Computational Methods and Deep Learning for Ophthalmology

Computational Methods and Deep Learning for Ophthalmology presents readers with the concepts and methods needed to design and use advanced computer-aided diagnosis systems for ophthalmologic abnormalities in the human eye. Chapters cover computational approaches for diagnosis and assessment of a variety of ophthalmologic abnormalities. Computational approaches include topics such as Deep Convolutional Neural Networks, Generative Adversarial Networks, Auto Encoders, Recurrent Neural Networks, and modified/hybrid Artificial Neural Networks. Ophthalmological abnormalities covered include Glaucoma, Diabetic Retinopathy, Macular Degeneration, Retinal Vein Occlusions, eye lesions, cataracts, and optical nerve disorders.

This handbook provides biomedical engineers, computer scientists, and multidisciplinary researchers with a significant resource for addressing the increase in the prevalence of diseases such as Diabetic Retinopathy, Glaucoma, and Macular Degeneration.

• Presents the latest computational methods for designing and using Decision-Support Systems for ophthalmologic disorders in the human eye
• Conveys the role of a variety of computational methods and algorithms for efficient and effective diagnosis of ophthalmologic disorders, including Diabetic Retinopathy, Glaucoma, Macular Degeneration, Retinal Vein Occlusions, eye lesions, cataracts, and optical nerve disorders
• Explains how to develop and apply a variety of computational diagnosis systems and technologies, including medical image processing algorithms, bioinspired optimization, Deep Learning, computational intelligence systems, fuzzy-based segmentation methods, transfer learning approaches, and hybrid Artificial Neural Networks

• Language: English
• Publisher: Academic Press
• Release date: Feb 18, 2023
• ISBN: 9780323954143

Book preview

1: Classification of ocular diseases using transfer learning approaches and glaucoma severity grading

D. Selvathi     Senior Professor and Head, Biomedical Engineering Programme, Department of ECE, Mepco Schlenk Engineering College, Sivakasi, Tamilnadu, India

Abstract

The recent advancement in medical imaging technologies resulted in an abundant availability of retinal images to analyze ocular pathologies. The ever-growing medical database in the form of images and the complexity of human analysis of these medical images made a way for the intrusion of Computer Aided Diagnosis in the medical field. Ocular diseases are analyzed using retinal fundus images by extracting optimal features for the particular disease. Successful identification and grading of ocular diseases demand rich human expertise. In addition, the process is time consuming and hard for nonexperts to identify the relevant features within the image. The ophthalmologist find the region of interest for the pathology analysis and grading, which is now eased by the automation of medical image analysis using machine learning techniques. This prompted the automation of ocular disease examination using image processing. Automatic image processing paved way for the experts as well as nonexperts to do the task. In this proposed method, deep convolutional neural network (DCNN) models classify the ocular pathologies into eight classes such as age-related macular degeneration, cataract, glaucoma, diabetic retinopathy, hypertensive retinopathy, myopia, normal, and others. The input sources are trained by DCNN and then validation is done by the trained DCNN for evaluating the accuracy. Among the DCNN networks ResNet50, AlexNet, VGGNet, and GoogLeNet, ResNet50 yielded a promising result on the classification of ocular diseases with a classification accuracy of 89.9%. The classification process is tailed by the grading of glaucoma realized by segmenting the region of interest that is, cup and disk. Based on the cup-to-disc ratio, the severity of glaucoma is graded. Severity grading is done to four labels, that is, no glaucoma, mild glaucoma, moderate glaucoma, and severe glaucoma.

Keywords

CNN; Glaucoma grading; Ocular diseases; Transfer learning

1.1. Introduction

The primary requisite of medical diagnosis is accuracy, which has a significant effect on mitigation and treatment. Medical diagnosis is the process of identification of the presence and severity of the pathology based on the symptoms and signs. Medical diagnosis is further assisted by medical imaging techniques that are drastically improved in the past few decades. These medical imaging techniques are predominantly used for creating the visual depiction of the organs or tissues that are hidden from the human eyes by the skin. Several imaging modalities are present in the medical field. These techniques portray the structure of the retinal fundus. Vision is the prominent sense for all living beings. Therefore, eye is considered as the most precious of all sensory organs because 80% of human perception is based on what is seen. So, visual health is given much importance nowadays. Retinal examination is often done regularly. This in turn adds a responsibility of proper monitoring, analysis, and diagnosis of ocular fundus images. If any careless is done during diagnosis, it may result in delayed treatment and sometimes even loss of vision.

Glaucoma is one such dread ocular disease that mostly affects adults above 40 having a serious risk of irreversible blindness. It is the second most ocular disease to cause blindness to a huge number. WHO surveys say that it affects nearly 12 million people annually and causes 1.2 million vision losses. Glaucoma analysis demands a rich human expertise. But human analysis has an issue of intra- and interobserver variability. Thus, automatic ocular disease identification and grading are emerging nowadays. But ophthalmic imaging is a challenging job. There are several imaging technologies for capturing eye images. Among that, color fundus imaging using fundus cameras is used for imaging the rear portion of the eye. So medical diagnoses using color fundus imaging modalities are having more advantages and less risk comparing other modalities. Color fundus photography is the best method for visualizing and analyzing ocular pathologies compared to other imaging techniques.

Ocular diseases are analyzed by fundus images (rear portion of the eye) rather than lens images because the retina, optic disc, fovea, posterior pole, and macula can be visualized better in the fundus images. Medical structures like hemorrhages, optical nerve heads, blood vessels, and their abnormalities can be conveniently extracted from the fundus image. Thus, the presence and progress of most deadly ocular diseases can be analyzed from the fundus image. But human diagnosis will take considerable time and it demands expert knowledge. Therefore, automation of image analysis is needed. The automation of medical image analysis has been a topic of interest over the past 2decades, which induced many researchers to work in this field.

1.2. Literature review

Over the past 3decades, the automation of medical image analysis paved the attention of everyone, which led to several useful improvements in medical image processing. Several works are done by researchers in this domain. Ophthalmic imaging is a subdomain in medical imaging which is automated by the CAD process. The following are such works that automated the ocular disease diagnosis and severity grading.

The work focuses on classifying the data into two classes—diabetic macular edema and age-related macular degeneration—and also improves the network's adaptability to datasets. Duke University and Noor Eye Hospital in Tehran SD-OCT imaging are used. The reusability of the networks is improved by using transfer learning based on CliqueNet, DPN92, DenseNet121, ResNet50, and ResNext101. CliqueNet's precision, recall, and accuracy scores are comparatively higher than other networks. It represents that CliqueNet has the highest adaptability to datasets [1].

The work proposed in Ref. [2] is an automatic diabetic retinopathy classification and grading system referred as Deep DR. The dataset used in this work is retinal fundus images from the Sichuan Academy of Medical Sciences. The first stage was a transfer learning network ResNet50 followed by the classifier Standard Deep Neural Network, which is a custom classifier. Then, the grading system works as a four class classifier (normal, NPDR, NPDR2PDR, and PDR), which utilizes ensemble learning. This strategy consistency and reproducibility for several diagnostics yielded promising results.

In [3], three Deep Neural Networks such as AlexNet, GoogLeNet, and ResNet50 performance are analyzed for image classification. Multiple datasets such as ImageNet, CIFAR10, CIFAR100, and MNIST are evaluated to prove the performance capability of the model. It is observed that increasing the training data increases the performance accuracy but also increases the complexity of the network.

In this work, the diabetic retinopathy fundus image classification using convolutional neural networks (CNNs)-based transfer learning is implemented. The publicly available retinal fundus images in DR1 and MESSIDOR datasets are used. The pretrained networks such as AlexNet, GoogLeNet, and VGGNet are evaluated for grading diabetic retinopathy (DR) in which VGGNet achieved a better performance in five classes such as DR, mild DR, moderate DR, severe DR, and proliferate DR). Two types of fine-tunings, layer-wise fine-tuning and all-layer tuning, are tested. The pretrained CNN's layer-wise will reduce the risk of overfitting and also obtain better results for small datasets [4].

This study explores the classification process using three CNNs, CifarNet, AlexNet, and GoogLeNet. The CNNs are examined for two different diseases Thoraco-abdominal lymph node and interstitial lung disease classification. This study shows that limited datasets can cause bottleneck problems. Therefore, large-scale annotated datasets can be beneficially classified using transfer learning models [5].

This work took a deep inspection on different requisites of deep learning in medical imaging. It suggested that deep convolutional neural networks (DCNNs) can auto-extract the mid and high-level features from the images. Increasing the number of iterations or epochs optimized the network parameters. When a medium-sized dataset is not available, then pretrained CNN is suggested and also the fine-tuning of the pretrained CNNs achieved better results. The challenges in medical imaging are need of a huge dataset, need of expensive medical expertise for high-quality annotation, and privacy issues in sharing the medical dataset [6].

This paper explained the processing operations to perform disease recognition using different approaches such as support vector machine (SVM), discrete cosine transform (DCT), hidden Markov model (HMM), and principal component analysis (PCA). The first step is the image acquisition followed by segmentation where the boundary of the iris is taken as circles, and they need not be cocentric. Then normalization of image is done to eliminate nonuniform illumination. Circular symmetric filter and grabber filter are used to extract features. Finally, a matching process is done by using encoding followed by hamming distance method [7].

This paper suggested a method that has a flow of region of interest (ROI) segmentation, image scaling, disc diameter calculation, cup diameter calculation, and cup-to-disc ratio (CDR) calculation from spectral domain OCT images of the Armed Forces Institute of Ophthalmology. This system uses the green channel of the preprocessed image for feature extraction. ROI extraction has two steps first extracting a circle with the center of the retina, and the next bilinear interpolation is done to improve spatial resolution and also for improving accuracy. This paper also indicates that CDR <0.5 is normal where CDR greater than or equal to 0.5 is glaucomatous and CDR increases to indicate different stages of glaucoma [8].

This work proposes a technique for glaucoma severity classification. The publicly available Online Retinal fundus Image database for Glaucoma Analysis (ORIGA) retinal fundus image dataset is used. The three steps include ROI extraction, deep feature extraction, and classification. A deformable shape model is used to localize the disc boundary, and then a rectangle with bounding box of the disc can be obtained. For deep feature extraction, AlexNet, VGG-19, and VGG-16 are used. Then, for the classification process, linear SVM is used. AlexNet gave the best performance among deep feature extractors [9].

The proposed method for an automatic glaucoma screening system is tested on retinal image (RIM) One retinal fundus image database. In preprocessing step, color channel selection and image filtering followed by illumination correction are performed. Then, there is a segmentation module that clusters the uniform super-pixels. Statistical analysis is used for feature extraction. Super-pixel classification module (cup and disc) is performed by using SVM. Then, CDR is calculated to grade glaucoma as a normal and glaucomatous image [10].

In this survey, tests are done on various publicly available retinal fundus image datasets. The segmentation process was done by two methods, optical cup and disc extraction together and optical cup and disc extraction separately. The green channel is selected for improving the segmentation speed [11]. This study shows that glaucoma assessment can be done using abnormal vision field, optic nerve damage (CDR), and intraocular pressure. Out of which CDR is the best for glaucoma assessment. This study uses the RIM-One database. CDR can be calculated using either vertical or horizontal length of both cup and disc. But vertical CDR is normally used for glaucoma assessment [12].

This work describes that glaucoma changes the optical nerve head structures such as optic disc diameter, optic cup diameter, and mean cup depth. The retinal fundus image dataset for this work is obtained from Erlangen Glaucoma Registry (EGR). The model flows as preprocessing where illumination correction is done followed by PCA for feature extraction. In the final stage, a probabilistic two-class classifier is used for glaucoma prediction [13].

This study developed an automated glaucoma risk evaluation model. EGR retinal fundus images database is used in this work. The PCA method is used to compress the extracted high-dimensional feature vectors before SVM-based classification is done [14]. In Ref. [15], Ocular Disease Intelligent Recognition (ODIR) with eight categories of diseases are classified using transfer learning approaches.

In the work [16], glaucoma grading is done using artificial intelligence in raw OCT images. In Ref. [17], deep learning-based depthwise separable convolution model is proposed to classify glaucoma from healthy images using OCT images.

In reference to these works, an automatic ocular disease classification system, which also grades glaucoma severity, is developed by implementing transfer learning using deep neural networks.

1.3. Proposed methodology

The proposed system classifies the fundus images of an eye based on eye diseases, and grading of glaucoma based on the severity of disease is shown in Fig. 1.1.

1.3.1. Dataset details

Grand Challenge Database consists of 7000 fundus images from 3500 patients (without repetition) in which 328 images correspond to age-related macular degeneration (AMD), 424 images for cataract, 430 images for glaucoma, 206 images for hypertensive retinopathy, 2256 images for diabetic retinopathy, 348 images for myopia, 2280 images for normal, and others. Table 1.1 shows the distribution of the dataset.