Добавил:
Sekretar
kiopkiopkiop18@yandex.ru
t.me/Prokururor I Вовсе не секретарь, но почту проверяю
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_145_библиотеки_им_акад_М_И_Перельмана
.pdf
Disease Prediction using Machine Learning, 2024, 103-130 103
https://t.me/med1917
CHAPTER 7
Applying Deep Learning and Computer Vision for
Early Diagnosis of Eye Diseases
Shradha Dubey
1
Department of Computer Science and Engineering, Madhav Institute of Technology and Science,
Gwalior, Madhya Pradesh, India
Abstract: Medical image processing has a significant role in clinical investigation and
recent medical research. An appropriate image-based medical assessment helps to
analyze or detect critical diseases early, as it has a high value of medical information.
In this study, medical imaging is reviewed for the diagnosis of eye diseases using
computational intelligence. However, the identification of these diseases using
traditional image processing is quite complicated. Nowadays, various machine learning
and deep learning approaches are developed for the detection of different eye diseases
which are helpful for the detection of the diseases at an early stage. Research showed
that eye disorders are more serious in emerging or underdeveloped nations due to
inadequate healthcare facilities and skilled health workers. An estimate of 45 million
people around the world are blind and the tragic fact is that only 75% of these cases are
curable. Moreover, the doctor-patient ratio around the globe is about 1: 10,000.
Therefore, it takes an hour to create a screening system for the identification of these
illnesses. Ophthalmology is close to making breakthroughs in evaluating, diagnosing,
and treating eye diseases. Additionally, many eye and vision problems show no
obvious signs. As a consequence, people are often unaware that problems exist. Early
detection of diseases is a primary concern as they could be easily cured before leading
to severity. This research paper focuses on detecting eye illnesses, such as Diabetic
retinopathy, Diabetic Macular Edema, Glaucoma, Age macular Degeneration, Retinal
Vascular Occlusions, and Retinal Detachment. The authors explore various algorithms,
imaging modalities, and challenges in this context. The study aims to raise awareness
about eye disorders leading to blindness using computer vision, image processing, and
deep learning techniques. It also investigates how these machine learning and deep
learning approaches can aid in early disease diagnoses for effective treatment before
vision loss occurs.
1,*
and Manish Dixit
1
Keywords: Age-related macular degeneration, Cataract, Deep learning, Diabetic
retinopathy, Glaucoma, Heidelberg retinal tomography, Optical coherence
tomography, Ultrasound imaging.
*
Corresponding author Shradha Dubey: Department of Computer Science and Engineering, Madhav Institute of
Technology and Science, Gwalior, Madhya Pradesh, India; E-mail: dubeyshradha29@gmail.com
Geeta Rani, Vijaypal Singh Dhaka & Pradeep Kumar Tiwari (Eds.)
All rights reserved-© 2024 Bentham Science Publishers
104 Disease Prediction using Machine Learning Dubey and Dixit
https://t.me/med1917
INTRODUCTION
Clinical research and contemporary medical research both benefit from medical
image processing. An appropriate image-based medical assessment helps to
analyse or detect critical diseases early, as it has a high value of medical
information. This study presents a review of medical imaging for diagnosing eye
diseases through computational intelligence. Traditional image processing
methods have proven challenging for disease identification. However, the
introduction of diverse machine learning and deep learning approaches has
revolutionized the detection of various eye diseases, enabling timely identification
and diagnosis. Further sections in this chapter delve into the details of these
approaches and their significant contributions to early detection of eye disorders.
MOTIVATION
Vision is a crucial human sense, but an increasing number of patients are affected
by retinal diseases daily. A healthy retina is essential for central vision [1].
Detecting these diseases early can protect the retina from damage, but it presents a
significant challenge to medical science. With the growing prevalence of ocular
diseases due to demographic aging and poor dietary habits, there is a pressing
need for effective detection methods. This research paper reviews various
computational intelligence techniques utilized to classify healthy and diseased
retinal images and categorize the severity of retinal diseases based on ocular
image abnormalities. Precise classification of retinal conditions can reduce
unnecessary hospital visits, particularly amid the rising cases of infectious
diseases. The proposed approach aids in diagnosing the type of retinal disease and
assessing its severity level, facilitating timely treatment and improved patient
care.
This chapter’s primary contribution lies in discussing the effectiveness of deep
learning and medical imaging for timely disease detection. Additionally, it
explores various methods and techniques to capture eye images using different
machines and approaches. This chapter extensively explores various imaging
systems used in diagnosing different diseases, each with its advantages and
limitations. It also provides a detailed analysis of numerous eye diseases,
including their causes, risk factors, and symptoms, aiming to facilitate early
diagnosis and proper treatment. The literature discusses several deep neural
networks tailored for specific disease detection, enabling swift and accurate
diagnoses even in urban or rural settings without the need for specialized
professionals. Following detection, patients can promptly seek medical advice for
appropriate treatment. These deep learning models offer rapid reports, unlike
Diagnosis of Eye Diseases Disease Prediction using Machine Learning 105
https://t.me/med1917
other cutting-edge methods that may take 2-3 days for results, which could be
potentially harmful.
The chapter is structured into several sections. Section 2 presents an introduction
to deep learning. In Section 3, a concise review of the existing literature is
provided. Section 4 delves into the discussion of various imaging equipment
necessary for capturing eye images to compile the dataset for disease diagnosis.
Section 5 offers a comprehensive examination of different eye diseases and the
applications of computer vision and deep learning in their detection. Moreover,
the chapter addresses research challenges in Section 6 before concluding with a
summary of key findings.
TECHNICAL ASPECTS OF DEEP LEARNING
The term “computer-aided diagnosis” refers to analyzing relevant clinical data
using computer-based filters or tools to identify disease-related patterns [2]. In
this study, various deep learning algorithms are explored, particularly in the field
of computer vision applied to eye images. For image processing applications [3],
Convolutional Neural Network (CNN) models are extensively used and studied
by researchers. These CNN models fall into different categories, including LeNet,
AlexNet, VGGNet, GoogleNet, InceptionV3, ResNet, and more [4].
Convolutional Neural Networks (CNNs) are deep neural networks used to
enhance traditional image processing tasks. They are widely popular due to their
reduced pre-processing requirements compared to other algorithms. CNNs
effectively capture temporal and spatial dependencies in an image using filters. A
typical CNN consists of three main layers: an input layer, hidden layers, and an
output layer. The network can have multiple hidden layers, each performing
different functions like feature extraction, classification, and feature flattening.
The basic block diagram of a CNN is shown in Fig. (1), illustrating its application
in disease classification. Initially, the convolution function is applied to the input
images, where a filter is used to activate specific features. Typically, a non-linear
ReLU (Rectified Linear Unit) activation function is employed to make the
network robust with various types of input images. Next, the pooling function is
used to reduce computation overhead by considering only relevant parameters.
Finally, the flatten function converts the spatial dimension into a channel
dimension, and a fully connected layer with an activation function, often softmax,
is used to obtain the classified image as output.
To construct a robust Deep Learning (DL) model, two key components are
essential: the 'brain' (CNN) and the 'dictionary' (the datasets). Initially, CNNs
faced computational challenges, but with the advent of GPUs, these methods
106 Disease Prediction using Machine Learning Dubey and Dixit
https://t.me/med1917
became more efficient, yielding better results in less time. There are two types of
DL solutions provided for biomedical image analysis as given below:
Image Classification: In this approach, images are provided to the DL network
along with their associated diagnoses, labels, or stages. The network learns to
classify these images into different categories based on the provided information
[4].
Semantic Segmentation: Semantic segmentation is another approach used in DL
for medical image analysis. Instead of classifying the entire image, this method
focuses on defining specific regions or areas within the image that correspond to
medical symptoms related to the disease's conditions. By segmenting the image,
the network can clearly identify and delineate different areas of interest.
Fig. (1). A typical functioning of CNN model.
Benefits of Deep Learning
Research has demonstrated that deep learning (DL) is highly effective in detecting
and diagnosing various ophthalmic diseases, often outperforming human graders.
DL represents an advancement over traditional artificial neural networks (ANNs)
as it involves creating networks with multiple layers, enabling more sophisticated
learning. In the past, researchers faced challenges in manually detecting diseases,
identifying regions of interest in images, and classifying characteristics like
microaneurysms and hemorrhages in diabetic patients. Writing hand-coded
instructions for algorithms was a time-consuming and laborious task. However,
the invention of computer-aided (CAD) systems has revolutionized this process.
With CAD systems, researchers can now input labeled images into the system,
and it automatically classifies them into categories like proliferative and nonproliferative Diabetic Retinopathy (DR).
Deep Learning models, which train automatically, require large datasets to
Diagnosis of Eye Diseases Disease Prediction using Machine Learning 107
https://t.me/med1917
achieve higher accuracy. In ophthalmology, significant amounts of diverse data
can be collected from different imaging equipment and diverse patients with
varying ages, sex, and ethnicity. While deep learning algorithms can recognize
classic characteristics of diseases like DR on their own, they may also discover
new patterns and offer valuable insights, leading to the concept of the “black box”
in deep learning [5]. Ongoing research aims to uncover how these algorithms
interpret and evaluate disorders.
Deep learning in imaging and clinical applications has been observed to produce
high-quality results and can handle unstructured data effectively [6]. In medical
imaging, many clinical applications are transitioning to deep learning due to its
superior performance over traditional approaches and its ability to process vast
amounts of complex data with great accuracy. This shift to deep learning is
particularly vital as the number of patients continues to rise, making it challenging
for doctors to diagnose rapidly. Deep learning aids in early disease detection,
allowing for timely treatment and better patient care [3].
LITERATURE REVIEW
Efforts have been dedicated to creating medical specialist systems that simplify
diagnostic procedures. These systems can offer accurate responses based on
predefined rules. However, relying on static rules hampers their ability to adapt to
new scenarios. To address this limitation, the focus shifted to deep learning,
where data is used to train deep learning-based algorithms. As a result, computer
research in various medical fields, especially ophthalmology, has expanded
rapidly.
Soheila Gheisari et al. (2021) proposed a study combining Convolutional Neural
Networks (CNN) and Recurrent Neural Networks (RNN) to detect spatial and
temporal features in sequential images. The dataset included 1810 fundus images
and 295 videos. By using both CNN and RNN, the researchers achieved better
accuracy compared to using either model alone, obtaining an impressive Fmeasure of 96.2% for classifying healthy and glaucoma images [7].
Turimerla Pratap and Priyanka Kokil (2021) introduced a reliable approach, the
Computer-Aided Cataract Diagnosis (CACD) method, for handling Additive
White Gaussian Noise (AWGN). This technique involved several independently
trained, local, and global support vector networks at different noise levels. The
network selection was based on the input image's noise level. Additionally, they
utilized a pre-trained CNN to extract features from fundus retinal images. The
research utilized the retina fundus image dataset from EYEPACS [8].
L. Cao et al. (2020) presented an approach for cataract assessment from retinal
108 Disease Prediction using Machine Learning Dubey and Dixit
https://t.me/med1917
images. They employed an enhanced version of Haar Wavelet to categorize retinal
data into non-cataract, mild, moderate, and severe cataract levels. The method
extracted more appropriate features, improving accuracy. To simplify
classification, they transformed the four-class problem into a two-class one using
a hierarchical approach. The final framework merged three sets of 2-class neural
networks. The study achieved accuracies of 85.98% for four-class and 94.83% for
two-class classification [9].
Abhishek Samanta et al. (2020) conducted research using the DenseNet model
with transfer learning on a small dataset of 3050 images from Kaggle for
classifying DR disease in four ways. Despite the limited number of images, their
CNN model exhibited excellent performance, particularly for real-time images.
The proposed model achieved high Cohens Kappa scores of 0.8836 and 0.9809 on
the validation and training sets, respectively [10].
In another research paper by Sambit S Mondal et al. (2020), the authors employed
the Generalized Improved Fuzzy Kohonen Clustering Network (GIFKCN)
algorithm to detect blood vessels in retinal images. They applied various preprocessing techniques such as Gaussian Filter and Morphological operations
before using the algorithm. The study, conducted on the DRIVE database, found
that the introduced model outperformed other models with better accuracy,
sensitivity, and low false-positive rates (0.979, 0.989, and 0.039, respectively)
[11].
Mary Dayana et al. (2020) worked on identifying anomalies in retinal fundus
images using deep learning algorithms. They developed a CNN-based model that
effectively segmented small lesions by analyzing image patches created by a
sliding window approach. The model used these patches to generate a probability
map, forecasting the various types of lesions. Compared to similar studies, the
proposed method demonstrated significantly higher accuracy and sensitivity [12].
Krishna Prasad et al. (2019) used a DNN model to detect glaucoma and DR in
retinal images, utilizing a pre-defined dataset (DR-Kaggle, Glaucoma Medimrg).
The research involved two phases: a training phase for evaluating the CNN model
and an implementation phase with a user-friendly GUI. The proposed model,
while not very complex, achieved an accuracy of 80% and provided a means to
recognize healthy or unhealthy input images through the GUI [13].
Mamta Juneja et al. (2019) conducted research focusing on the segmentation of
glaucoma images and non-glaucoma images using CNNs. The study utilized an
enhanced version of U-net called Glaucoma Network (G-net) for improved
segmentation, leading to good accuracy. Two different CNNs were employed,
Diagnosis of Eye Diseases Disease Prediction using Machine Learning 109
https://t.me/med1917
with one separating the optic disc and the other segmenting the optic cup from
retinal fundus images. The framework was tested on a set of 50 fundus images,
achieving high accuracy in disc segmentation (95.8%) and cup segmentation
(93%) [14].
In a study by Bilal Khomri et al. (2018), they developed a technique using the
PSO algorithm to extract blood vessels from both low and high-resolution datasets
containing healthy and DR images. Compared to the MSLD approach, their
defined technique effectively captured many small and tortuous vessels using
fewer sizes with varying pixel resolutions. This method proved valuable in
comparison to other methods, achieving promising performance on highresolution images. The work obtained sensitivity, specificity, and accuracy of
74.5%, 97.1%, and 94.2% for the DRIVE dataset, and 85.4%, 94.4%, and 93.5%
for the STARE dataset [15].
Suvajit Dutta et al. (2018) designed a model for the diagnosis of DR, which
incorporated a series of backpropagation, DNN, and CNN layers for training.
During the testing phase, the developed model outperformed other CPU-trained
models, primarily due to the one hidden layer concept. The network not only
achieved high accuracy but also successfully located various features like
exudates, microaneurysms, optic discs, and blood vessels, contributing to the
development of an optimal model for DR detection [16].
IMAGING MODALITIES
In this section, we explore the utilization of imaging equipment to capture images
of the eye's interior, specifically focusing on the retina. Various computer visionbased imaging techniques are discussed, which empower specialists to assess
retinal health and diagnose a wide range of eye diseases. Eye imaging constitutes
a crucial branch of medical imaging, involving specialized equipment for
photographing, scanning, or capturing images of the eye. The chapter delves into
diverse eye imaging methods, presenting their respective advantages and
disadvantages.
Ultrasound Imaging
Ultrasound imaging employs sound waves to create visual representations of the
interior of the human body. In the context of eye and orbit examination, highfrequency sound waves are utilized to produce high-resolution images of the eye
and its surrounding structures, even in the presence of opaque media [17].
Ophthalmic ultrasound typically operates at a frequency of 8-10 MHz, providing
much clearer images of the internal parts of the eye compared to a regular eye
110 Disease Prediction using Machine Learning Dubey and Dixit
https://t.me/med1917
checkup. This examination employs three modes of display: A-mode, B-mode,
and M-mode.
A-scan, also known as amplitude scan, is an ultrasound technique used for ocular
evaluation [18]. In this method, the tear film acts as an effective acoustic
transmission agent, eliminating the need for ultrasound coupling jelly. A single
sound beam is emitted from the transducer in A-scan, offering a more detailed,
level-by-level perspective of the eye compared to a standard eye exam. The
frequency of ultrasound used in A-scan is 8MHz. This type of ultrasound
examination is straightforward and simple.
On the other hand, B-mode, or brightness modulation, provides cross-sectional or
two-dimensional images of the eye and focuses on real-time evaluation of ocular
pathology. It operates at a frequency rate of 10MHz. Ultrasound imaging is
widely used to examine various medical conditions such as lungs, bone defects,
breast cancer, pulse rate, blood circulation in blood vessels, maternity
visualization, and ophthalmic evaluation [19].
Advantages
●
Ultrasound imaging is a valuable tool for rapidly assessing traumatic eye
injuries.
●
By utilizing 3D reconstruction, ultrasound imaging allows for accurate
determination of the size and location of objects.
●
It is particularly effective for evaluating the posterior segment in cases of hazy
media.
●
Ultrasound imaging can detect and distinguish between intraocular and orbital
lesions.
Disadvantages
●
Mastering the development of a three-dimensional image from two-dimensional
B-scan images is considered the most challenging step [20].
●
To achieve optimal imaging, increasing the depth requires the use of lower
frequency.
Optical Coherence Tomography (OCT)
OCT, a non-invasive imaging technique, captures cross-sectional retinal images
by combining low-coherence interferometry and near-infrared light with specific
coherence width [21]. By measuring the intensity of reflected light and echo time
delay, OCT enables ophthalmologists to differentiate between different layers of
the retina and determine the retinal nerve fiber layer's thickness, facilitating early
detection and diagnosis of retinal diseases. Even for healthy eyes, OCT can be
Diagnosis of Eye Diseases Disease Prediction using Machine Learning 111
https://t.me/med1917
utilized to establish a baseline image.
Similar to ultrasonography, OCT operates on the principle of using light instead
of sound waves. The current OCT imaging systems offer resolutions ranging from
1 to 15 µm. OCT can be categorized into two types: spectral domain and timedomain OCT. In time-domain OCT, image resolution and acquisition speed are
inversely proportional. Spectral domain OCT overcomes this limitation, enabling
simultaneous improvements in imaging speed and resolution [22]. Table 1
provides a comparison between time-domain and spectral domain OCT,
highlighting the advantages of spectral domain OCT.
Table 1. Types of OCT specification.
- Spectral Domain OCT
Detector Spectrometer Single Detector Faster acquisition Less Motion
Light Source 840 nm
Transverse Resolution
Axial Resolution 6-7 microns 10 microns
Scanning Speed
Maximum A-Scans
Scan Depth 2mm 2mm
10 microns
About 2800 A-scans per
second
8000 512
Time Domain
OCT
820 nm
20 microns
400 A-scans per
second
Advantages of Spectral Domain
High definition
-
Visualization of retinal layers are
good
Better registration
High number of scans with better
visualization
Slightly better penetration of light
For a typical OCT machine, its working principle and complete specifications can
be seen in Figs. (2) and (3), respectively.
Light
Source
Grating
Line
Camera
Fiber
Coupler
Transverse
scan
Spectral Interferogram
Fourier
Transform
Stationary
Reference
Mirror
Axial scan
Fig. (2). Principle SD-OCT [48].
112 Disease Prediction using Machine Learning Dubey and Dixit
https://t.me/med1917
Advantages
●
OCT takes short scanning time and also scans various ocular structures and
contactless.
●
Resolution up to 10 µm which is helpful for quantitative information.
Disadvantages
●
Scan quality depends on the operator's skills. Each scan must be taken in the
range.
●
Measurement is not accurate if the scan is not taken at the centre of the fovea
[23].
●
Penetration Depth is limited. It required a greater pupil diameter.
Color Fundus Photography
Color Fundus Retinal Photography is a technique used to capture photographs of
the inner surface of the eye using a fundus camera. This imaging method allows
for the observation and monitoring of diseases and their progression over time.
Fundus photography involves imaging the posterior part of the eye to examine
abnormalities associated with eye diseases and to monitor their development [24].
The specialized camera used for this purpose is commonly referred to as a fundus
camera or retinal camera, as illustrated in Fig. (4). It consists of a low-power
microscope and a camera to capture images of the interior portion of the eye,
including the optic nerve, macula, central, and peripheral retina.
Fig. (3). Specification of OCT.
Соседние файлы в папке Библиотека им академика М.И. Перельмана