International Journal of

INTELLIGENT SYSTEMS AND APPLICATIONS IN

ENGINEERING
ISSN:2147-67992147-6799 www.ijisae.org Original Research Paper

Cad System for the Detection and Classification of Alzheimers from MR

Images Using Deep Learning Techniques
Nithya V. P.1, Dr. N. Mohanasundaram2 and Dr. R. Santhosh3

Submitted: 21/08/2022 Accepted: 23/11/2022

Abstract:Alzheimer's disease, also known as Alzheimer's degeneration, is an irreversible brain disorder affecting over 65-year-olds.AD
cannot currently be cured, but its progression can be slowed by certain treatments.It is imperative to diagnose AD accurately and early to
improve patient care and future treatment options.Many ways evaluate picture features to retrieve handmade characteristics and then
design a classifier to differentiate AD from other groups. The Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset is used to
propose a new method for categorizing AD and MCI.Preprocessing of brain pictures, including image segmentation, is essential for these
methods.To reduce noise and factor fluctuations, the pictures are preprocessed using several complex methods.The preprocessed images
are then used to segment the images.This study presents a novel classification algorithm based on deep learning and ResNet 50, inspired
by deep learning's success in image analysis.The suggested classifier approach allows the selection of relevant characteristics while
avoiding overfitting issues.When the proposed algorithm was compared to the existing models, the results proved that it had better
accuracy values.

Keywords: CLAHE, BADF, K-means clustering, ResNet 50.

1. Introduction improve the quality of life of an AD patient even in their old

The alarming statistics about the prevalence of AD are also an age. Along with this, classifying the cognitive impairment of
important factor for focusing the attention on predicting and the patients in advance helps the physician provide the required
finding AD patients. Alzheimer's disease, a chronic and optimal treatment for dealing with the disease. Figure 1
neurological illness, causes irreversible cognitive function loss shows the existence of AD on neuron cells.
and dementia. There is no cure for this ailment, and its
aetiology is unknown. According to Alzheimer's disease facts
and numbers 2022 [1], 6.5 million senior citizens will have
Alzheimer's disease in 2022.They're all over 75, with 73% of
them being over 75. Alzheimer's disease can affect around one
out of every nine seniors over the age of 65. (10.7 per cent).The
report is also giving alarming information about the high
mortality rate of AD patients, especially among the aged
population around the globe. The presence of AD can make a
significant negative impact on the health of an aged person who
is already suffering from other severe diseases. However, the Figure 1:Neurons in the brain display microscopic plaques and
report emphasizes the early detection of AD for helping the tangles in Alzheimer's illness
physicians cope with this because it helps them in designing Computer-aided systems are very helpful in diagnosing medical
appropriate and better medication strategies like prescribing images, which is otherwise a difficult task.The most popular
appropriate drugs for reducing the rate of progression of AD imaging techniques for diagnosing neurodegenerative disorders
[2]. Moreover, the related family and friends of the AD patient include Computer Tomography (CT), Magnetic Resonance
can give better care to them. Such pre-planned strategies Imaging (MRI), and positron emission tomography (PET).
Among these, CT scanning takes a shorter time and the images
1
Research Scholar, Department of Computer Science and Engineering, are clearer.Various diseases can be examined using CT
Karpagam Academy of Higher Education,Coimbatore, Tamil Nadu, India, scanning.MCI(Mild Cognitive Impairment), could be mistaken
[email protected]
2
for a natural result of ageing due to its limited resolution in the
Professor, Department of Computer Science and Engineering,Faculty 0f
middle lobe.With Pet, you'll get improved sensitivity and
Engineering, Karpagam Academy of Higher Education, Coimbatore,Tamil
Nadu, India, [email protected] resolution, and your images won't be absorbed by tissue, as
3
Professor, Department of Computer Science and Engineering,Faculty 0f there is no tissue absorption [3].The inside architecture of the
Engineering, Karpagam Academy of Higher Education, Coimbatore,Tamil human body is imaged using an MRI scanner. With a rapidly
Nadu, India, [email protected] changing gradient magnetic field, MRI can be speeded up
significantly, and good soft tissue resolution can be achieved
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2022, 10(2)s, 96–104 | 96
without ionizing radiation.In picture analysis, feature extraction subject's physique will certainly change to some degree. To get
plays a crucial role.Several previously studied regions such as the actual picture, a series of preprocessing processes must be
the amygdala, hippocampus and other brain regions must be done to satisfy the demand for feature extraction, feature
manually removed [4].Because manual labour is required, there selection, and classification.Preprocessing is essential before
are restrictions as well as the possibility of human error. data operations in most studies, particularly in machine
Because of the rapid advancement of artificial intelligence, learning. Some research may not require preprocessing if deep
computer vision has become increasingly crucial in diagnosing learning is used. The majority of the research still uses raw data
Alzheimer's disease. Deep learning is a popular machine preprocessing processes such as intensity normalization,
learning branch that can solve problems that old techniques registration, skull dissection, and motion correction [9,10].
can't. In recent times, deep learning technology has dominated Researchers Gao and Lio [11] examined deep learning
medical imaging, and it has been effectively employed to algorithms for detecting Alzheimer's illness. The authors
extract the features from medical pictures to identify described biomarkers and features for detecting the disease, as
Alzheimer's illness [5]. Alzheimer's illness is anticipated to well as methods for detecting it.As per the research, deep
grow more widespread by 2022, according to Figure 2 of learning technology is effective for detecting Alzheimer's.Heba
Alzheimer's illness Facts and Figures, 2020. et al. [12] employ deep learning methods to identify END. The
system is made up of four different processes: transfer learning,
deep feature extraction, feature reduction, and classification.
The fusion of features is crucial to the framework's success.A
method of analyzing embryonic MRI images at various
gestations was found to be effective at diagnosing ENDs
(embryonic neuro-developmental disorders).The results are
then compared to comparable studies that employed embryonic
photos to confirm the efficacy of the suggested framework.The
proposed application's efficiency was comparable to that of
other frameworks. As a result, the proposed framework may be
utilized to detect END in a step-by-step fashion. For AD
Figure 2. In Alzheimer's illness Facts and Figures, 2022, the prediction, Fei et al. [13] build a deep CNN based on a multiple
prevalence of Alzheimer's illness is reported. attention approach. First, cyclic convolution is used to improve
Importance of DL algorithms in AD classification the feature information in the MRI picture. Second, a multiple
DL algorithms are used to get more insights from the health attention technique is used to re-calibrate features and
care data [6]. Likewise, the application of DL algorithms helps adaptively learn feature weight to discover brain regions that
a physician by assisting in the crucial decision-making process are particularly important for disease detection. Lastly, for the
for distinguishing MCI (Mild Cognitive Impairment), AD, and backbone network, an enhanced VGG model is given. The
HC (Healthy Controls) [7]. The Healthy Controls (HC) are the accuracy, sensitivity, and specificity of AD prediction are all
individuals who do not have any cognitive impairment. 99.8%, 99.9%, and 99.8%, respectively, according to the
However, some HCs would have Mild Cognitive Impairment trial.Loddo et al. [14] conducted Alzheimer's disease studies
(MCI) at some stage of their life. MCI is a state of cognitive and research using computer-aided diagnosis while assessing
impairment where the individual can do daily activities like numerous deep learning models (CAD). Based on patient brain
bathing, and brushing without difficulty, but, MCI patients face scans, the suggested method is designed to give significant
difficulty in solving logical problems, and recalling historical support for clinical care.
or past events. As the exact reasons responsible for developing
AD is unknown, the physician infers whether the patients are Problem Statement
most likely to develop AD or MCI after observing the The major challenge faced in AD classification is
characteristics that are related to MRI, PET, SPECT medical preprocessing and feature extraction which results in an
images, cognitive memory assessment tests, CSF and inaccuracy classification of MRI images. The existing methods
demographic data. However, this is a challenging task for provide inaccurate results in segmentation as well as feature
physicians because it is difficult to find a pattern between the selection. Also, it has a problem along with computational
MCI and AD patients after observing these data [8]. This is complexity due to the number of iterations. Thus, the overall
where DL algorithms are deployed on these characteristic data MRI classification performance is reduced significantly.
of the patients for finding the hidden patterns and crucial Consequently, advanced methods for the removal of noise from
features that are playing a major role in distinguishing MCI, MRI images were developed, including preprocessing, feature
AD, and HC. The DL algorithms make use of information from extraction, feature selection, and classification.
various other domains such as statistics and mathematics to
accomplish this. Thus, the overall application of the DL The objective of the Research
algorithm in the AD domain makes a good impact on the The primary goal of this study is to create and implement
physician in assisting them. automated multiclass diagnostic detection techniques on MRI
scan pictures using image processing and data mining.The
2. Literature Review major objectives of this research work have been:
Since picture acquisition is done by machines made by various • To study in detail, research on Alzheimer's Disease &
manufacturers, the medical staff's operation procedures may Image Processing.
vary. Furthermore, due to the lengthy acquisition time, the
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2022, 10(2)s, 96–104 | 97
 • To analyze the existing Algorithms for detecting study is presented in part 3. The result analysis is presented in
Alzheimer's Disease from MRI Images. part 4, and the work is concluded in part 5.
• To develop an algorithm for preprocessing the MRI
images for the Noise Removal. 3. Proposed Work
• To design & develop an algorithm for effective A computer-assisted automatic detection system is offered as a
classification of Alzheimer's Disease. way of detecting Alzheimer's disease. Figure 3 depicts the
• To maximize the performance measures of whole detection technique. The original MRI picture is first
Preprocessing and Classification of Alzheimer's preprocessed, after which it is improved using CLAHE and
Disease. BADF, and finally turned into K-means clustering. The neural
• To determine the most effective improved picture for network classifier is then trained using deep learning-based
Alzheimer's disease detection. ResNet 50 structures to identify the picture as normal or
The following sections provide the framework for the paper: abnormal, using the trained characteristics as a guide. Finally,
Alzheimer's disease is explained in part 1 as an introduction. the MRI image categorization system's performance is
The literature reviews are explained in part 2, and the planned evaluated.

Contrast limited Contrast limited

Adaptive Histogram Adaptive Histogram
Equalization Equalization

Input MRI Pre-processing

images

Normal
ResNet-50 K-means Clustering

AD

Figure 3: Overall Block Diagram

a. Preprocessing Step 2: Depending on the grey scale in the above photo, make a
Pre-main processing's goal is to improve the picture's histogram for each contextual area.
uniformity. It removes or minimizes non-related and step 3: The contrast restricted histogram of the platform can be
unnecessary components in medical data to prepare them for calculated based on the CL score.
subsequent processing.
Contrast Limited Adaptive Histogram Equalisation 𝑁𝑎𝑣𝑔 = (𝑁𝑟𝑋 × 𝑁𝑟𝑌)/𝑁𝑔𝑟𝑎𝑦 (1)
In the first place, CLAHE was designed to aid in low-contrast Navg is the average number of pixels in the platform,N gray is
medical imaging.Regular AHE has different limitations.Noise the number of grayscale pixels, and NrX and NrY are the pixels
amplification was combated by implementing clipping per direction in the platform.
limitations on the CLAHENoise amplification was combated Step 4: Re-distribute the remaining pixels until none remain.
by implementing clipping limitations on the CLAHE.There are The following formula is used to determine the redistributing
multiple contextual components in the CLAHE picture, pixel stage:
including sub-images, tiles, and blocks. Two parameters are 𝑠𝑡𝑒𝑝 = 𝑁𝑔𝑟𝑎𝑦 /𝑁𝑟𝑒𝑚𝑎𝑖𝑛 (2)
critical in CLAHE: the Block Size (BS) and Clip Limit N_ pixels are still cut.A step is a one-digit positive integer. The
(CL).Image quality is determined by these two variables.Due to value indicates how many snipped pixels were left. A phase is
the low intensity of the input picture and the flattening effect of an integer that is at least one. Whilethe system scans from the
a larger CL, the picture appears brighter as CL is lowest to the highest grey level, as it did in the previous
increased.This dynamic range and visual contrast increase as step,the software will add one pixel to grey levels that are less
the BS grows larger. With the entropy of the picture [15, 16], than N CL.If there is unequal dispersion of pixels at the end of
the two factors set at the place with the greatest entropy the research, the algorithm will calculate a new path using
curvature give subjectively good picture quality. Eq.(2) and repeat the lookup until all leftover pixels are evenly
The phases in the CLAHE approach for enhancing the actual scattered.
photo are as follows:
Step 1: By converting the original pixel intensity into non- Step 5: Each region's brightness distribution is boosted by the
overlapping contextual zones.MN refers tothe "total number of Rayleigh transform.The accumulated probability
image tiles," and an 88 percent figure is optimum for 𝑃𝑖𝑛𝑝𝑢𝑡 (𝑖)function is employed on the trimmed histogram.When
preserving chromatic knowledge in photographs. the Rayleigh distribution is utilized, the underwater picture

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2022, 10(2)s, 96–104 | 98
looks to be more natural. The forward transform of Rayleigh is Gaussian mixture filtering, which provides a consistent filter in
calculated as follows: various situations. Both the expected and posterior densities of
the state are considered to be Gaussian mixes when using a
1 Gaussian mixture filter.
𝑦(𝑖) = 𝑌𝑚𝑖𝑛 √2𝛼 2 𝑙𝑛 ( ) (3) 𝐿∗
1−𝑃𝑖𝑛𝑝𝑢𝑡 (𝑖) ∗ ∗ ∗
𝑓𝑘∗ (𝑥𝑘 ) = ∑𝑖=1
𝑘
𝜔𝑘,𝑖 . 𝑁(𝑥𝑘 ; 𝑥𝑘,𝑖 , 𝐶𝑘,𝑖 ), 𝑤𝑖𝑡ℎ ∗∈ [𝑒, 𝑝] (8)
Where Y min is the pixel value's lower bound, and α is a
L_k^* is the number of mixture components for every time
Rayleigh distribution scaling parameter that is determined by
step is k.
the input image. The Rayleigh function has an α value of 0.04
b. Segmentation
in this investigation. For each brightness point, the
Medical image segmentation, also known as natural image
outcome probability density is:
segmentation, is the process of manually, semi-automatically,
or fully automating the extraction of the desired object (organ)
𝑦(𝑖)−𝑌𝑚𝑖𝑛 (𝑦(𝑖)−𝑌𝑚𝑖𝑛 )2
𝑃(𝑦(𝑖)) = 𝑒𝑥𝑝 (− ) for𝑦(𝑖) ≥ 𝑌𝑚𝑖𝑛 (4) from a medical image (2D or 3D). In a variety of applications,
𝛼2 2𝛼 2
medical picture classification is crucial. Microscopy,
A greater number will give in more significant picture contrast dermoscopy, X-ray, ultrasound, Computed Tomography (CT),
enhancement while also raising saturation and noise levels. Magnetic Resonance Imaging (MRI), and positron emission
Step 6: Mitigating the effect of a sudden change, Linear tomography are just a few of the imaging techniques that have
intensity stretch is used to rescale the result of the transfer recently gotten a lot of attention and funding. One of the most
function in Eq (9). The linear contrast stretch is calculated as significant medical imaging techniques is picture segmentation,
follows: which retrieves the ROI using a semiautomatic or automatic
process. It used border identification, tumor tracking, and mass
𝑥(𝑖)−𝑥𝑚𝑖𝑛 recognition in clinical uses to segment body organs and
𝑦(𝑖) = (5)
𝑥𝑚𝑎𝑥 −𝑥𝑚𝑖𝑛 tissues.Because classification divides a picture into logical
The transfer function's lowest and maximum variables are xmin sections, clustering algorithms can be used to professionally
and xmaxrespectively, and input data for this function is x(i). separate the ROI from the background by extracting the image's
Step 7: The grey scale distribution of pixels within a sub-matrix global properties [18].
system can be calculated using a bi-linear interpolation of four
separate mappings to remove border artefacts. • K-means Clustering
K-means is a clustering method or strategy that quickly and
Boosted Anisotropic diffusion filter easily organizes a massive quantity of information [19].In
contrast, the k-means algorithm relies on the initial cluster of
By maintaining fine details in the image, it primarily focuses on values to locate the centre and thus has a flaw.K- means
removing noise. After obtaining the scattered picture, this clustering produces optimal topical treatments as an outcome of
BADF includes the Partial Differential Equation (PDE), which the trials. The testing process, on the other hand, should result
gives the anisotropic scattering filter [17] an extra advantage. A in data similarities or proximity. As a result, it can be divided
type of diffusion that is lacking at the edges and borders can into multiple clusters, each with a high degree of similarity
also be used to smooth the surface. among cluster locations. As cited by (celebi et al. 2013), K-
Means is a multifaceted approach due to the ease with which it
The smoothing scheme in anisotropic filtered is based on the can be modulated at each stage, the ease with which it can
PDE regulating system. quantify distance, and its reliance on iteration termination
criteria.The beginning parameters obtained from the cluster's
𝜕𝐼𝑚
= 𝑑𝑖𝑣(𝑇∇𝐼𝑚 ) (6) midpoint are significant because the K-mean cluster is a local
𝜕𝑡
Im is for image intensity, ∇ stands for gradient, div stands for optimization.These changes are made to attain the highest
dispersion, and T stands for time in the weighting direction. accuracy and quickest convergence. Furthermore, starting from
Smoothing orientation is determined by the structure tensor T. a cluster's midway limits the K-mean cluster method to the best
It is constructed using a common gradient tensor G, which is possible site. The k-mean cluster method chooses a beginning
obtained by correlating the total of outer ∇Im goods with a point from the centre at randomized from a range of styles up
Gaussian kernel Kρ across all weighted instructions: to k. Based on the original centroid cluster, which is selected at
𝐺 = 𝑘𝑝 ∗ ∑𝑚(∇𝐼𝑚 ⊕ ∇𝐼𝑚 ) (7) random, the number of iterations with the centroid cluster is
determined. As an outcome, the centroid cluster in the large
The outer product operator is represented by ⊗. The gradient starting data points can be improved. Data points [1,2,3,...,x] n
tensor's spatial scale is ρ, determined by the standard deviation are typically clustered into k-means using k-means clustering.It
of the Gaussian kernel. An anisotropic diffusion filter's main supports multi-dimensional vectors and features high-
benefit is that it smooths out the image's homogeneous sections performance computing. Figure 4 depicts the flow chart of K-
while also enhancing the edges. means clustering. As a result, the distortion measure is reduced
While Gaussian filters function well in many practical uses, by lowering the cost function, which includes:
𝑗 2
they are significantly constrained when the true state |𝑥𝑖 − 𝑐𝑗 | (9)
distribution is complicated, such as when there are several 𝑗
The 𝑥𝑖 metric is used to calculate the distance between both the
modes, high skewness, or large tails. In multi-target tracking or cluster centre and a piece of data.(n) data points are measured
financial forecasting, for example, such circumstances may by (cj) from their cluster centres.The following parts make up
develop. The obvious extension of Gaussian filtering is the algorithm:
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2022, 10(2)s, 96–104 | 99
1. Set the k points in the clustered object's space. These are different forms, each with a varied number of nodes but the
the initial C_ k centroids. same underlying concept.Resnet50 [21] is a variation that
𝒅 = ‖𝑷(𝒙, 𝒚) − 𝑪𝒌 ‖ (10) supports up to 50 neural network layers.
2. Allocate each object's group to the subcategory with the Deeper convolutional neural networks are used by machine
most nearby centres. learning engineers to solve an issue in computer vision since of
3. While all objects are allocated, compute the locations of Resnet's significance. Since the individual levels can be taught
the k centroids. for various jobs to generate accurate results, these extra stages
aid in the speedier resolution of complex problems. While
𝟏 increasing the number of stacks can improve the model's
𝑪𝒌 = ∑𝒚∈𝑪𝒌 ∑𝒙∈𝑪𝒌 𝑷(𝒙, 𝒚) (11)
𝒌
characteristics, a deeper network can cause performance
4. Keep repeating steps 2 and 3 till the centroids stop
reduction. To put it another way, as the number of channels
moving.
increases, a neural network's accuracy rates may get saturated
The K-means clustering algorithm has two well-known
and begin to decline. As a result, both the training and testing
attribututes in which minimal attributes are divided into groups
data accuracy of the model degrades. Overfitting isn't to blame
for such deterioration. It could otherwise be owing to the
network's initialization, optimization algorithm, or, more
crucially, the issue of vanishing or ballooning biases [22].This
is why ResNet was created. Deep residual networks add
residual blocks into the model to increase their accuracy.
Skipping interconnections can be done in two ways. They start
by generating a new path for the gradient to follow, which
solves the problem of vanishing gradients. They also give the
algorithm the ability to learn a function that defines its identity.
As a result, the model's upper layers operate just as well as its
bottom layers.So, to conclude, residual blocks simplify the
layers' learning of identity values. ResNet improves deep neural
network performance in addition and more neural levels with
minimizing failure rates. To put it differently, skip connections
mix the results of the preceding layer with the outcomes of
stacked layers, enabling the training of far deeper networks
than was previously available.

Figure 4: steps of k-means clustering

c. Deep learning technique using ResNet
Due to the introduction of new technology, the field of
computer vision has seen significant changes in recent years.
As a result of these improvements, computer vision models can
now outperform humans in tackling many tasks such as picture
identification, object detection, face recognition, image
categorization, and so on. Deep convolutional neural networks,
sometimes known as CNNs, need special consideration. For
accurate analysis of visual imagery, artificial networks are
widely used [20].However, while it allows us to add more
layers to the CNNs to perform increasingly difficult computer
vision tasks, it also introduces its own set of problems. It has
Figure 5: An illustration of a Residual block presented by (He
been found that as the number of added layers increases,
et al. 2016), with x as the input and F(x) + x as the output, is
training neural networks gets more challenging, and the
shown before the rectified linear unit (ReLU).
accuracy of the neural networks decreases. The use of ResNet
As the number of layers stacked increases, the number of
becomes critical in this situation. Deep neural networks are
"levels" of attributes increases Deep networks extract small,
more complex to learn. Resnet has made it possible to train
middle, and high-level attributes.The network is built using the
incredibly deep neural networks without trouble.
leftover block. A macro-architecture is formed by a series of
"Residual Network" refers to the system that isn't connected to
convolutions and pooling layers. The entire architecture, as
everything else. Kaiming He, Xiangyu Zhang, Shaoqing Ren,
well as the description of this block, was published by (He et
and Jian Sun initially illustrated it in their 2015 computer
al. 2016). They showed that using the typical Stochastic
vision study entitled "Deep Residual Learning for Image
Gradient Descent (SGD) optimizer and residual blocks,
Recognition." The groups that took first place in the ILSVRC
extraordinarily deep networks can be trained. As indicated in
2015 classification competition with a 3.57 percent error rate
figure5, ResNet50's architecture is divided into four parts.
demonstrated that this approach was a huge success. ImageNet
Pictures with heights and widths multiples of 32, as well as a
identification, ImageNet localization, COCO detection, and
channel width of three, can be handled by the network. For
COCO segmentation were all first place winners in the 2015
simplicity, we'll suppose that the input size is 224 x 3. Each
ILSVRC and COCO contests. ResNet is available in many
ResNet design uses kernel sizes of 7x7 and 3x3, respectively,
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2022, 10(2)s, 96–104 | 100
for initial convolution and pooling.In Phase 1, there were three In table 1, we see that Resnet 50 is composed of the following
Residual blocks, each with three levels. components:
Table 1.ResNet-50 architecture elements • A layer from convolution with 64 unique kernels and
Layer Names Output size Layers a stride of 2. The kernels have a size of 7 × 7.
Conv1 112×112 7×7, 64, stride 2 • The next step is max pooling with a stride size of 2.
Conv2_x 56×56 3×3 max_pool, stride 2 • That this next convolution has a 1 × 1,64 kernel, a 3
1 × 1, 64 × 3,64 kernel, and a 1 ×1,256 kernel. This phase has
[ 3 × 3, 64 ] × 3 nine layers because these three stages are performed
1 × 1, 256 three times in total.
Conv3_x 28×28 1 × 1, 128
• After that, there's a 1 × 1,128 kernels, 3 × 3,128
[3 × 3, 128] × 4
1 × 1, 512 kernels, and 1 × 1,512 kernels. Four stages every
Conv4_x 14×14 1 × 1, 256 time were repeated, totalling 12 of them.
[ 3 × 3, 256 ] × 6 • Following that, there are 1 × 1,256 kernels, followed
1 × 1, 1024 by 3 × 3,256 and 1 × 1, 1024 kernels, all of which
Conv5_x 7×7 1 × 1, 512 are duplicated six times for a total of 18 levels.
[ 3 × 3, 512 ] × 3
• There were 9 levels in total: a 1 × 1,512 kernel, a 3 ×
1 × 1, 2048
3,512 kernel, and a 1 × 1,2048 kernel.
1×1 Avg_pooling, 1000-d
fully-connected, softmax • After that, we conduct an averaged pool and conclude
with a fully connected layer, accompanied by a
softmax function, which results in a single line.
Each of the three levels of stage 1's block uses 64, 64, and 128-
bit kernels to accomplish the convolution operation. Curved
arrows show the identity relationship. Dashed linking arrows 4. Experimental Results
indicate that the Residual Block performs the convolution • Experimental Setup
operation with stride 2 (half its height and width while The suggested method has been implemented in the Python
doubling its width) during the convolution phase.By the end of environment on an Intel Core i5 CPU running at 1.6 GHz. The
the process, the input size has been halved and the channel recommended system was tested using the ADNI data set [17].
width has been doubled.Bottleneck design is used for deeper The Alzheimer's Disease Neuroimaging Initiative (ADNI)
networks, such as ResNet50 and ResNet152.The residual brings together specialists and research to better understand
function F is layered three times.The three layers are made up how Alzheimer's illness progresses (AD). Through MRI and
of 11%, 33%, and 11% convolutions. The downsizing and PET imaging, genetics, cognitive tests, CSF, and blood
restoration of proportions are handled by the 11 convolution biomarkers, ADNI researchers collect, validate, and use
layers. The 33rd stage becomes a bottleneck as input/output information as disease indicators. Along with Research
dimensions decrease. Lastly, the network has a 1000-neuron materials and data from the North American ADNI project, this
fully connected layer and an Average Pooling layer (ImageNet paper offers information about Alzheimer's disease, mild
class output). The architecture of ResNet 50 is shown in figure cognitive impairment, and ageing.
6.
Simulation output:
As part of this study, 150 brain MRI images have
been used to classify Alzheimer's disease patients into two
categories: affected and unaffected.In the training phase, 75
images were used, and in the testing phase, 75 images.The
results obtained by using the previously mentioned strategies
are significantly more precise than those obtained by
alternative techniques. Figure7 shows the input image.

Figure 6: Architecture of ResNet-50

The ResNet-50 model, in particular, is made up of five phases,
each having a residual block. Each residual block is made up of
three layers with 1×1 and 3×3 convolutions. Residual blocks
are a basic concept. Each layer of classic neural networks feeds
onto the next. An identification network with residual blocks
feeds directly into the next stage and on to the next identification
connection level within 2–3 hops.

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2022, 10(2)s, 96–104 | 101
 Sam 69. 34.5 24. 0.97 59. 35.4 29. 0.97
ple 3 6 5 69 35 76 5 69 39
Sam 25. 31. 0.97 20. 36.6 34. 0.97
33.6
ple 4 0 13 60 07 6 1 66
Sam 43. 23.0 32. 0.97 41. 36.0 37. 0.97
ple 5 1 4 77 42 17 4 7 45
Sam 40. 33.1 33. 0.97 30. 36.1 39. 0.97
ple 6 9 5 01 43 91 0 71 47

Segmentation is performed using the K-means algorithm after

preprocessing.The various performance metrics of the
segmentation process are detailed in Table 3. Dice Coefficient,
Jaccard Coefficient, accuracy, and MCC (Matthews' correlation
coefficient) are all assessed.

Table 3: Performance analysis of segmentation results.

METRICS K-means
clustering
Image 1 Dice Coefficient 0.7964
Jaccard Coefficient 0.6617
MCC 0.8117
Accuracy(%) 0.9957
Dice Coefficient 0.9088

Image 2 Jaccard Coefficient 0.8328

MCC 0.9114
Accuracy 0.9974
Figure 7: Samples from the ADNI dataset
Dice Coefficient 0.9818
Figure 8 illustrates how the input MRI Alzheimers images are
Image 3 Jaccard Coefficient 0.9642
preprocessed using the CLAHE and BADF algorithms, which
provide greater image clarity after post-filtering, and less detail MCC 0.9814
loss.The mean squared error (MSE), Peak Signal-to-Noise Accuracy 0.9991
Ratio (PSNR) and structural similarity index (SSIM)are only a Dice Coefficient 0.7846
few of the performance measures listed in Table 2. If the MSE
Image 4 Jaccard Coefficient 0.6456
and PSNR are lower, the signal-to-noise ratio in the extracted
image will be higher. The dice factor calculates how much MCC 0.7647
overlap is there between the automatic and human Accuracy 0.9989
segmentation process for a particular dataset. Images with unclear boundaries are separated by k-Means
clustering.In most cases, these are used to define Alzheimer's
region shapes.Marking severity in this way will be efficient. As
can be seen in Table 4, this algorithm is used to calculate the
area parameter centroid diameter.

Table 4: Calculation of The area parameter centroid diameter

K-mean clustering
Samples
Standard
Mean Variance
error

Figure 8: Preprocessed output 1 0.582031 0.460936 0.020371

Table 2: Performance analysis of pre-processing results. 2 0.694118 0.49991 0.025185

Filte BADF 3 0.740741 0.496041 0.023866

CLAHE
rs
Ima MS PS SN SSI MS PS SN SSI 4 0.705078 0.455954 0.02015
ges E NR R M E NR R M
Sam 59. 32. 0.97 55. 35.6 37. 0.97
33.2 5 0.582031 0.460936 0.020371
ple 1 2 41 51 1 6 41 56
Sam 38. 39. 0.97 33. 36.1 40. 0.97
31.7
ple 2 7 26 33 57 7 26 35

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2022, 10(2)s, 96–104 | 102
 obtains the minimum training and validation losses as around
0.12 and 0.20 respectively.In this figure, the blue colour
indicates the validation and the red colour indicates the loss.
Table5 shows the accuracy, sensitivity, and specificity
performance measures of the proposed model compared to the
existing model.
Table 5: Comparison of performance metrics
References Accuracy Sensitivity Specificity
[23] 93.01 89.13 96.8
[24] 91.02 92.7 89.05
[25] 91.95 89.5 93.8
[26] 92 90.9 93
Figure 9: Classification output [27] 91.41 89.06 93.75
[28] 94.7 94.3 95.8
The updated ResNet-50 architecture has been used in the
Proposed 95.9 96.2 97.4
suggested framework for extracting features, identification, and
categorization of MRI images on Alzheimer's disease. Figure 9 100
depicts the detection and classification output, the statistical
98
value under the output represents the probability of positive as
96 Beheshti, et al [45]
well as the probability of negative outputs. Thus the proposed
framework is accurately identifying the affected and non- 94
Suk, et al [46]
affected MRI images. Shi, et al [47]
92
Liu, et al [48]
90
Aderghal, et al [49]
88
Islam, et al [50]
86
Proposed
84
Accuracy Sensitivity Specificity
Figure 12: Comparison of performance metrics

Based on the performance parameters such as accuracy,

sensitivity, and specificity our proposed framework[45-50], is
compared with other existing techniques. Figure12 depicts the
graphical depiction, and table 5 shows the statistical
representation. We can see from the table and graph that the
proposed framework outperforms strategic approaches for
Figure 10:Training and validation accuracy
detecting and classifying Alzheimer's disease.
5. Conclusion
MRI has evolved in neurology and oncology over the last
several years.MRI has improved soft tissue delineation, which
makes it the most commonly used imaging technique in the
early stages of primary Alzheimer's illness. Researchers used
Resnet 50, a deep-learning technique previously used to
diagnose Alzheimer's illness.The results suggested that the
deep learning based ResNet 50 is promising for classification
and diagnosing diseases with neuroimaging data. The proposed
method uses K-means clustering to extract regions of interest
from MR Images.Utilizing standard datasets, we test the
proposed model's performance and categorize MRI as normal
Figure 11:Training and validation loss AD.As a result, it produces more precise results with greater
efficiency.
The proposed system of Alzheimer's disease identification has
training and validation accuracy, as shown in Figure 10. It's
References
worth noting that testing precision exceeds 95%, while
[1] Zetterberg, Henrik, and Barbara B. Bendlin. "Biomarkers for
validation precision hovers around 93%. In this figure, the blue Alzheimer’s disease—preparing for a new era of disease-
colour indicates the validation and the red colour indicates the modifying therapies." Molecular psychiatry 26, no. 1 (2021):
training. The suggested architecture for Alzheimer's disease 296-308.
identification also suffers from training and validation loss, as [2] Sanchez, Justin S., J. Alex Becker, Heidi IL Jacobs, Bernard J.
seen in Figure 11.It is worth noticing that the proposed system Hanseeuw, Shu Jiang, Aaron P. Schultz, Michael J. Properzi et al.

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2022, 10(2)s, 96–104 | 103
 "The cortical origin and initial spread of medial temporal Unnatural light,” International Journal of Engineering Research
tauopathy in Alzheimer’s disease assessed with positron emission and Development, vol. 8, no. 9, pp. 57–61, 2013.
tomography." Science translational medicine 13, no. 577 (2021): [18] Pizer SM, et.al (1990) “Contrast-limited adaptive histogram
eabc0655. equalization: speed and effectiveness”, In Proceedings of the first
[3] Johansson, Maurits, Erik Stomrud, Philip S. Insel, Antoine Leuzy, conference on visualization in biomedical computing. IEEE, pp
Per Mårten Johansson, Ruben Smith, Zahinoor Ismail et al. "Mild 337–345.
behavioural impairment and its relation to tau pathology in [19] Aimi Salihai Abdul, MohdYusuffMasor and Zeehaida Mohamed,
preclinical Alzheimer’s disease." Translational psychiatry 11, no. Colour Image Segmentation Approach for Detection of Malaria
1 (2021): 1-8. Parasite using Various Colour Models and k-Means Clustering, In
[4] Leng, Kun, Emmy Li, Rana Eser, Antonia Pierogies, Rene Sit, WSEAS Transaction on Biology and Biomedicine., vol. 10,
Michelle Tan, Norma Neff et al. "Molecular characterization of January (2013).
selectively vulnerable neurons in Alzheimer’s disease." Nature [20] Hossain, Md Belal, SM Hasan Sazzad Iqbal, Md Monirul Islam,
neuroscience 24, no. 2 (2021): 276-287. Md Nasim Akhtar, and Iqbal H. Sarker. "Transfer learning with
[5] Wightman, Douglas P., Iris E. Jansen, Jeanne E. Savage, Alexey fine-tuned deep CNN ResNet50 model for classifying COVID-19
A. Shadrin, Shahram Bahrami, Dominic Holland, ArvidRongve et from chest X-ray images." Informatics in Medicine
al. "A genome-wide association study with 1,126,563 individuals Unlocked (2022): 100916.
identifies new risk loci for Alzheimer’s disease." Nature [21] .Alsabhan, Waleed, and TurkyAlotaiby. "Automatic Building
genetics 53, no. 9 (2021): 1276-1282. Extraction on Satellite Images Using Unet and
[6] Nguyen, Phuong H., AyyalusamyRamamoorthy, Bikash R. ResNet50." Computational Intelligence and Neuroscience 2022
Sahoo, Jie Zheng, Peter Faller, John E. Straub, Laura Dominguez (2022).
et al. "Amyloid oligomers: A joint experimental/computational [22] Saadna, Youness, AnouarAbdelhakimBoudhir, and Mohamed
perspective on Alzheimer’s disease, Parkinson’s disease, type II Ben Ahmed. "An Analysis of ResNet50 Model and RMSprop
diabetes, and amyotrophic lateral sclerosis." Chemical Optimizer for Education Platform Using an Intelligent Chatbot
reviews 121, no. 4 (2021): 2545-2647. System." In Networking, Intelligent Systems and Security, pp.
[7] Karikari, Thomas K., Andrea L. Benedet, Nicholas J. Ashton, 577-590. Springer, Singapore, 2022.
Juan Lantero Rodriguez, AnniinaSnellman, Marc Suarez-Calvet, [23] Beheshti, I., Demirel, H., Matsuda, H., Initiative, A.D.N., et al.:
Paramita Saha-Chaudhuri et al. "Diagnostic performance and Classification of Alzheimer’s disease and prediction of mild
prediction of clinical progression of plasma phosphor-tau181 in cognitive impairment-to-Alzheimer’s conversion from structural
the Alzheimer’s Disease Neuroimaging Initiative." Molecular magnetic resource imaging using feature ranking and a genetic
psychiatry 26, no. 2 (2021): 429-442. algorithm. Comput. Biol. Med. 83, 109–119 (2017).
[8] Neff, Ryan A., Minghui Wang, SezenVatansever, Lei Guo, Chen [24] Suk, H.I., Lee, S.W., Shen, D.: Deep ensemble learning of sparse
Ming, Qian Wang, Erming Wang et al. "Molecular subtyping of regression models for brain disease diagnosis. Med. Image Anal.
Alzheimer’s disease using RNA sequencing data reveals novel 37, 101–113 (2017).
mechanisms and targets." Science advances 7, no. 2 (2021): [25] Shi, B., et al.: Nonlinear feature transformation and deep fusion
eabb5398. for Alzheimer’s disease staging analysis. Pattern Recognit. 63,
[9] Qiu, Chengxuan, MiiaKivipelto, and Eva Von Strauss. 487–498 (2017).
"Epidemiology of Alzheimer's disease: occurrence, determinants, [26] Liu, M., Zhang, D., Shen, D., Alzheimer’s Disease Neuroimaging
and strategies toward intervention." Dialogues in clinical Initiative: Hierarchical fusion of features and classifier decisions
neuroscience (2022). for Alzheimer’s disease diagnosis. Hum. Brain Mapp. 35(4),
[10] Dubois, Bruno, Gaetane Picard, and Marie Sarazin. "Early 1305–1319 (2014).
detection of Alzheimer's disease: new diagnostic [27] Aderghal, K., Benois-Pineau, J., Afdel, K.: Classification of
criteria." Dialogues in clinical neuroscience (2022). SMRI for Alzheimer’s disease diagnosis with CNN: single
[11] Gao, Shuangshuang, and Dimas Lima. "< PE-AT> A review of Siamese networks with 2d+? approach and fusion on ADNI. In:
the application of deep learning in the detection of Alzheimer's Proceedings of the 2017 ACM on International Conference on
disease." International Journal of Cognitive Computing in Multimedia Retrieval, pp. 494–498. ACM (2017).
Engineering (2021). [28] Islam, Jyoti, Yanqing Zhang, and Alzheimer’s Disease
[12] Attallah, Omneya, Maha A. Sharkas, and HebaGadelkarim. Neuroimaging Initiative. "Deep convolutional neural networks for
"Deep learning techniques for automatic detection of embryonic automated diagnosis of Alzheimer’s disease and mild cognitive
neurodevelopmental disorders." Diagnostics 10, no. 1 (2020): 27. impairment using 3D brain MRI." In International Conference on
[13] Liu, Fei, Huabin Wang, Yonglin Chen, Yu Quan, and Liang Tao. Brain Informatics, pp. 359-369. Springer, Cham, 2018.
"Convolutional neural network based on feature enhancement and
attention mechanism for Alzheimer's disease prediction using
MRI images." In Proc. of SPIE Vol, vol. 12083, pp. 120830X-1.
2022.
[14] Loddo, Andrea, Sara Buttau, and Cecilia Di Ruberto. "Deep
learning based pipelines for Alzheimer's disease diagnosis: a
comparative study and a novel deep-ensemble
method." Computers in biology and medicine 141 (2022):
105032.
[15] A. Kaur and C. Singh, “Contrast enhancement for cephalometric
images using wavelet-based modified adaptive histogram
equalization,” Applied Soft Computing, vol. 51, pp. 180–191,
2017.
[16] Shehroz S. Khan and Amir Ahmad, Cluster Centre Initialization
Algorithm for K-means Cluster, In Pattern Recognition Letters,
pp. 1293–1302, (2004).
[17] D. Sonker, “Comparison of Histogram Equalization Techniques
for Image Enhancement of Grayscale images in Natural and

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2022, 10(2)s, 96–104 | 104