Neural Networks & Medical Image Segmentation in Deep Learning

Arham Tanzeel Muhammad Amjad Khan

F24-3090 F24-3046
Arhamtanzeel@[Link] amjad97@[Link]

Department of CS and IT
Lahore Leads University

Abstract Introduction
Image segmentation is a sort of digital An input picture is segmented into numerous
image processing with applications in image segments that have a strong correlation with the
analysis, augmented reality, and computer image's area of interest (ROI). The goal of
vision. The discipline of medical image medical image segmentation is to make sense of
an input image so that it can be used to examine
analysis is growing, making it increasingly
anatomy and identify the region of interest
difficult to segment organs, illnesses, or
(RoI), measure tissue volume to determine
anomalies in medical pictures. Tumors, drug tumour size and aid in drug dose decisions, plan
dosages, and radiation exposure doses may treatment before applying radiation therapy, or
all be managed with the use of medical calculate radiation dose. Image segmentation
image segmentation. Medical images are improves analysis by defining the region of
difficult to segment because of the many interest in medical imaging. Segmentation
artefacts present. Deep neural models have techniques can be used for brain tumor
recently been shown to be effective in a boundary extraction on MRI images[1],
number of image segmentation tasks. This cancer detection on biopsy images, mass
massive increase is due to the success and segmentation on mammography, edge
high performance of deep learning detection on coronary angiograms,
algorithms. This study presents a review of segmentation of pneumonia affected area on
the research on the topic of medical image chest X-rays, etc. Various medical picture
segmentation using deep convolutional segmentation methods have been developed.
neural networks. The study examines a New technologies have been created in
number of commonly used medical image response to a lack of trained personnel, and
datasets, as well as the criteria used to they are in great demand.
evaluate segmentation tasks and the
outcomes of different CNN-based Previous image segmentation models relied
comparisons. The present study looks at a on traditional image processing methods like
variety of challenges in the area of medical thresholding and edge-and-region-based
image segmentation, as well as current algorithms. Using the thresholding
reviews and surveys, as well as several state- approach, pixels were allocated to different
of-the-art methodologies available in the groups depending on the range of values
literature. they possessed. [2]. In region-based
segmentation algorithms, pixels with similar
values and groups of pixels with differing following are some of the contributions
values were shared. Segmenting medical found in the document:
images is a difficult process due to multiple
First and foremost, the current research
limits imposed by medical imaging
examines the current status of deep neural
modalities, the nature of the disease, and
network topologies utilized for medical
unique biological variations. [6].
picture segmentation, as well as their
Professionals can undertake medical image
strengths and drawbacks. Second, the study
analysis, and there is a scarcity of medical
discusses the datasets for medical picture
imaging experts. Deep learning networks
segmentation that are publicly accessible.
have aided in the creation of new pictures in
Finally, it lists the many performance
recent years.[3][4] On a variety of
indicators that may be used to assess deep
prominent datasets, depth pictures from
learning segmentation models. Finally, the
neural networks have reached great quality.
publication discusses the most pressing
Semantic segmentation and instance
difficulties in picture segmentation, as well
segmentation are two types of image
as state-of-the-art solutions.
segmentation algorithms. Semantic
segmentation may be thought of as a Neural Network Structures in Deep
problem of pixel categorization. Each pixel Learning
in the picture is tagged with a certain class
using this segmentation approach. Each item The most essential method for artificial
of interest present in the input picture is intelligence is deep learning. A deep
detected and delimited using instance learning method employs numerous layers
segmentation. to create an artificial neural network. An
artificial neural network (ANN) is made up
The current literature on medical picture of three layers: input, hidden layer (s), and
segmentation is reviewed in this publication. output. The signal is received by the
The paper gives an overview of several deep network's input layer; an output layer makes
learning-based picture segmentation judgments about the input; and hidden layers
algorithms and discusses their architecture. execute calculations between the input and
Medical image segmentation challenges output layers. A deep neural network has
have been reviewed by a number of writers. numerous hidden layers between the input
All of the review literature mentioned above and the base layer.
covers the numerous deep neural networks.
This overview paper not only summarizes Convolutional neural networks (CNNs) are a
the various deep learning methodologies but kind of neural network. A CNN is made up
also includes information on the various of three primary neural layers: a
medical imaging datasets used to train deep convolutional layer, a pooling layer, and a
neural networks, as well as an explanation of fully connected layer. Each layer has a
the metrics used to train deep neural distinct purpose. The convolution layer
network models. The current paper also recognizes characteristics in an image such
examines the different issues that DL-based as edges and other visual aspects. The
image segmentation models encounter, as mathematical action of local neighbor
well as their current solutions. The multiplication of an image pixel by kernels
is performed by the convolution layer. To
produce its feature maps, CNN employs a convolutional layer. By making the final
variety of kernels that convolve the input fully connected layer of CNN a fully
picture. The pooling layer decreases the convolutional layer, existing differences in
input data's spatial dimensions (width and CNN topologies may be adjusted in FCN.
height) for the neural network's succeeding Instead of providing patchy predictions, the
layers. This has no effect on the depth of the developed model may provide a spatial
data. Subsampling is the term for this segmentation map and a pixel-dense
process. The computing needs for the prediction from the full-size input picture.
following several layers are reduced as a The model upscale feature maps from the
result of this size decrease. previous level and merges them with feature
maps from prior levels using skip joins. As a
On NN, fully linked NN layers relax at a
result, the model generates comprehensive
high level. To produce the final results,
segmentation at the same time. The typical
these layers combine the numerous feature
FCN model, on the other hand, has the
responses of the provided input picture.
following limitations: [6] [7]
Some of the CNN models that have been
It's too slow for real-time inference and
reported in the literature include Alex Net,
doesn't account for global context
Google Net, VGG, Inception, SequeezeNet,
information well. The produced feature
and Dense Net. Each network employs a
maps in FCN have poor resolution, and the
varied number of convolutions and pooling
output is lowered as a result of propagation
layers, with key processing blocks
via alternating convolution and pooling
sandwiched in between. For the
layers. In FCN, this leads to low-resolution
classification jobs, CNN models were
predictions with blurring at object borders.
mostly employed. The photographs of the
[8].
brain were divided into three categories: In,
Squeeze Net, and Google Net. The
following factors restrict the performance of
CNN segmentation models:
CNN's fully linked layers can't manage
varying input sizes. Because the existence of
a number of items of interest is not fixed in
the picture segmentation challenge, the
length of the output layer cannot be constant
in a convolutional neural network with a
fully connected layer, so it cannot be utilized
for object segmentation tasks.
Convolutional network in its entirety. Only
layers of convolution exist in a completely
convoluted network (FCN).[5] Existing
differences in CNN architectures can be
modified in FCN by making the last fully
connected layer of CNN a fully
 are often used for scientific picture
segmentation:

U-Net. A down-sampling and up-sampling

component are included in the U-Net
version. The down sampling phase, which
uses an FCN-like structure to extract
functions, uses three convolutions to capture
context. To reduce the huge range of
calculated function maps, the up-sampling
component uses DE convolution. The
function maps created by the down sampling
or contracting component are provided as
inputs to the up sampling component to
avoid any information gaps. Precision
localization is possible because of the
symmetrical up sampling component. The
E-version creates a segmentation map that
classifies each pixel in the picture. Every
pixel present inside the picture has been
classified.
The following are the benefits of the U-Net
version:
 The U-Net version can do green
segmentation of images using a
limited number of labelled school
The Parse Net, a more advanced FCN, has images.
also been mentioned; it employs  The u-Net structure creates a true
international common pooling to create segmentation map by combining area
transnational context. In addition, strategies data from the down sampling
for adding conditional random fields and direction with contextual data from
Markov random fields into the DL structure the up sampling direction.
have been disclosed.  U-Net structure also have certain
limits. The maximum length of an
Encoder-Decoder Models are a kind of input picture is 572X572 pixels.
encoder-decoder. Encoder-decoder-based
modelling creates a two-level model to map The learning frequently slows down in the
record factors from the input to the output middle layers of deeper UNET designs,
areas. The encoder compresses the input x to which is why the community ignores the
a latent area representation, and the decoder levels with summary functions. The
predicts the output from that representation. version's pass connections enforce a
The following are the several types of stringent fusion approach, resulting in the
encoder-decoders based on total models that aggregation of the encoder and decoder
networks' scale function maps that are rules also predicts offset values. The R-CNN
identical. structure's methods have been proven. The
usage of the main RCN version is limited
Unique variations of the U-Net structure
due to the following factors:
have been suggested in the literature to
overcome these limitations: U-Net++, It can't be done in real time since it takes
Attention U-Net, and SD-UNet. around 47 seconds to train the community
for the sort of project of 2000 location ideas
It is also an FCN-based variant that is used
in a look at images.
for scientific image segmentation. There are
three aspects to the VNet structure: The selective seek set of rules is a set of
compression, decompression, and rules that have been predefined. As a result,
community. Convolution layers with knowledge acquisition is no longer possible
residual functions are used by the at that point. This should lead to a period of
compression community at all levels. unfavorable candidate placement
Volumetric kernels are used in these suggestions.
convolution layers. The decompression
The selective seek set of rules is a set of
community collects features from low-
rules that have been predefined. As a result,
decision characteristic maps and increases
knowledge acquisition is no longer possible
their spatial depiction. It provides multi-
at that point. This should lead to a period of
channel probabilistic segmentation for each
unfavorable candidate placement
foreground and heritage area.
suggestions.
It is also an FCN-based variant that is used
R-CNN in a hurry. The suggested regions of
for scientific image segmentation. There are
the snapshot overlap in R-CNN, and the
three aspects to the VNet structure:
same CNN calculations are repeated over
compression, decompression, and
and over. An inputted image and a hard and
community. Convolution layers with
fast list of item ideas are fed into the quick
residual functions are used by the
R-CNN stated above. Convolutional
compression community at all levels.
function maps are then generated by the
Volumetric kernels are used in these
CNN. After that, the ROI pooling layer
convolution layers. The decompression
reshapes each item idea into a constant-size
community collects features from low-
function vector
decision characteristic maps and increases
their spatial depiction. It provides multi- The function vectors are delivered to the
channel probabilistic segmentation for each model's final fully connected layers. Finally,
foreground and heritage area. The output the calculated ROI function vector is sent to
dense layer contains capabilities gathered the Softmax layer, which predicts the
from the image, which are then fed into the recommended elegance and offset values.
category set of rules for classification of the Because of the employment of a selective
item's mendacity in the location set of rules, the R-CNN is slower.
recommendation community. In order to
improve the precise level of the location Deep Lab Model
recommendation or bounding box, the set of
To extract capabilities from an image, the computer vision. Figure 5 depicts a deep
Deep Lab version uses pre-skilled CNN learning system as a block diagram. Data
versions, ResNet-101/VGG-sixteen, with collection is the initial stage in developing a
atrous convolution. The following are the deep learning system. After that, the data is
advantages of using atrous convolutions: parsed and preprocessed so that it may be
used in the following block in an appropriate
It controls the decision of function responses
manner. Training, validation, and test data
in CNNs
sets are separated from the preprocessed
It translates the image category community data. The model is chosen and trained using
into a dense function extractor without the deep neural networks. The model that has
need to learn any additional parameters and been trained is put to the test and assessed.
uses conditional random fields (CRF) to The system analysis is completed in the
provide nicely segmented output. conclusion. [3]
The architecture of this basic deep learning
model is employed in a variety of medical
applications, including picture segmentation.
Image objects are subdivided in the
segmentation process. The purpose of
medical image segmentation is to find
regions of interest (RoI) in images, such as
cancers and lesions. Automatic
segmentation of medical pictures is a tough
problem owing to the existence of numerous
artefacts such as artefacts, intensity
inhomogeneity, and so on. The literature has
presented a number of deep learning models.
Choosing a deep learning model is
Various deep learning-based segmentation dependent on a number of parameters,
methods are compared. The above- including the body part to be segmented, the
mentioned outstanding deep neural networks imaging modalities employed, and the kind
are recruited for exceptional applications. of illness, since various body parts and
Each version has its own set of advantages disorders have distinct needs.
and disadvantages. [1]. Table three provides
To segment cardiac MR images into left and
a short comparison of the best deep learning
right ventricular cavities and myocardium,
algorithms based entirely on image
TA's completely automated 2D and 3D
segmentation techniques.
CNN-based frameworks were demonstrated.
Deep Neural networks applications To partition brain tissue in MR images, the
in medical image segmentation scientists created a deep CNN with layers
that perform convolution, pooling,
Deep learning networks have aided in the normalization, and other functions.
detection and classification of images, object
recognition, picture segmentation, and
Christ et al. proposed a design in which the Image Segmentation Data Sets
liver was subdivided using two NCFs in the (Medical Image)
cascade, as well as the lesions being
segmented inside the ROI. A dense 3D Data is critical in deep learning models. A
conditional random field was used to create lot of data is needed for deep learning
the final segmentation. models. The use of data is crucial. Data
protection requirements for data collection
The scientists determined that by employing and labelling make medical imaging data
FCN, the network's speed is increased, and difficult to get, and it also necessitates
output scores are generated more quickly. extensive professional explanations.[10]
On CT scans, the authors segmented the Two-dimensional (2D) pictures, 2.5-
liver using FCN. The authors presented a dimensional (2D) images, and three-
full-channel spatial convolution and dimensional (3D) images are the three types
excitation module for the segmentation of of medical imaging data sets. Each piece of
pneumothorax in chest radiographs.[9] information in a 2D medical picture is
referred to as a pixel. Each piece in 3D
medical imaging is referred to as a voxel.
RGB pictures are referred to as 2.5D. 3D
pictures are occasionally displayed as a
succession of 2D slices in a sequential order.
[11] 3D voxels are represented by pixels in
CT, MR, PET, and ultrasound scans. JPEG,
PNG, and DICOM images are all
acceptable.
CT scans, ultrasound, MRI, mammography,
positron emission tomography (PET), and
X-rays of various regions of the body are
examples of medical imaging modalities.
MRI allows for the creation of images with
varying contrast by using various pulse
sequences. The interior anatomy of the
chest, liver, brain, pelvis, and abdomen may
all be seen via an MRI. X-rays are used in
computed tomography to collect information
about the structure and function of bodily
components.
Brain, abdomen, liver, pelvic, chest, and
spine diseases, as well as computed
tomography-based angiography
Mammography is a procedure that employs
X-rays to capture images of the breast's
interior anatomy. Chest X-rays (CXR) are
photographic pictures that show the internal
composition of the breast. They are created
by transmitting X-rays through the chest and Recall.
having those rays absorbed by varied The recall number shows what percentage of
quantities of various components in the all relevant outcomes the model successfully
chest. A list of the most important publicly categorized.
accessible medical imaging datasets is
provided.

Evaluation Matrices Recall = TP/TP+FN

A metric aids in the evaluation of any
planned model's performance. The
F1 result.
measurements show how accurate the
planned model is. The following are some of As stated in the equation below, the F1 score
the most frequently used metrics for offers information about the models'
evaluating the efficacy of any segmentation accuracy. The harmonic mean of the
method. accuracy and recovery numbers is what it's
called.
 True positives (TP) are removed,
which indicates that both the actual
and anticipated data classes are
correct. F1 Score = 2* Precession* Recall/
 The True Negative (TN) indicates Precession + Recall
that both the actual and anticipated
data classes are incorrect.
 The term "false positive" refers to a Precision in pixels Returns the proportion of
situation in which the actual data pixels successfully categorized by the model
class is false while the projected data in a given input picture.
class is true.
 The term "false-negative" (FN)
describes a situation in which the Pixel Accuracy = No. of Pixel properly
actual data class is true while the classified/ Total number of pixels
projected data class is false.

Different challenges related to the dataset

Precision. are the following:
Precision is an assessment statistic that Annotated dataset with a small number of
provides the percentage of genuine input entries. The data needed for training has
data situations that are represented in a been meticulously annotated. Several DL-
graph. based medical techniques rely heavily on the
dataset. Collecting a large number of
Precession = TP/ TP+FP
annotated medical pictures is challenging in
medical image processing. Annotating imbalance in the data set.
modern medical imaging is also time- Second, the problem of record
consuming, costly, and demands skill. imbalance may be addressed by
Several major datasets are open to the modifying the rating or
public. Table 2 contains a list of some of performance measure.
these documents. More advanced datasets, (b) New data samples may be
capable of handling dense objects and created using data augmentation
allowing for improved training of DL methods.
models, are still needed.[12] Existing 3D (c) The data set's class imbalance
datasets are often small and few are issue may be addressed by
synthetic, necessitating the development of merging minority classes.
more complex datasets.[13] Image
Several methods are utilized to adjust
magnification changes, such as B., rotating
intensity in homogeneities, and numerous
photos from various angles, flipping images
non-parametric strategies are suggested in
vertically or horizontally, or cropping and
the literature. [15]. To eliminate
slicing images, may be used to enhance the
homogeneities, utilize the pre-filter
size of existing medical image databases.
procedure before segmentation. By
These improvements using some
enhancing the scanning equipment, the
approaches, you may improve the
intensity of homogeneities may be resolved.
performance of your system. (b) The data
limitation issue may be solved by using The image's texture has a lot of subtleties.
transfer learning from efficient models. (c) During the image processing process, a
Finally, data from various sources is variety of artefacts might appear in medical
combined.[14] pictures. Noise is introduced into the picture
as a result of the many sensors and electrical
Classes in records are unbalanced. Several
components that are employed to collect it.
publicly accessible medical imaging datasets
The gray levels of the acquired picture
include an inherent class imbalance.
might be quite close together, and the image
Training the LD model with highly skewed
borders can be very weak. On dermoscopic
data is challenging, and the model's
pictures, there may be overlap and
accuracy is deceiving. Inpatient data, for
abnormalities in the appearance of skin and
example, shows that the condition is rather
hairlines. In medical pictures, locating key
uncommon, with just 10% of patients tested
areas may be challenging because of the
having it. Because most patients do not have
many variables involved. Prior to
the illness and attain local minima, the
segmentation, several image enhancing
overall accuracy of the suggested model
methods are employed to eliminate various
would be high.
artefacts and noise from the picture. With
The issue of class imbalance may be this image enhancement technology, the
remedied by picture noise is reduced and the integrity of
the image edges is preserved.
(a) Oversampling the data, with the
quantity of oversampling varying Challenges[16] with DL models. ([17] The
according to the degree of most pressing issues in training a deep neural
network (DNN) for effective segmentation of models, the leak gradient issue is more
medical pictures include: apparent. The issue of a diminishing
Over fitting of the model. A model is said gradient has a number of solutions.
to be over fit if it is able to learn more (a) The extra losses and the initial loss of the
accurately from the ensemble of training hidden layer are combined to raise the
data than it could if it were taught from the gradient value by scaling the intermediate
raw data alone. For example, while the hidden layer output using DE convolution
model is being trained with tiny amounts of and softmax.
data, this might happen. Data augmentation
strategies may be used to increase the data (b) In addition, the gradient vanishing issue
set's size in order to avoid over fitting. A may be avoided by properly initializing the
fraction of the random set of neurons in the network's weights. The intricacy of the
network may be dropped out of each computer program. [20]
iteration in order to prevent over-fitting. The feature analysis method used by deep
Memory Efficient Models [18]. Memory learning must be very computationally
requirements for medical picture efficient. [21] Powerful computers and
segmentation models are high. These models graphics processing units are needed to run
must be simplified in order to be compatible these algorithms. Supercomputers are
with certain devices, such as mobile phones. required to train some of the most advanced
The storage needs of a DL model may be algorithms, which may not be accessible. In
reduced by using simpler models and model order to deal with these issues, the
compression methods. Workout time. researcher must take into account a
Training a deep neural network takes a long restricted number of characteristics.
time. Fast convergence of the training time
for deep NN is essential in picture
Future
segmentation.[19] Automated picture segmentation using
machine learning and deep learning has
(a) Batch normalization is the answer to
gone a long way from manual segmentation.
this issue. If you want to find the
A vast number of photos may be segmented
pixel values around 0 in a picture,
using ML/DL-based techniques. In the
you may subtract the pixel values
photos, it aids in the identification of key
from the image's mean value. Rapid
things and the diagnosis of illnesses. Image
convergence may be achieved by
segmentation methods mentioned in this
using this method
work may be used on a variety of datasets
(b) To further speed up convergence,
by future researchers. Using publicly
consider using pooling layers to
accessible datasets, future research might
reduce the number of parameters in
compare the various deep learning models
your model.
presented in the article against each other.
The gradient disappears. Leaky gradients are The image segmentation model may also be
a concern for deep neural networks. As a improved by varying the number of layers
result, the ultimate gradient loss can't be and classifiers used in the picture
transmitted to the preceding layers. In 3D processing. Even though picture
segmentation has been improved, there is [3] A. Sengur, U. Budak, Y. Akbulut, M.
still room for improvement. In the future, Karabatak, and E. Tanyildizi, “A
researchers will be able to experiment with survey on neutrosophic medical
new deep learning model architectures. image segmentation,” Neutrosophic
Set Med. Image Anal., pp. 145–165,
Conclusion Jan. 2019, doi: 10.1016/B978-0-12-
818148-5.00007-2.
Disease diagnosis based on medical imagery
has been the focus of a recent study. For [4] C. D. Mathers and D. Loncar,
“Projections of global mortality and
medical picture segmentation, we describe
burden of disease from 2002 to
the most commonly used DL-based models, 2030,” PLoS Med., vol. 3, no. 11, pp.
highlighting their strengths and weaknesses. 2011–2030, Nov. 2006, doi:
Additionally, an overview of the various 10.1371/[Link].0030442.
medical picture datasets utilized for illness
segmentation and several performance [5] K. L. Ng, J. Yazer, M. Abdolell, and
metrics used to assess the performance of P. Brown, “National survey to
identify subspecialties at risk for
the image segmentation algorithm is given
physician shortages in canadian
here. Deep networks are used to segment
academic radiology departments,”
medical images, and the research also covers Can. Assoc. Radiol. J., vol. 61, no. 5,
the many hurdles and state-of-the-art pp. 252–257, Dec. 2010, doi:
solutions to these issues. 10.1016/[Link].2010.02.007.
The results of this study will aid medical [6] S. Roy, A. Carass, P.-L. Bazin, and J.
researchers in developing neural network L. Prince, “Intensity inhomogeneity
topologies for diagnosing illness. Research correction of magnetic resonance
in the area of medical picture segmentation images using patches,”
based on deep learning and state-of-the-art [Link]
technologies has also highlighted possible vol. 7962, pp. 444–449, Mar. 2011,
issues. Medical image segmentation is an doi: 10.1117/12.877466.
important area of study, and the information [7] Q. Dou, H. Chen, Y. Jin, L. Yu, J.
in this review contains both reference Qin, and P. A. Heng, “3D deeply
material and new research. supervised network for automatic
liver segmentation from CT
References volumes,” Lect. Notes Comput. Sci.
(including Subser. Lect. Notes Artif.
[1] “A survey on neutrosophic medical Intell. Lect. Notes Bioinformatics),
image segmentation - ScienceDirect.” vol. 9901 LNCS, pp. 149–157, 2016,
[Link] doi: 10.1007/978-3-319-46723-
e/article/pii/B9780128181485000072 8_18/TABLES/2.
(accessed May 26, 2022).
[8] “P. Malhotra and E. Garg, ‘Object
[2] S. Yuheng and Y. Hao, “Image detection techniques: a comparison,’
Segmentation Algorithms Overview,” in Proceedings of the 2020 7th
Jul. 2017, doi: International Conference on Smart
10.48550/arxiv.1707.02051. Structures and Systems (ICSSS), pp.
1–4, IEEE, Chennai, India, July 2020
 - Google Search.” segmentation,” Lect. Notes Comput.
[Link] Sci. (including Subser. Lect. Notes
hl=en&q=P.+Malhotra+and+E. Artif. Intell. Lect. Notes
+Garg, Bioinformatics), vol. 9351, pp. 234–
+“Object+detection+techniques: 241, 2015, doi: 10.1007/978-3-319-
+a+comparison,”+in+Proceedings+of 24574-4_28.
+the+2020+7th+International+Confer
ence+on+Smart+Structures+and+Syst [14] N. Navab, J. Hornegger, W. M.
ems+(ICSSS),+pp.+1–4,+IEEE, Wells, and A. F. Frangi, “Medical
+Chennai,+India,+July+2020 Image Computing and Computer-
(accessed May 26, 2022). Assisted Intervention - MICCAI
2015: 18th International Conference
[9] N. Dhungel, G. Carneiro, and A. P. Munich, Germany, October 5-9, 2015
Bradley, “Deep Learning and proceedings, part III,” Lect. Notes
Structured Prediction for the Comput. Sci. (including Subser. Lect.
Segmentation of Mass in Notes Artif. Intell. Lect. Notes
Mammograms,” Lect. Notes Comput. Bioinformatics), vol. 9351, no. Cvd,
Sci. (including Subser. Lect. Notes pp. 12–20, 2015, doi: 10.1007/978-3-
Artif. Intell. Lect. Notes 319-24574-4.
Bioinformatics), vol. 9349, pp. 605–
612, Oct. 2015, doi: 10.1007/978-3- [15] L. J. Muhammad, E. A. Algehyne, S.
319-24553-9_74. S. Usman, A. Ahmad, C.
Chakraborty, and I. A. Mohammed,
[10] A. Andreopoulos and J. K. Tsotsos, “Supervised Machine Learning
“Efficient and generalizable statistical Models for Prediction of COVID-19
models of shape and appearance for Infection using Epidemiology
analysis of cardiac MRI,” Med. Dataset,” SN Comput. Sci., vol. 2, no.
Image Anal., vol. 12, no. 3, pp. 335– 1, pp. 1–13, Feb. 2021, doi:
357, Jun. 2008, doi: 10.1007/S42979-020-00394-7/FIGU
10.1016/[Link].2007.12.003. RES/11.
[11] P. F. Christ et al., “Automatic Liver [16] E. Goceri, “Challenges and Recent
and Tumor Segmentation of CT and Solutions for Image Segmentation in
MRI Volumes using Cascaded Fully the Era of Deep Learning,” 2019 9th
Convolutional Neural Networks,” Int. Conf. Image Process. Theory,
Feb. 2017, Accessed: May 26, 2022. Tools Appl. IPTA 2019, Nov. 2019,
[Online]. Available: doi: 10.1109/IPTA.2019.8936087.
[Link]
[17] N. Sharma et al., “Automated medical
[12] “CVPR 2017 Open Access image segmentation techniques,” J.
Repository.” Med. Phys., vol. 35, no. 1, p. 3, Jan.
[Link] 2010, doi: 10.4103/0971-6203.58777.
_cvpr_2017/html/Chollet_Xception_
Deep_Learning_CVPR_2017_paper.h [18] F. N. Iandola, S. Han, M. W.
tml (accessed May 26, 2022). Moskewicz, K. Ashraf, W. J. Dally,
and K. Keutzer, “SqueezeNet:
[13] O. Ronneberger, P. Fischer, and T. AlexNet-level accuracy with 50x
Brox, “U-net: Convolutional fewer parameters and <0.5MB model
networks for biomedical image size,” Feb. 2016, doi:
 10.48550/arxiv.1602.07360.
[19] D. S. Kermany et al., “Identifying
Medical Diagnoses and Treatable
Diseases by Image-Based Deep
Learning,” Cell, vol. 172, no. 5, pp.
1122-1131.e9, Feb. 2018, doi:
10.1016/[Link].2018.02.010.
[20] “Faster R-CNN: Towards Real-Time
Object Detection with Region
Proposal Networks.”
[Link]
015/hash/14bfa6bb14875e45bba028a
[Link] (accessed
May 26, 2022).
[21] B. Van Ginneken, B. M. Ter Haar
Romeny, and M. A. Viergever,
“Computer-aided diagnosis in chest
radiography: A survey,” IEEE Trans.
Med. Imaging, vol. 20, no. 12, pp.
1228–1241, Dec. 2001, doi:
10.1109/42.974918.