Chapter 1 Introduction

1.1 Background
Based on the advanced digital technology and communication networks,
humans are more conscious in obtaining enhanced health information along with EHR
[1]. The highly advanced hospitals with specialized equipment are rapidly increasing
worldwide. Hence, patients pay more attention in visiting these medical care service
providers and, they are interested in consulting popular doctors for better disease
cal, pharmacy, personal and account
ting data are needed to be stored properly in segmented way. As days goes on, the
amount of data stored has grown rapidly. Hence, there is a need for upgrading the old
storage system, which is considered as a complicate process, especially in security
issues [2]. Also, there is a need for interchanging of information regarding health issues
between the providers in the healthcare environment when the patients prefer various
hospitals for treatments. This is considered as a complicated process. Hence, there
exists an urge to covert the health in formations into digital form along with proper
network for communication for sharing the generations regarding to the health issues
of patients. A consolidated healthcare data management is required among the doctors,
patients, hospitals, pharmaceutical manufactures, insurance firms and medical labs,
which lie at several geographical locations [3]. For obtaining a solution for the above-
mentioned complications, a new system is required to be implemented with efficient
storage space, safely stored EHR and a properly validated retrieval mechanism. Such
-
sharing and patient-provider relationship with more relevant to patient-centric data
management in health care environment.
In day-to-day medical platforms, there exists a rapid development in medical
imaging and detection techniques which have become most important element in
disease detection. Medical images embrace a significant role in storing information of
nerves, brain, heart etc. for obtaining the interior vision of human body [4]. The medical
images consist of enormous mathematical applications to analyze, whether the health
tissue is affected or not. Although, missing of any specific data or area during the
process of medical imaging may lead to death. Hence, this is considered as a main
complication in medical imaging that is to utilize the images without any kind of

1
information or data loss. It is highly feasible and during the process of acquiring of
medical images, the corruption of network by artifacts and noise may takes place. The
term noise is defined as changing of original pixel values of images, randomly. Due to
this interruption of noises, the quality of the images is reduced and is highly remarkable
when the images comprise of small size and lower contrast in comparison with the
images obtained from the conventional image processing [5]. Damaging the medical
image noises is severely critical because these noises will minimize the image quality
and also confuse the disease detection processing. Hence, the process of de-noising in
medical images is unavoidable and is considered as essential pre-processing stage in
the medical imaging systems.
The examining of the medical image model is generally carried out for
identification of single disease. But, the demand for the detection of multi-level diseases
in medical applications and the multi-function solutions are needed to be explored.
Many health care analysis systems consider only one disease at a time [8]. Several
articles are also focused on one specific disease. When, the organizations need to look
after the reports of their patients, they must establish many models, but the existing
process is utilized for analyzing only a specific disease [7]. The death rate is increasing
now-a-days due to the improper identification of the specific disease exactly. The cured
patient from one disease may tend to affect from another disease due to this improper
detection [9]. Some of the available systems utilize several parameters during the
process of analysis of disease. Due to this, it is not possible to identify one disease,
which is caused by the side effect of another disease [6]. For example, due to diabetes,
there exists the chance of neuropath, heart disease, hearing loss, dementia and
retinopathy. Hence, there is a need for multi-disease model.

1.2 Overview of Medical Data Processing

The medical data recording is considered as one of the best clinical tools. The
data collection and its usages are midway to several kinds of activities in health care
environments. These data are provided during the routine clinical regimen or utilized
directly from the patient by using the PRO measures. Even though, the accuracy and
other reasons justify the availability of data for a wide range of potential uses, these
data are mostly collected for the single particular purpose. There are only a smaller
number of literatures available for the combined use of clinical research, health data for

2
clinical use, administrative purposes, and surveillance. EHR is important for

confidentiality, and data processing. The communication of the service providers is

promoted by the clinical records available. The quality of the records is essential in
improving the quality of patient care. The analysis of the outcomes obtained from the
treatments is typically entrapped by physical examination, communication, and
combination of biological data with the patient. Although, the analysis of treatment
outcomes at a group level namely administratively, epidemiologically, geographically,
or at the facility level will mainly focus on the mortality. But these outcomes may fail
to gather the treatment outcomes entirely. The process of medical images data
processing by using HIS and EHR is shown in Figure 1.1.

Data Data
collection preprocessing

EHR

Data
HIS Feedback iteration mining

Application Evaluation

Figure 1. 1 Process of medical image data-processing

1.3 Need of removal of noises in Medical Images

The noise is the unnecessary signal present in the images and it was caused by
various resources such as the slow shutter, low light, sensor heat and sensor fill factor
etc. The generation of white or dark spots in the images takes place due to the sudden
disturbance in the signal of the images, and it is termed as Impulse noise. It is typically
in terms of dispersed black and white pixels. To reduce these effects, many filters have
been designed and implemented to provide an image approximation like the original
image. The low clarity of noise reduction or improper de-noising algorithm makes the
process of obtaining original image to be complicated. For example, a tiny spot
indicated in black color consists of one or more pixels in an image at uniform green

3
area will not definitely be noise. Furthermore, these pixels are available in the shape of
regular or irregular form. Regular shapes are created because of the sensors defects and
at the same time irregular shape is hopefully the real features or elements of the image.
Apart from that, there are few cases, where the reduction of the impulse is not
considered as a huge problem to collect small detailed information, which are hidden
in a bunch of noise such as astronomical images. Moreover, the imperfect algorithms
induce issues such as leaving some of the noise or removing important features during
processing [10]. The median filtering technique have also comprised of some defects as
well. It is not capable in preserving the edges of the image processed and is relatively
complex and expensive with high means of time. Either the pixels are corrupted or
uncorrupted, it computes each pixel in the same way only. Wide variety of efforts has
been taken by various investigators for reduction and removal of impulse noise in the
medical imaging systems. The algorithms based on the median filter are one of the
benefactions based on the reduction of noise or its removal. These entire algorithms
based on median comprises of their available advantages and complications in pursuing
of images de-noising through requesting later investigation to provide huge
developments in the study area. Median filters are typically applicable even in the day-
to-day industries, and these filters can be utilized in locations, where a small delay can
be authorized in medical image processing. Hence, the complete ignorance of them was
not possible as they are objective and still used. A good denoising algorithm is needed
based on the following factors listed below.
1. Protection of edges: The abnormalities in medical images were effectively
detected by using the medical imaging technique. These complications are mostly
occurred by the mass replication of the cells present inside the human body. To analyze
at what extent the disease could spread, a clear demarcation of cells is needed.
Additionally, the surgeries done by using guidance images, it is essential that the edges
of the region of disease should be recognized accurately. Several post-processing
methodologies will tend to smooth the edges of an image that is not applicable in
medical images. Hence, there is a huge need for the best medical image de-noising
algorithm for securing those edges of the medical images.
2 Preserving the structural similarity: The algorithm for de-noising will not
establish any kind of artifacts, while processing de-noising. It is essential that during
entire minute information are
equivalence with the real image.

4
 3. Difficulty: Several applications of medical imaging are considered as the
instantaneous process such as surgeries done under the guidance image information by
utilizing the CT scan images or which needs the instant results present likely in US
imaging. Consequently, it has high significance, where the considered algorithm will
be effective in calculation capacity and comprise of minimum difficulty.
4. Unnecessary of prior databases: The acquiring of medical imaging data is
complex because of its highly privileged nature and related ethical issues. Eventually,
a best de-noising algorithm needs to utilize reduced number of databases to attain a
better de-noising performance. The image comprises of noise, and the image after
removing noise by using de-noising filter is represented in the following Figure 1.2.

(a)Poisson noise (b)Filtered image

Figure 1. 2 Poisson noise removal
The imaging modalities generally include MR images, US images, PET images
and CT images. The previous studies disclose that several imaging systems were
disturbed by various categories of noises comprises of its own distribution functions.
Therefore, a consolidated technique will not establish the problems of de-noising the
entire medical imaging techniques.

1.4 Techniques in Noise Removal of Medical Images

The techniques involved in the process of reduction of the noises in medical
images are broadly classified into two ways such as before and after the image
acquisition [11]. The prior indicates the special operating modules, which are applied
to the image collecting system to squash the noise and also to improve the information
of contents that takes place in created image. The hindmost marked the image
processing techniques in digital form to minimize the noise and to enhance the image
quality. The following block diagram shows the basic noise removal from data images
in Figure 1.3.

5
 Input image Noise addition

Restored Filtering
image algorithm

Figure 1. 3: Steps involved in basic noise removal of image processing system

1.4.1 Spatial domain filters
Spatial domain filtering technique is a deep-rooted denoising methodology. In
this technique, the spatial filters directly put on to the digital images to the process of
removal of noise from the image.
Median filter: Median filter is considered as a very famous methodology for
the process of noise removal based on its best de-noising power and computational
efficiency. The pixel values are removed by using an intensity median in the
surrounding of that pixel by using the median filter. For the outcome of surrounding
pixels, an irreplaceable filtering size window is utilized by the median filter. The
consistent implementation of the median filters is carried out along the images and it
alters both of the noise free and noisy pixels present in the image.
Bilateral filter: A bilateral filter represents an edge-preserving, noise-reducing, and
non-linear filter for smoothing of images. It changes the number of pixels intensity by
using weighted average intensity values obtained from the closely placed pixels based
on the Gaussian distribution.
LMMSE: It is familiar that the wavelet-indicated images are scales. In the
wavelet domain, there is a rapid decrease in noise level along scales during the signal
structures are hardened with increasing scale. Hence, they use information of the
coarser scale for improving the finer estimation of scale.
Non local means filtering: The classical NLM filter averages the pixels
weighted by the likeliness of their surroundings that is utilized and illustrated to
minimize the Rician noise effectively without influencing the information of edges in
the MR images.
1.4.2 Transform Domain Filters
The transform domain methodologies work based on the significance that, at
first the image is transferred into a transform domain. After that, the process of

6
despeckling was carried out by using various data elements of image present in
transform domain. The techniques based on multi-resolution or the transform domain
are established for minimizing the points in US image processing.
PCA: The PCA denoising of the MRI relaxometry data is considered as an
enhanced technology for developing the validity of parameters without surrendering
the image resolution. This simple and important step for processing makes the way for
wider applicability of the multi-exponential MRI relaxometry.
WSM: This method was considered as an elaboration of the above-mentioned
analysis of MRI relaxometry. This model proposes a inter-mixing of a 3D sub-bands
wavelets among this filters to enhance the performance of noise removal later, while
maintaining the time for calculation comparable.
Contourlet Transform: The Contourlet transform has been implemented to
conquer the controls of the wavelets transform [12]. It allows several numbers of elastic
directions at every scale, while mostly succeeding the critical sampling.
Wavelet based thresholding techniques: The wavelet thresholding is
considered as simplest non-linear methodology, which processes on single wavelet
coefficient at allotted time. Wavelet thresholding is utilized in Bayesian framework to
process on the statistics of the network signal in non-Gaussian noises. The main
disadvantage in many of the thresholding approaches is the previous understanding of
the number of noise available in the image to calculate an optimum threshold that was
in high demand by them.
1.4.3 Hybrid filters
For connecting the above explained techniques, the latest investigators
significantly aims on the hybrid filters [13]. The flexible filter degrades the noise
present in the images into maximum and minimum frequency elements and it was

minimum amount of speckle content. Contrary method employed in this filtering is the
OBNLM along with equaling of blocks, which is considered as a patch-based recovery
methodology. The advantages of spatial domain filtering and multi-resolution schemes
methodology were utilized by the fast bilateral filter for connecting the wavelet
thresholding. By using the wavelet-based and guided bilateral filter denoising
methodologies, enhanced performance accuracy is achieved.
1.4.4 Low Rank based techniques
The Low-rank techniques are also used as a methodology for denoising. The

7
exploration of the issues related to recovering the despeckled US image using NNM
method was carried out based on inspiration of low rank recovery techniques in
denoising [14]. The difficulties are initially noted by using the WNNM methodology,
which is then altered by including a stage of pre-processing to décor-relate edges of the
ultrasound data in video form by comprising its statistical properties.
1.4.5 Sinogram based methodologies
Sinogram filtering methodologies is utilized to minimize the mixed Poisson
Gaussian noise available in the projected medical data [15]. In this method the process
of filtering is carried out based on projections by utilizing the sonogram or the image
space.
ResNet: This ResNet focuses on the elimination of both artifact and noise by
utilizing the left-over component in-between the LDCT and HDCT images
correspondingly in the training data.
1.5 Medical Storage Systems
The medical information is contained in the form of records accurately and in
detailed manner along with safety assurance, which is considered as an essential part in
medical data management. With the incoming of renewed technologies such medical
records, electronic computing, similar to various application, the domains are highly
depended on computerized storage systems for the process of secured storage and
preservation of health information. Although, in the digital domain, the attackers and
adversaries are normally different than the physical world digitally stored
computerized data can be copied word for word, and records may be revealed to several
types of competitors [16]. To secure the privacy and protection of such computerized
electronic medical records, the privacy and protection laws for consumers are
composed in many countries. This law comprises of rigorous guidelines and needs for
regulations of the medical data storage management. Unluckily, the available storage
features are unable to provide strong safety and privacy assurance mandated by the laws
related to the new digital information domain [17]. For example, various regulations
are needed along with data integrity for the period level of 30 years. But storage of such
records for huge time will need unavoidable change of the hardware and format of the
storage [18]. The resulting transportation to new network servers is loyal and
remarkable. Likewise, if the records require to be replaced after the retention period,
the system of storage must assure safe deletion [19]. Such variety of features is not
available in current storage systems for storing the medical data.

8
 The personal medical data is a kind of personal data, which was possessed by
an individual. For protecting the personal private data, the users are allowed to access
the medical record along with proper authorization [20]. The protection of the private
medical records is not only moral liability but also a need from legal guidance. The
traditional computerized medical data storage system comprises of centralized scheme,
which rely on the dependable firewall, such as increased elements of authentication or
acquiring most powerful scheme of encryption and so on [21]. Although, a hacker can
access entire medical data once he/she have got into the health information system that
implies the availability of one point of failure [22]. Meantime, patients have the
mentality to be faithful towards the medical organizations storing medical records. The
securing of medical data of patients has become challenging task to the medical
organizations.
The modern-generation medical data management systems are based on text and
documents. The future-generation medical data storage systems are now under the
process of improvement and it comprises of computerized medical records known as
CPR. The CPR comprises of all sorts of information of medical data and is accessible
online at any time to the care providers with proper authority [23]. The data elements
of the CPR comprise of all types of information objects such images, documents, text,
audio, video between others. The technical complication of this attempt involves the
transmission, visualization, storage, and acquisition of these objects features in an
effective manner [24]. The need for medical data storage management is considered as
a main problem, which is being mentioned by the medical information platforms. The
CPR ensures to provide confidential, accurate, and accessible information to the
relevant healthcare providers. When the patient medical records are stored and retrieved
once digitally, the available and future decision system based on computerized data
systems are used to assist the patients care and liability.
1.6 Benefits of Secured Medical Storage Systems
The health data management is described as the technique of collecting, storing,
retrieving, transferring, and protecting medical data. It allows the healthcare providers
to access the information and utilize it to the workable perceptions [25]. But that is only
half side of it. Let us take a deep study at the various benefits and chances that effective
data storage management can provide to the staffs and patients. The following figure
shows the processing of medical data management system.

9
 Medical data Collecting

Storing

Retrieving

Transferring

Protecting

Figure 1. 4 Process of medical data management system

Coordinated health care: The specialty of the medical data management
system lies in the capability to allow and provide health data of patients-without
considering the place, where the healthcare related services are given. Proper and valid
data transported safely in-between the healthcare managements assures readily
available services and treatments with high accuracy.
Lower operational costs: Proper data management helps in minimizing the
healthcare expenses. For example, a consolidated health care management system
provided in the cloud network can minimize the cost in certain regions, comprising of
hardware maintenance and healthcare data management. It paves way for the

data or consult
Better data analysis: Clear data management and HIE have considered as the
stepping stone for utilizing creative technologies like AI. For example, machine
learning helps in evalua
in particular areas or population groups. It helps the medical care providers to create a
complete view of heath condition of the patients and to carry on the preventive
measures if required.
Improved patient experience
enhanced heath care management systems. The entire data is readily available hence;
they do not need to repeat things again and again. For instance, filling the similar

10
medical forms repeatedly or having the same medical diagnosis repeatedly.
Consolidated storage provides broad entry to their medical care information. The
improved clarity can motivate the activities of patient and increase the performance
accuracy of records. The clear collection of data, storage of data and management in
medical care utilities aims on these three factors to gain advantage from the entire data
collected.
1.7 Multi-disease Prediction Models
Health is considered as main element for the human to survive peacefully. As
the population increases, the disease also increases. Hence, clinicians encountered
difficulties in detecting many numbers of diseases and to accurately identify if the
patients is suffering from a particular disease or not, against the rapid population
growth. Therefore, the huge requirement for the enhanced healthcare systems has been
raised to fulfill the medical requirement of the growing populations. Such technologies
are created with methodologies, data and healthcare that must be preceded in an
accurate, effective, and smart action. Prediction system is considered as the best
procedure that satisfies the expectation level in the required field [26]. There are various
models have been established for detection of single disease. But there exist only few
technologies for the prediction of multi-level disease. The disease identification
technique has aimed to detect multi-level diseases by utilizing the data from medical
records of patients [27]. Apart from the single disease detection problems that only aim
on disease risk management in a model, this study focuses in technically determining
all the available disease of the individual including the chances of disease, which may
develop in future. The figure 1.5 represents the schematic of multi-disease prediction
model.
In day-to-day world, the detection of disease based on the computerized format
by utilizing the clinical data with DL approaches is considered as a broad area of
investigation. But there exist some difficulties in this specific field, which is considered
more remarkable. The multi-level disease prediction based on RNN achieves high
prediction accuracy and it comprises the capability to resolve any size of data, but the
mathematical calculations are difficult to be carried out because of its recurrent nature,
and the learning of RNN is analyzed as complicated one during the operation of large
medical data environments [28]. The multi-task learning is implemented for
and the accurate information results are obtained through

11
this learning process. It also avoids the over fitting problems. Even though, this
proposed method is default to be implemented in huge data experiments. It will succeed
only on the single data experiments.
ensemble learning and deep polynomial networks are found to be helpful even in data,
which are unstructured and this methodology achieves valid predictions but still the
identification and execution of problems is considered as difficult for huge datasets.
When there is a deeper network available, there are issues in optimization. By utilizing
the network analysis and data mining methodologies such as parameter optimization,
tree classification and predictive training-regression, it gains clear accuracy in
predicting diseases and it enhances the information quality. Still, it undergoes some
complications during the processing of real-time datasets and several numbers of
essential resources for the huge management datasets.

Start

Data cleaning and

processing

Risk factor
identification

Classificatio
Training set
n

Trained Future
Training set prediction
model

Disease No disease

Figure 1. 5: Multi-disease prediction model

The ML is capable in obtaining calculation time and low robust for over fitting
issues. It maintains more accuracy and it determines the data easily. Although, the
correcting of issues is complicated in particle swan optimization and it comprises of

12
reduced rate for convergence. The GA is considered as computationally expensive,
when compared with the decision trees, and it also has large time consumption. The
FNN will not rely on input data and by reading datasets; the networks are capable to
create the whole network. But, the FNN needs long training portions and there exists
some practical problems [29]. For the determination of CKD, the SVM comprises of
better prediction than ANN and both the methods are utilized in storing the data
information on each cloud network. Eventually, the SVM is considered effective but, it
will not perform huge data applications. Their dependent characteristic for hardware
affects the performance validity. These disadvantages encourage future investigators in
identifying more enhanced techniques for the multi-level disease prediction.

1.8 Data Mining for Multi-disease Prediction

The data mining techniques are implemented in wide variety of applications. In
healthcare environments, it plays a remarkable role in disease prediction for the
identification of a particular disease and several test result data are required from the
patients. But the data mining technique effectively reduces the number of tests that
needs to be performed. This reduction in tests plays a remarkable role in time
consumption and validation accuracy.
Naive Bayes: The Naive Bayesian method proceeds based on Baye
theorem with some assumptions, which are not dependent between the predictors. The
generation of Naïve Bayesian model is considered simple with no complicated
parameters of iteration in estimation, which turns the process useful for computing huge
datasets. In spite of its simplicity, it fulfills all the sufficient requirements and is broadly
used because it frequently performs several sophisticated models.
Decision Tree (J48): Decision tree models are widely used in data mining to
determine data and to influence the tree and its rules are used in preparing the
predictions. This prediction is used in the identification of classification values, when
occasions are to be located in divisions or classes. Decision trees are in the shape of the
tree, in which nodes are considered as leaves, that denotes the value of target domain
or example classes or a decision node. Some remarkable tests are carried out on a single
value by comprising one branch, and the sub-tree is considered for each available
outcome of the test. A decision tree is utilized to categorize the case by looking after
the stem of the tree and shifting towards the location of the leaf node, which gives the

13
categorization of the instance.
K-means clustering: It [30] was a quite abandoned learning that will be utilized
for data information during the presence of unlabeled data. The key element utilized for
the understanding of input sample data is clustering process. The important purpose of
this algorithm is to identify for groups present in data and finding the number of
clusters. For every data point, one of K groups is allotted on the basis of iteration. The
clustering carried out on the points of data is considered based on the identical features.
The output obtained from the K-means algorithm is in the form of clusters and labels
for the data, which are trained.
ANN: ANN is enrolling systems motivated towards the biological neural
networks. ANN comprises of many number of nodes. Each node is referred as
biological nodes present in the brain of humans. These neurons may communicate with
each other by utilizing huge number of links. These considered nodes agree the input
data and they execute straight forward processing on the data. The outcome of this
process is moved to several neurons. The output obtained from each node is known as
node values.
Bayesian Algorithm: The modeling of data mining based on prediction are also
done by utilizing the Bayesian algorithm, it includes a smaller number of mathematical
computations, when compared with the other available algorithms. This algorithm
analyses the probability condition of the input and the output monitored and process
faster analysis indicating the connection between the input and output. This algorithm
is remarkable in its starting stage of detection and utilized in treatment of the harmful
infections for early diagnosis. Figure 1.5 shows the techniques involved in data mining.

Naive Decision
Bayes Tree

Data mining techniques

K-means
Bayesian
clustering
Algorithm

Artificial Neural
networks

Figure 1. 6: Data mining techniques.

14
1.9 Deep Learning and Machine Learning in Prediction of Multi-
diseases
The prediction of various diseases is considered as main factor in healthcare
organizations for taking the clear decisions. In health care field, improper affirmations
may leads to delay in the diagnosis or even result in death. Several models for the
purpose of prediction were analyzed and many challenges are affecting the disease
diagnosis. The main challenge considered in healthcare industries is variable
accumulations, where the data resources will be non-simultaneous into remarkable
signs of personal healthcare data. Initially, the healthcare professionals face the
difficulties to collect and estimate huge variety of data for victorious diagnosis and
treatments based on the availability of minimum tools or methodologies. Hence, for the
effective analysis of disease prediction on the basis of the available approaches, only
small number of variables is considered such as gender, height, weight, age and more.
On the contrary, the machine learning utilizes more variables on the basis of calculating
gadgets. Thus, for the disease prediction, the machine learning achieves high
performance in the healthcare environment. The further disease status for future, where
analyzed by utilizing several algorithms. This algorithm paves way to create models for
analyzing data and executing of the obtained results by use of ancient data and the data
from real-time. By implementing machine learning, clinicians choose enhanced
medical record and diagnosis availabilities, which
makes the services of medical industry to improve efficiently. DL is the main
development progress of ML and it is applicable for the collection of main features
from the difficult datasets efficiently from the utilized and ordered and stored learning
approaches. DL provides enhanced performance accuracy in several sectors such as
identification of images, natural language processing, and identification of speech.
LR: LR is an effective and well-known technique for monitored classification
[22]. It is considered as an expansion of normal regression and only the diverging
variable, which generally denotes the appearance or non-appearance of the event. LR
also supports in determining the probability that a renewed occasion included to a
particular class. Since it is a probability, the result takes place in-between 0 and 1.
Hence, for using the LR binary classifier, a threshold is required for allotment to
categorize two classes.

15
 SVM: SVM may differentiate both non-linear and linear data. Initially, the
mapping of each data item was carried out with n-dimensional elements space, in which
n is represented as the number of features. Then, they determine the hyperplane that
splits the data into two different sections during the process of maximization of distance
on edges for both the sections and reducing the classification errors. The edge distance
for the sections is considered as the available distance from the decision hyperplane till
its closest specimen, that itself an element of that section. Moreover, every point of the
data was noted first in terms of points in n-dimension space along with the feature value
of each element. To execute the classification, the hyperplane are determined, which
differentiates the two sections by the maximum edge.
CNN: When a neural network comprises of only a fully joined layer, which
arrives a deeper level, the number of parameters will be increased gradually, which will
then maximize the memory footprint and also computational cost is maximized. To
fight against the issues caused by entirely connected layers and improves the
enhancement of neural networks to a deep level, a remedy has identified and
implemented gradually. By utilizing the concept of weight sharing and local correlation,
a CNN is developed, which not only minimizes the amount of parameters, but also
efficiently enhances the training accuracy. Although, CNN was typically established
for computerized vision and it has typically achieved as very powerful tool in many
other fields and the well-known paths including the health field. CNN precedes several
data, where the leading is the operation of images along with the spatial local correlation
data namely health care elements such as CT, chest X-ray and various image data.
Despite, there exists a single data image for the process of disease diagnosis [16], the
CNN model also recognize the medical data processing by utilizing the fusion
technology and feature extraction of various difficult data. Then, the entirely inter-
connected layers are used to succeed in the categorization of disease and it also has high
performance accuracy when comparison of other traditional models.
RNN: The process of task-based dialogue and time series forecasting etc. and
need models to train the features from the arrangements. The complete inter-connection
of neurons is not possible in the fully-connected layers, which can affect the time
management in sequencing. The CNN can resolve the issues related to spatial
dimensions, but the temporal dimension issues will not be resolved effectively. Based
on the above-mentioned complications, the RNN was considered in the network, which
comprises of feedback loops and neurons. For situations such as the prior and next input

16
has dependent in relationship, the RNN indicates specific advantages. The performance
accuracy of RNN is good when correlated machine learning algorithms for the
processing of big data comprised of abnormal noise and delays. But, during the learning
process of RNN, it faces certain issues such as disappearance and explosion of gradient
and also its short-term memory loss, which paves way to introduced improved GRU
and LSTM models.
AE: The availability of huge numbers of unnamed data is originally presented.
It encourages the generation of the unsupervised training algorithms. In real medical
situations, the labeling of some specific names is slight complicated. Hence, some
improved technology is required to categorize or aggregate the medical data to establish
future potential values of the data. Moreover, there are several ways to gain the health
data of patients and it is connected to the real healthcare environment. Some temporary
data loss will occur, and the information will be deformed and incomplete [27]. The
auto-encoder performs effective feature extraction on the unnamed data through the
supervised learning and also operates well in noise removal, feature dimensionality
reduction and restoration of real data [28].
GAN: GAN is a type of model, which is utilized to learn synthetic data on the
basis of the game theory. The concept of this game theory is specifically incorporated
similar to the network learning. Two types of sub-networks are arranged namely
discriminant and generation network. The process utilized by them is similar to the
conflict, and for the continuous optimization of the generation and discriminant
network, the back propagation algorithm were obtained and used. The generation
networks are mainly implemented in training the real distribution samples and
discriminant network is in charge of clearly elaborating the created samples by the
network generation of the true samples. The nature of the creative network is a deep-
CNN and the discrimination network is the CNN.
1.10 Research Motivation
While the specific transformations in the medical environment are provided by
the obtained data, it generates certain challenges. Actually, the inflow of computerized
medical records locates an extra tension on medical caregivers for data management in
an effective manner by assuring the integrity, security and interoperability during the
process of arranging the regulations and policies correspondingly. Few challenges
faced by health clinicians in day- today are discussed below.

17
 Fragmented Data: The medical image data is available in terms of structured
or non-structured form and exists in several formats such as digital formats,
images/video/multimedia files, specialized formats, papers or spreadsheets named as
DICOM format, which is used in the process of imaging of MRIs X-rays, and ECGs
MRI scans. The collection of data and aggregation is equally segmented, which makes
the collection and extraction of the data features a real complication. The data are
gathered by the payers, patient, and healthcare providers, but centralizing that
information collected from data has some complications. The duplicate or false data
may evolve, which makes the profiles of healthcare givers to be invalid. The lack of
factors such as supporting, monitoring, and understanding leads to high risks.
Changes to Data and Quality: There will be a constant change in health-care
data. The patients may have several treatments and diagnosis along a particular mode
of time. The diagnosis and operations comprise of overtime, for those new models
known as Telehealth models and new data were generated. The data is the fuel for
operation of medical information system, improving the data integrity means improving
the working of the system potential. The data classified by duplication, invalidity,
inaccuracy, incompleteness and poor standardization is known as dirty data, which may
lead to mistake in data processing and it c -
being, affecting the specialty of healthcare providers and leads to costly lawsuits.
Additionally, the poor-quality data may tend to improper communication internally and
externally, treatment delay and improper payments. The false data may also lead to
affect the potential of proceeding the health research and improvement steps by giving
wrong information.
Data Interoperability: For the proper utilization of the data, it requires
sequential meanings and it needs to transfer among peoples, systems and managements,
this ideology is named as data interoperability. It further, arranges the data of patients
from a range of computerized healthcare systems and allows the patients to take over
the entire control in the sharing of medical records, by improving the research and
creativity. The standardized HER data is needed, which was provided by the medical
care providers but, the medical testing providers working externally uses alternative
systems and protocols. Hence, the sharing of the entire medical record and creating

for information sharing.

18
 Lack of Integration between Clinical and Administration Systems: There
exist a consolidation gap in-between data of patient and management, internally.
Health-care records obtained by clinicians and on the hospital floor is considered in
insurance claims and billing sections accurately. The configuration of data management
system is carried out to assure that there is a match of diagnosis codes and proper care
provided is accurately monitored for both management purposes and analysis.
Data Security: Data security is considered as the major objective for any kind
of business, specifically the medical care industry. During every movement phase from
one place to another, the data should be resolved. This process makes the data security
a broad-scale concern in day-to-day industries. The data security involves the training
the data security using a potential hardware and software methodologies and the
accuracy and validity of the considered data are assured. The data security concept with
high potential capacity comprises of features such as collection of the needed
information only, protecting the stored data and demolishing any data information that
is not needed further. Due to the increasing cyber-attack cases, invasion of personal
data and data breaches, the management should admit the new and involving privacy
and safety regulations. The breach of sensitive contents leads to criminal and civil
penalty.
Regulations and Compliance: Medical data is a sensitive platform and should
obey the government regulations. The challenges in finding of data and improper
quality of data will make it more complicate to process the audits needed, to meet
regularity needs and control the data diversity of medical care providers, who use the

data governance in medical care organizations helps in better management and usage
of data creates operations for solving the data problems and also allows the users to
make decisions on the basis of highly qualified data. Still, all of this starts with
improved data collection and assuring that the collected data are accurate, updated, and
complete and in reliance with HIPAA regulation standards. A well-implemented
HIPAA issues web form solution can be indispensable in allowing the healthcare
managements to manage data collection, along with new forms for patients, HIPAA
release forms, patient medical history forms, consent forms and contact update forms.
The process of digital imaging plays a significant process in analysis of medical
data, which supports in identifying information that is free from mistakes and provides
precise identification and diagnosis. Various image processing methodologies have

19
been used in the health care environment for data analysis. The selection of proper
methodology to perform effective de-noising without any information loss or changing
of edges is considered as a complicate process. The aim lies behind the attainment of
noise free medical images is to preserve the edges of tissues and boundaries at same
time in securing small changes in the same tissue areas. There is no method obtained to
be effective till now considering all factors such as edge preservation, robustness, noise
reduction, elimination, cost of computation and user interaction. The algorithms such
as adaptive filters create promising results for images of brain disease but gives bad
results for other disease images known as cancer images and X-Rays. Hence, the
process of noise removal from these data images is an essential and major step in the
digital image processing technique. The type of noise present in the image analyzes the
model in denoising. The method of denoising mainly aims in enhancing the image
quality by improving indication to noise ratio and securing the minute details
specifically. For eliminating the degradation effects takes place in the captured image,
several image restoring methodologies and techniques are obtained, which is used in
minimizing, eliminating or removing the noise present in data image. The image quality
tends to decompose due to low clarity and due to noise available or by blurring effects.
Blurring takes place because of the imperfect, incomplete or incorrect formation of
images, where the noise minimizes the visual quality of data image by disturbing the
original signals.

1.11 Research Objectives

The significant objectives of this thesis are elaborated as follows,
Propose Optimized Deep Learning-enabled Hybrid Logistic Piece-Wise Chaotic
Map for Secured Medical Data Storage System.
Identify an Adaptive Noise Removal Framework for Medical Images
Apply Hybrid Deep Learning models to Optimize multi-disease prediction

1.12 Contributions of the thesis

The main purpose of the noise removal techniques is described below.
To identify and de-noise all types of noises from the MR images and to secure the
medical data information. At first, the problem formulation in de-noising of images
and data securing are analyzed by utilizing the research outcomes and for further

20
 processing, an adaptive noise removal framework is implemented for enhancing the
performance of the model. The comparative analysis is processed between SNR and
peak SNR and as a result, it achieved high accuracy in noise removal and also
ensures the future scope of the project.
To develop a new medical image storage system with improved security and
privacy protection during the process of transferring the medical images by utilizing
the optimized deep learning architecture along with the support of an enhanced
optimization algorithm and to implement a hybrid chaotic map named HLPWCM
by combining the 2DLCM and PWLCM for image encryption with a deep learning-
based key generation for securing the medical images. Then, enhanced optimization
algorithm named BF-CSHO is introduced for optimizing the weights of RNN-based
key generation for improving the overall performance by minimizing the difficulty
in image transmission. Thus, the experimental assured effective performance in
noise reduction.
To develop the process of multi-disease prediction by utilizing the improved DL
concept. For this, several datasets in relation with different diseases are collected
and pre-processing was carried out with data normalization avoiding the redundant
data. Then, the proposed JA-MVO algorithm is implemented by combined meta-
heuristic algorithms such as MVO and JA and the weighted feature extracted and
the weights of DBN and RNN were optimized. Hence, the resultant analysis
provides high performance in developed method, when comparing the available
traditional models for multi-disease prediction.

1.13 Organization of the thesis

The organization of the work of noise removal model is explained as shown
below.
`Chapter 1 describes the introduction of medical data processing in data images and
its available methodologies, then it further describes about the need for noise
removal, techniques used in noise removal, medical data storage systems and their
benefits and the multi-disease prediction model. It also gives the explanation about
its research motivations and objectives, contribution, and organization of the thesis.
Chapter 2 reviews the literature survey for the medical data storage systems, where
it briefly elaborates the noise removal approaches. Further, the categorization,

21
research gaps and challenges related to the medical data storage system were
provided.
Chapter 3 describes the improved de-noising techniques for medical images by
collecting images and analyzing them by using proposed models and then, the
results were compared for accuracy and further recommendations were provided.
Chapter 4 explains the Optimized Deep Learning model in medical data storage and
processing and the HLP-WCM were done to secure the stored medical data. The
experimental results provide high performance in effective noise removal of
medical data images.
Chapter 5 implements an improvised DL concept for multi-disease prediction
model. Additionally, the JA-MVO algorithms were proposed to provide better
results for the noise removal of medical data storage system. The results provide
superior performance in proposed model.
Chapter 6 concludes the medical data processing model along with its future scopes
and objectives.

22