. Kabani, N., MacDonald, D., Holmes, C.J., Evans, A., 1998. A 3D atlas of the human
brain. Neuroimage 7, S717.
53. Kim, Y., Kim, J., 2004. Gradient LASSO for feature selection. International Conference
on Machine Learning.
54. Korolev, S., Safiullin, A., Belyaev, M., Dodonova, Y., 2017. Residual and Plain
Convolutional Neural Networks for 3D Brain MRI Classification. IEEE International
Symposium on Biomedical Imaging.
55. Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. ImageNet classification with deep
convolutional neural networks. International Conference on Neural Information Processing
Systems, pp. 1097-1105.
56. Lécun, Y., Bottou, L., Bengio, Y., Haffner, P., 1998. Gradient-based learning applied to
document recognition. Proceedings of the IEEE 86, 2278-2324.
57. Leung, K.K., Barnes, J., Ridgway, G. R., Bartlett, J. W., Clarkson, M. J., & Macdonald,
K., et al. , 2010. Automated cross-sectional and longitudinal hippocampal volume measurement
in mild cognitive impairment and Alzheimer's disease. Neuroimage 51, 1345-1359.
58. Lian, C., Liu, M., Zhang, J., Shen, D., 2018. Hierarchical Fully Convolutional Network
for Joint Atrophy Localization and Alzheimer's Disease Diagnosis using Structural MRI. IEEE
Transactions on Pattern Analysis and Machine Intelligence PP, 1-1.
59. Lindberg, O., Walterfang, M., Looi, J.C., Malykhin, N., Ostberg, P., Zandbelt, B., Styner,
M., Paniagua, B., Velakoulis, D., Orndahl, E., 2012. Hippocampal shape analysis in
Alzheimer's disease and frontotemporal lobar degeneration subtypes. Journal of Alzheimers
Disease Jad 30, 355.
60. Liu, M., Cheng, D., Wang, K., Wang, Y., 2018a. Multi-Modality Cascaded Convolutional
Neural Networks for Alzheimer’s Disease Diagnosis. Neuroinformatics, 1-14.
61. Liu, M., Zhang, D., Shen, D., 2012. Ensemble sparse classification of Alzheimer's
disease. Neuroimage 60, 1106-1116.
62. Liu, M., Zhang, J., Adeli, E., Shen, D., 2017. Landmark-based deep multi-instance
learning for brain disease diagnosis. Medical Image Analysis 43, 157.
63. Liu, M., Zhang, J., Nie, D., Yap, P.T., Shen, D., 2018b. Anatomical Landmark based

160
Deep Feature Representation for MR Images in Brain Disease Diagnosis. IEEE Journal of
Biomedical & Health Informatics PP, 1-1.
64. Liu, S., Liu, S., Cai, W., Pujol, S., Kikinis, R., Feng, D., 2014. Early diagnosis of
Alzheimer's disease with deep learning. IEEE International Symposium on Biomedical
Imaging.
65. Milletari, F., Navab, N., Ahmadi, S.A., 2016. V-Net: Fully Convolutional Neural
Networks for Volumetric Medical Image Segmentation. Fourth International Conference on 3d
Vision, pp. 565-571.
66. Morey, R.A., Petty, C.M., Xu, Y., Hayes, J.P., II, H.R.W., Lewis, D.V., LaBar, K.S.,
Styner, M., McCarthy, G., 2009. A comparison of automated segmentation and manual tracing
for quantifying hippocampal and amygdala volumes. Neuroimage 45, 855-866.
67. Ng, Y.H., Hausknecht, M., Vijayanarasimhan, S., Vinyals, O., Monga, R., Toderici, G.,
2015. Beyond short snippets: Deep networks for video classification. 16, 4694-4702.
68. Ortiz, A., Jorge, M., Gorriz, J.M., Ramírez, J., 2016. Ensembles of Deep Learning
Architectures for The Early Diagnosis of Alzheimer's Disease. International Journal of Neural
Systems 26, 1650025.
69. Platero, C., Tobar, M.C., 2016. A fast approach for hippocampal segmentation from T1-
MRI for predicting progression in Alzheimer's disease from elderly controls. Journal of
Neuroscience Methods 270, 61-75.
70. Shen, D., Davatzikos, C., 2002. HAMMER: hierarchical attribute matching mechanism
for elastic registration. IEEE Transactions on Medical Imaging 21, 1421.
71. Shen, K.K., Fripp, J., Mériaudeau, F., Chételat, G., Salvado, O., Bourgeat, P., 2012.
Detecting global and local hippocampal shape changes in Alzheimer's disease using statistical
shape models. Neuroimage 59, 2155-2166. Silveira M, M.J., 2015. Boosting Alzheimer disease
diagnosis using PET images. 20th IEEE International Conference on Pattern Recognition
(ICPR),, pp. 2556-2559.
72. Simonyan, K., Vedaldi, A., Zisserman, A., 2013. Deep Inside Convolutional Networks:
Visualising Image Classification Models and Saliency Maps. Computer Science. Simonyan, K.,
Zisserman, A., 2014. Very Deep Convolutional Networks for Large-Scale Image Recognition.
Computer Science.
73. Sled, J.G., Zijdenbos, A.P., Evans, A.C., 1998. A nonparametric method for automatic

161
correction of intensity nonuniformity in MRI data. IEEE Transactions on Medical Imaging 17,
87.

74. Suk, H.I., Lee, S.W., Shen, D., 2015. Latent feature representation with stacked auto-
encoder for AD/MCI diagnosis. Brain Structure & Function 220, 841-859.

75. Thyreau, B., Sato, K., Fukuda, H., Taki, Y., 2018. Segmentation of the hippocampus by
transferring algorithmic knowledge for large cohort processing. Medical Image Analysis 43,
214-228.

76. Wang, X., Gao, L., Song, J., Shen, H., 2017. Beyond Frame-level CNN: Saliency-Aware
3-D CNN With LSTM for Video Action Recognition. Ieee Signal Processing Letters 24, 510-
514.

77. Wang, Y., Nie, J., Yap, P.T., Shi, F., Guo, L., Shen, D., 2011. Robust deformable-surface-
based skull-stripping for large-scale studies. Med Image Comput Comput Assist Interv 14, 635-
642.

78. Xin, L., Hong, X., Zhen, Z., Tong, L., 2012. 3D texture analysis of hippocampus based on
MR images in patients with alzheimer disease and mild cognitive impairment. Journal of
Beijing University of Technology 38, 942-948.

79. Ye, T., Zu, C., Jie, B., Shen, D., Zhang, D., 2016. Discriminative multi-task feature
selection for multi-modality classification of Alzheimer's disease. Brain Imaging & Behavior
10, 739.

80. Zhan, L., Zhou, J., Wang, Y., Jin, Y., Jahanshad, N., Prasad, G., Nir, T.M., Leonardo,
C.D., Ye, J., Thompson, P.M., 2015. Comparison of nine tractography algorithms for detecting
abnormal structural brain networks in Alzheimer's disease. Frontiers in Aging Neuroscience 7,
48.

81. Zhang, D., Wang, Y., Zhou, L., Yuan, H., Shen, D., 2011. Multimodal classification of
Alzheimer's disease and mild cognitive impairment. Neuroimage 55, 856.

82. Zhang, J., Gao, Y., Gao, Y., Munsell, B., Shen, D., 2016. Detecting Anatomical
Landmarks for Fast Alzheimer's Disease Diagnosis. IEEE Transactions on Medical Imaging 35,

162
2524-2533.

83. Zhang, J., Liu, M., An, L., Gao, Y., Shen, D., 2017. Alzheimer's Disease Diagnosis using
Landmark-based Features from Longitudinal Structural MR Images. IEEE Journal of
Biomedical & Health Informatics 21, 1607.

84. Zhang, Y., Brady, M., Smith, S., 2001. Segmentation of brain MR images through a
hidden Markov random field model and the expectation-maximization algorithm. IEEE Trans
Med Imaging 20, 45-57.

85. Zhu, X., Suk, H.I., Zhu, Y., Thung, K.H., Wu, G., Shen, D., 2015. Multi-view
Classification for Identification of Alzheimer’s Disease. Machine Learning in Medical Imaging
Mlmi Author 9352, 255.

86. Diamond E, Miller S, Rentz D, Sperling R, Dickerson B, Atri A, Sperling RA.

Relationship of fMRI activation to clinical trial memory measures in Alzheimer disease.
Neurology 2007; 69: 1331-1341.
87. Pah-Lavan Z. Alzheimer's disease: the road to oblivion. J Commun Nursing 2006; 20: 4-
10. Pies RW. Alzheimer's Redux: A preliminary take on the new diagnostic criteria. Psychiatric
Times 2012; 4: 24-30.
88. Ashburner J, Friston KJ. Voxel-Based Morphometry-the methods. Neuroimage 2000; 11:
805-821.
89. Wolz R, Julkunen V, Koikkalainen J, Niskanen E, Zhang DP, Rueckert D, Soininen H,
Lötjönen J; Alzheimer's Disease Neuroimaging Initiative. Multi-method analysis of MRI
images in early diagnostics of Alzheimer's disease. PLoS One 2011; 6: e25446.
90. Mangin JF, Rivière D, Cachia A, Papadopoulos-Orfanos D, Collins DL, Evans AC, Régis
J. Object-based strategy for morphometry of the cerebral cortex. Inf Process Med Imaging
2003; 18: 160-171.
91. Toews M, Wells W, Collins DL, Arbel T. Feature-Based Morphometry: Discovering
Group-related Anatomical Patterns. Neuro Image 2010; 49: 2318-2327.
92. Mrudula,D. and G.N. Lakshmi, Joint DWT, SWT based MRI image enhancement
technique and denosing based on Gaussian filter,International Journal of Electronics &
Communication (IIJEC), 2014, 2(9): 64-69.

163
93. Jumah,A.A., Denoising of an Image Using Discrete Stationary Wavelet Transform and
Various Thresholding Techniques, Journal of Signal and Information Processing, 2013, 4:33-
41.
94. Robert Manzke, Carsten Meyer, Olivier Ecabert, Jochen Peters, Niels J. Noordhoek,
Aravinda Thiagalingam, Vivek Y. Reddy, Raymond C Chan, and Jürgen Wee, “Automatic
Segmentation of Rotational X-Ray Images for Anatomic Intra-Procedural Surface Generation in
Atrial Fibrillation Ablation Procedures”, IEEE, . 0278-0, 2010.
95. Liu, X., Chen, X., Zheng, W., Xia, M., Han, Y., Song, H., 862 Wang, Z. (2018). Altered
functional connectivity of insular subregions 863 in Alzheimer’s disease. Frontiers in aging
neuroscience, 10, 107. 864
77. Power, J. D., Cohen, A. L., Nelson, S. M., Wig, G. S., Barnes, 865 K. A., Church, J. A.,
Petersen, S. E. (2011). Functional network 866 organization of the human brain. Neuron, 72(4),
665-678
96. Isa,I. S., Sulaiman,S.N., Mustapha, M.and N. K. A. Karim, Automatic contrast
enhancement of brain MR images using Average Intensity Replacement based on Adaptive
Histogram Equalization (AIR-AHE), Biocybernetics and Biomedical Engineering, 2017, 37:
24-34.
97. Peter A. Freeborough and Nick C. Fox, “MR Image Texture Analysis Applied to the
Diagnosis and Tracking of Alzheimer’s Disease”, IEEE, 0278–0062, 2014.
98. Attali, D., Montanvert, A.,“Semicontinuous skeletons of 2D and 3D shapes”, proceedings
from the Second International Workshop onVisual Form pp 32 – 41, 1994.
99. T. Kalaiselvi, P.Sriramakrishnan and K. Somasundaram, “Brain Abnormality Detection
from MRI of Human Head Scans using the Bilateral Symmetry Property and Histogram
Similarity Measures”, IEEE, 978-1-5090-4420-7, 2016.
100. Ms. Suchita Goswami, Mr. Lalit Kumar P. Bhaiya, “A Hybrid Neuro-Fuzzy Approach for
Brain Abnormality Detection Using GLCM based Feature Extraction”, IEEE, 2013.
101. Gao ting, Kamil Moydin, Askar Hamdulla, “An Overview Of Feature Extraction Methods
For Handwritten Image Retrieval”, IEEE, 978-1-7281-1366-1, 2018.
102. Akshaya Pisal, Ravindra Sor, Dr.K.S.Kinage, “Facial Feature Extraction Using
Hierarchical MAX(HMAX) Method”, IEEE, 978-1-5386-4008-1, 2017.

164
103. A. Keleş A. Keleş, and U. Yavuz, “Expert system based on neuro-fuzzy rules for
diagnosis breast cancer”, Expert Syst. Appl. 38(5), pp. 5719-5726, 2011.
104. Sakthivel K, Jayanthiladevi A, Kavitha C. “Automatic detection of lung cancer nodules
by employing intelligent fuzzy cmeans and support vector machine”. Biomed Res, P: 27, 2016
105. Kamil A., Rustamov S., Clements M.A., Mustafayev E. “Adaptive Neuro-Fuzzy
Inference System for Classification of Texts”, Studies in Fuzziness and Soft Computing, vol
361. Springer, 2018.
106. Karaboga, D. & Kaya, E. “Adaptive network based fuzzy inference system (ANFIS)
training approaches: a comprehensive survey”, Artificial Intelligence Review, Springer, 2018.
107. Chiaki Tokunaga, Hidetaka Arimura, T Akashi Yoshiura Yasuo Yamashita, Taiki
Magome, Hiroshi Honda Hideki Hirata , Fukai Toyofuku, Masafumi Ohki, “Fuzzy -Based
Segmentation Of Brain Parenchymal Regions With Alzheimer's Disease Into Cerebral Cortex
And White Matter In 3.0-T Magnetic Resonance Images”, IEEE, 2014.
108. Ain, Q.,Mehmood,I.,Naqi,S.M.and M.A.Jaffar, Bayesian Classification Using DCT
Features for Brain Tumor Detection, International Conference on Knowledge-Based and
Intelligent Information and Engineering Systems,ISBN:3-642-15386-0 978-3-642-15386-0,
September 2010, Cardiff, UK:340-349.
109. Alzheimer’s Association. Alzheimer's disease and dementia. http://www.alz.org/;
2015. [accessed April 5, 2015].
110. Liu S, Cai W, Wen L, Feng DD, Pujol S, Kikinis R, et al. Multi-channel
neurodegenerative pattern analysis and its application in Alzheimer’s disease
characterization. Comput Med Imaging Graph 2014;38:436–44. http://dx.doi.
org/10.1016/j.compmedimag.2014.05.003.
111. Weiner MW, Veitch DP, Aisen PS, Beckett LA, Cairns NJ, Green RC, et al. 2014
Update of the Alzheimer’s Disease Neuroimaging Initiative: a review of papers published
since its inception. Alzheimers Dement 2015;11:e1-120. http://dx.doi.
org/10.1016/j.jalz.2014.11.001.
112. Hanyu H, Sato T, Hirao K, Kanetaka H, Iwamoto T, Koizumi K. The progression
of cognitive deterioration and regional cerebral blood flow patterns in Alzheimer’s
disease: a longitudinal SPECT study. J Neurol Sci 2010;290:96–101. http://dx.doi.
org/10.1016/j.jns.2009.10.022.

165
 113. Gray KR, Wolz R, Heckemann RA, Aljabar P, Hammers A, Rueckert D.
Multiregion analysis of longitudinal FDG-PET for the classification of Alzheimer’s
disease. Neuroimage 2012;60:221–9. http://dx.doi.org/10.1016/j.neuroimage.
2011.12.071.
114. Chen YJ, Deutsch G, Satya R, Liu HG, Mountz JM. A semi-quantitative method
for correlating brain disease groups with normal controls using SPECT: Alzheimer’s
disease versus vascular dementia. Comput Med Imaging Graph 2013;37:40–7.
http://dx.doi.org/10.1016/j.compmedimag.2012.11.001.
115. Górriz JM, Segovia F, Ramírez J, Lassl A, Salas-Gonzalez D. GMM based
SPECT image classification for the diagnosis of Alzheimer’s disease. Appl Soft Comput
2011;11:2313–25. http://dx.doi.org/10.1016/j.asoc.2010.08.012.
116. Zhang D, Wang Y, Zhou L, Yuan H, Shen D. Multimodal classification of
Alzheimer’s disease and mild cognitive impairment. Neuroimage 2011;55: 856–67.
http://dx.doi.org/10.1016/j.neuroimage.2011.01.008.
117. Attali, D., Montanvert, A., 1994. “Semicontinuous skeletons of 2D and 3D
shapes”, proceedings from the Second International Workshop on Visual Form pp 32 – 41.
118. Bedi, S.S. and R. Khandelwal, Contrast Enhancement for PCA Fusion of Medical Images,
Journal of Global Research in Histogram Computer Science,2013, 4(3): 25-29.
119. Aguilar C, Westman E, Muehlboeck JS, Mecocci P, Vellas B, Tsolaki M, et al. Different
multivariate techniques for automated classification of MRI data in Alzheimer’s disease and
mild cognitive impairment. Psychiatry Res Neuroimaging 2013;212:89–98.
http://dx.doi.org/10.1016/j.pscychresns.2012.11.005.
120. Westman E, Muehlboeck JS, Simmons A. Combining MRI and CSF measures for
classification of Alzheimer’s disease and prediction of mild cognitive impairment conversion.
Neuroimage 2012;62:229–38. http://dx.doi.org/10.1016/j.neuroimage.2012.04.056.
121. Li M, Qin Y, Gao F, Zhu W, He X. Discriminative analysis of multivariate features from
structural MRI and diffusion tensor images. Magn Reson Imaging 2014;32: 1043–51.
http://dx.doi.org/10.1016/j.mri.2014.05.008.
122. Moradi E, Pepe A, Gaser C, Huttunen H, Tohka J. Machine learning framework for early
MRI-based Alzheimer’s conversion prediction in MCI subjects. Neuroimage 2015;104:398–
412. http://dx.doi.org/10.1016/j.neuroimage.2014.10.002.

166
123. Bron EE, Smits M, van der Flier WM, Vrenken H, Barkhof F, Scheltens P, et al.
Standardized evaluation of algorithms for computer-aided diagnosis of dementia based on
structural MRI: the CADDementia challenge. Neuroimage 2015;111: 562–79.
http://dx.doi.org/10.1016/j.neuroimage.2015.01.048.
124. Andersen AH, Rayens WS, Liu Y, Smith CD. Partial least squares for discrimination in
fMRI data. Magn Reson Imaging 2012;30:446–52. http://dx.
doi.org/10.1016/j.mri.2011.11.001.
125. Fan Y, Resnick SM, Wu X, Davatzikos C. Structural and functional biomarkers of
prodromal Alzheimer’s disease: a high-dimensional pattern classification study. Neuroimage
2008;41:277–85. http://dx.doi.org/10.1016/j.neuroimage.2008.02.043.
126. Dinesh E, Kumar MS, Vigneshwar M, Mohanraj T. Instinctive classification of
Alzheimer’s disease using FMRI, pet and SPECT images. 7th Int Conf Intell Syst Control;
2013. p. 405–9. http://dx.doi.org/10.1109/ISCO.2013.6481189.
127. Qi Zhou, Mohammed Goryawala, Mercedes Cabrerizo, Jin Wang, Warren Barker, David
Loewenstein, Ranjan Duara and Malek Adjouadi, “An Optimal Decisional Space for the
Classification of Alzheimer’s Disease and Mild Cognitive Impairment”, IEEE, 0018-9294,
2013.
128. Satheesh, S., Reddy,A.S. and K.V. Prasad, “Noise suppression using Wiener filtering in
the nonsubsampledcontourlet domain for Magnetic Resonance brain images”, IEEE, DOI:
10.1109/IHCI.2012.6481836, 27-29, 2012.
129. Fred,A.L. and S.N. Kumar, “A Novel Filter Selection for Efficient and Robust
Segmentation in MR Human Brain Images Through Comparative Analysis of Various Spatial
and Transform Domain Filters”, IEEE, 3(5):582-586, 2015
130. Ashburner J, Friston KJ. Voxel-based morphometry—the methods. Neuroimage
2000;11:805–21. http://dx.doi.org/10.1006/nimg.2000.0582.
131. Nakatsuka T, Imabayashi E, Matsuda H, Sakakibara R, Inaoka T, Terada H.
Discrimination of dementia with Lewy bodies from Alzheimer’s disease using voxel-based
morphometry of white matter by statistical parametric mapping 8 plus diffeomorphic anatomic
registration through exponentiated Lie algebra. Neuroradiology 2013;55:559–66.
http://dx.doi.org/10.1007/s00234-013- 1138-9.

167
132. Dinesh E, Kumar MS, Vigneshwar M, Mohanraj T. Instinctive classification of
Alzheimer’s disease using FMRI, pet and SPECT images. 7th Int Conf Intell Syst Control;
2013. p. 405–9. http://dx.doi.org/10.1109/ISCO.2013.6481189.
133. Westman E, Muehlboeck JS, Simmons A. Combining MRI and CSF measures for
classification of Alzheimer’s disease and prediction of mild cognitive impairment conversion.
Neuro image 2012;62:229–38. http://dx.doi.org/10.1016/j.neuroimage.2012.04.056.
134. Zhou Q, Goryawala M, Cabrerizo M, Wang J, Barker W, Loewenstein DA, et al. An
optimal decisional space for the classification of Alzheimer's disease and mild cognitive
impairment. Biomed Eng IEEE Trans 2014;61:2245–53. http://dx.doi.
org/10.1109/TBME.2014.2310709.
135. Hinrichs C, Singh V, Xu G, Johnson SC. Predictive markers for AD in a multi-modality
framework: an analysis of MCI progression in the ADNI population. Neuroimage 2011;55:574–
89. http://dx.doi.org/10.1016/j.neuroimage.2010.10.081.
136. R.Brijesh, Dr.G.Ravindran, “A Spiking neural network of the CA3 of the Hippocampus
can be a Neural Prosthesis for lost cognitive functions”, IEEE, 1-4244-0788-5, 2007.
137. Nakatsuka T, Imabayashi E, Matsuda H, Sakakibara R, Inaoka T, Terada H.
Discrimination of dementia with Lewy bodies from Alzheimer’s disease using voxel-based
morphometry of white matter by statistical parametric mapping 8 plus diffeomorphic anatomic
registration through exponentiated Lie algebra. Neuroradiology 2013;55:559–66.
http://dx.doi.org/10.1007/s00234-013- 1138-9.
138. Matsuda H, Mizumura S, Nemoto K, Yamashita F, Imabayashi E, Sato N, et al.
Automatic voxel-based morphometry of structural MRI by SPM8 plus diffeomorphic anatomic
registration through exponentiated lie algebra improves the diagnosis of probable Alzheimer
disease. Am J Neuroradiol 2012;33:1109–14. http://dx.doi.org/10.3174/ajnr.A2935.
139. Liu M, Zhang D, Shen D. Ensemble sparse classification of Alzheimer’s disease.
Neuroimage 2012;60:1106–16. http://dx.doi.org/10.1016/j.neuroimage.2012. 01.055.
140. Chaves R, Ramírez J, Górriz JM, López M, Salas-Gonzalez D, Álvarez I, et al.
SVMbased computer-aided diagnosis of the Alzheimer’s disease using t-test NMSE feature
selection with feature correlation weighting. Neurosci Lett 2009;461: 293–7.
http://dx.doi.org/10.1016/j.neulet.2009.06.052.

168
141. Wang D, Zhang H, Liu R, Lv W, Wang D. T-test feature selection approach based on
term frequency for text categorization. Pattern Recognit Lett 2014;45:1–10.
http://dx.doi.org/10.1016/j.patrec.2014.02.013.
142. Ashburner J, Friston KJ. Voxel-based morphometry—the methods. Neuroimage
2000;11:805–21. http://dx.doi.org/10.1006/nimg.2000.0582.
143. Nakatsuka T, Imabayashi E, Matsuda H, Sakakibara R, Inaoka T, Terada H.
Discrimination of dementia with Lewy bodies from Alzheimer’s disease using voxel-based
morphometry of white matter by statistical parametric mapping 8 plus diffeomorphic anatomic
registration through exponentiated Lie algebra. Neuroradiology 2013;55:559–66.
http://dx.doi.org/10.1007/s00234-013- 1138-9.
144. Ashburner J. A fast diffeomorphic image registration algorithm. Neuroimage 2007;38:95–
113. http://dx.doi.org/10.1016/j.neuroimage.2007.07.007.
145. Haq NF, Kozlowski P, Jones EC, Chang SD, Goldenberg SL, Moradi M. A datadriven
approach to prostate cancer detection from dynamic contrast enhanced MRI. Comput Med
Imaging Graph 2015;41:37–45. http://dx.doi.org/10.1016/j. compmedimag.2014.06.017.
146. Tang L, Wen Y, Zhou Z, von Deneen KM, Huang D, Ma L. Reduced field-of-view DTI
segmentation of cervical spine tissue. Magn Reson Imaging 2013;31: 1507–14.
http://dx.doi.org/10.1016/j.mri.2013.07.003.
147. Prati RC. Combining feature ranking algorithms through rank aggregation. Proc Int Jt
Conf Neural Networks 2012;10–5. http://dx.doi.org/10.1109/IJCNN.2012. 6252467.
148. Kamkar I, Gupta SK, Phung D, Venkatesh S. Stable feature selection for clinical
prediction: exploiting ICD tree structure using Tree-Lasso. J Biomed Inform 2014;53:277–90.
http://dx.doi.org/10.1016/j.jbi.2014.11.013.
149. Gao Q, Liu J, Zhang H, Hou J, Yang X. Enhanced Fisher discriminant criterion for image
recognition. Pattern Recognit 2012;45:3717–24. http://dx.doi.org/10.
1016/j.patcog.2012.03.024.
150. Dimitrovski I, Kocev D, Kitanovski I, Loskovska S, Džeroski S. Improved medical image
modality classification using a combination of visual and textual features. Comput Med
Imaging Graph 2015;39:14–26. http://dx.doi.org/10.1016/j.compmedimag.2014.06.005.

169
151. Al-Kadi OS. A multiresolution clinical decision support system based on fractal model
design for classification of histological brain tumours. Comput Med Imaging Graph
2014;41:67–79. http://dx.doi.org/10.1016/j.compmedimag. 2014.05.013.
152. Xue T, Bai L, Chen S, Zhong C, Feng Y,Wang H, et al. Neural specificity of acupuncture
stimulation from support vector machine classification analysis. Magn Reson Imaging
2011;29:943–50. http://dx.doi.org/10.1016/j.mri.2011.03.003. 262 I. Beheshti, H. Demirel /
Magnetic Resonance Imaging 34 (2016) 252–263
153. Song X, Chen NK. A SVM-based quantitative fMRI method for resting-state functional
network detection. Magn Reson Imaging 2014;32:819–31. http://dx.
doi.org/10.1016/j.mri.2014.04.004.
154. Van Der Heijden VF, De Ridder D. Classification, parameter estimation, and state
estimation: an engineering approach using MATLAB, vol. 32. John Wiley and Sons; 2004.
155. Narasimhamurthy A. Theoretical bounds of majority voting performance for a binary
classification problem. IEEE Trans Pattern Anal Mach Intell 2005;27: 1988–95.
http://dx.doi.org/10.1109/TPAMI.2005.249.

170
 LIST OF PUBLICATIONS
[1]. Viswanathan, R. and K. Perumal, Support Vector Machine based classification into Groups
using Disriminant Function, International Journal of Innovative Research in Computer and
Communication Engineering (IJIRCCE), ISSN No: 2320-9798, DOI: 10.15680 /
IJIRCCE.2016.0411201, Page No: 20316 – 20322.
[2]. Viswanathan, R. and K. Perumal, Support Vector Machine based classification of AD on
Bootstrap Method, International Journal of Innovative Research in Science, Engineering and
Technology (IJIRSET), ISSN No: 2347-6710, DOI: 10.15680 / IJIRSET.2016.0512081, Page
No: 21313 – 21318.
[3]. Viswanathan, R. and K. Perumal, Detection and Classification of Alzheimer’s Disease in
Magnetic Resonance Imaging using Region Growing, proc: International Conference on
Artificial Intelligence and Evolutionary Computations in Engineering System (ICAIECES 2017)
and Power, Circuit and Information Technologies (ICPCIT 2017), Madanapalle, Andhra
Pradesh, India, April 2017, (paper presented) and published the paper in Scopus indexed journal
on International Journal of Applied Engineering Research (IJAER), ISSN No: 0973-4562, Page
No:47-53.
[4]. Viswanathan, R. and K. Perumal, Evaluation and Classification of Alzheimer’s Disease in
Magnetic Resonance Images on Hippocampus Region using support vector machines, proc:
IEEE International Conference on Power, Control, Signals and Instrumentation Engineering
(ICPCSI 2017), Saveetha Engineering College, Thandalam, Chennai, (paper presented) and the
paper published in IEEE Digital Library, INSPEC No: 17860956, DOI:
10.1709/icpcsi.2017.8391821, Page No: 144-150.
[5]. Viswanathan, R and K. Perumal, Segmentation Technique to detect the Alzheimer’s Disease
in hippocampus using Region Growing, International Journal of Scientific research in computer
science and engineering (IJSRCSE), the paper published in IEEE citation, MLA citation, APA
citation, Google Scholar and DPI digital library, ISSN No: 2320-7639, DOI:
15680/IJSRCSE.208.0411201, Page No: 55-59.
[6]. Viswanathan, R. and K. Perumal, Classification of Alzheimer’s Disease from Magnetic
Resonance Image for Feature Extraction using Discrete Wavelet Transform, proc: International
Conference on Trends in Electronic and Informatics (ICOEI 2018), Scad College of Engineering

171
and Technology, Tiruneveli, (paper presented) and published the paper in IEEE digital library,
IEEE xplore ISBN: 978-1-5386-3570-4, Page No: 584-592.
[7]. Viswanathan, R. and K. Perumal, Segmentation for hippocampus in Alzheimer’s Disease
using Region Growing and Level Set, proc: International conference on ISMAC 2018, Scad
college of engineering and technology, palladam, Coimbatore, (paper presented) and the paper
published in Springer lecture series ISBN No: 978-3-030-00664-8.
[8]. Viswanathan, R. and K. Perumal, Ranking based classification of Alzheimer’s Disease in
hippocampus region from structural magnetic resonance imaging using support vector machine,
IPASJ, International journal of Computer Science (IJCS), the paper published in Science Citation
Journal, Thomson Reuters, ISSN No:2321-5992, Page No:01-14.
[9]. Viswanathan, R. and K. Perumal, Segmentation in Hippocampus for Alzheimer’s affliction
utilizing Region Growing and Level Set Technique, the paper has published in Scopus indexed
journal on International Journal of Advanced Research in Engineering and Technology –
IJARET, Volume – 11, Issue – 7, Page No: 489 – 495, July 2020, published by IAEME
Publication.

172