Retinal Image Analysis and Diagnosis of Retinal Blood Vascular Diseases Using Deep Learning Model

Retinal Image Analysis and Diagnosis of Retinal Blood Vascular Diseases using Deep Learning Model Bismita Choudhury (Student ID: 100063702) A thesis submitted in fulfilment of the requirements for the degree of Doctor of Philosophy (PhD) (Computer Science and Engineering) Faculty of Engineering, Computing, and Science Swinburne University of Technology Sarawak Campus May/2019 “Research is what I’m doing when I don’t know what I’m doing.” – Wernher von Braun THESIS – DOCTOR OF PHILOSOPHY (PHD) Abstract The retina is not only an important part of the visual system, but it also has the potential to indicate the general health of the other parts of the human body. In addition to the eye disease, various retinal abnormalities can be indicative of other health issues. Recent studies have show n that these retinal abnormalities associated with the blood vascular disease are predictive to several major diseases, viz., Diabetes, Cardiovascular diseases like Hypertension and Coronary heart disease, Kidney disease, and Stroke. Among the various blood vascular diseases, Diabetic Retinopathy (DR) and Retinal Vein Occlusion (RVO) are the two leading causes of blindness worldwide. The main causes for both of these sight-threatening retinal diseases are the age, obesity and sedentary lifestyle of people. As these factors are beyond controllable to avoid such diseases, therefore, it is particularly important to detect these retinal abnormalities as early as possible and prevent the visual imparity. The recent years have seen the increased interest in diagnosing various diseases through Computer Aided Diagnosis (CAD) of the digital images. In past decades, several such methods have been proposed for diagnosing DR. Majority of the methods face the challenge in detecting DR at the earliest stage. Again, there is a substantial lack of research in automatic detection of RVO given the fact that it is the second most popular reason of vision loss and an indication to the possible blockage in cardiac veins and nerves in the human brain as well. The literatures on CAD for diagnosing retinal abnormality are disease specific. The algorithms for detecting one type of disease either cannot or fail to detect other types of disease due to similar intra and interclass variability of the features. And, the majority of these methods are dependent on hand designed feature engineering. Therefore, the performance of these methods for classifying diseases depends on the performance of the segmentation and feature extraction methods. To overcome these problems, in this dissertation, deep learning methods have been proposed to analyse retinal image and diagnose retinal abnormalities to detect blood vascular diseases, particularly DR and RVO. The proposed Deep Learning methods utilize the advantages of Convolutional Neural Swinburne University of Technology Sarawak Campus | Abstract i THESIS – DOCTOR OF PHILOSOPHY (PHD) Network (CNN) to analyse the retinal image and detect possible retinal blood vascular diseases associated with the abnormal appearance of the retina. A novel architecture of CNN has been proposed to particularly work with retina image and detect DR and RVO. A design hypothesis has been set to design CNN from scratch for the particular problem at hand. The main focus has been on overcoming the barrier of using deep models since the popular, widely used CNN models introduce extra complexity. Generally, the deep learning models are complicated and require a huge amount of training samples, memory, and time for providing supervised learning. In this research, a simple CNN architecture has been carefully designed to extract and analyse the normal and abnormal features from the raw image pixels of the whole color fundus image, and detect possible retinal blood vascular disease as early as possible. The careful design, careful selection of hyperparameters, and learning algorithm have made the proposed model a simple yet a potential model to diagnose retinal abnormalities without compromising the performance. This CNN can detect DR at the earliest stage and grade DR into the mild to moderate Non-proliferative DR (NPDR) and severe NPDR to Proliferative DR (PDR). The experiments have been conducted for two databases, viz., Messidor and Kaggle database. The proposed model can detect early stage DR or mild-NPDR with an accuracy of 98.11%, sensitivity of 100%, and specificity of 96.2% for Messidor database. For Kaggle database, the proposed model has achieved 96.6% accuracy, 99.7% sensitivity, and 93.2% specificity. For DR severity grading, the experiments have been conducted on STARE and Messidor Database. The proposed model can classify DR into three classes; viz., normal, mild to moderate NPDR, and severe NPDR to PDR with an accuracy of 98.2%, sensitivity of 100%, and specificity of 98%. Again, the proposed model can detect RVO and its two types, viz., Central Retinal Vein Occlusion (CRVO) and Branch Retinal Vein Occlusion (BRVO). The RVO images have been collected from two databases, viz. STARE and Retinal Image Bank and as per the conducted experiments, the proposed model can classify normal, BRVO and CRVO images with accuracy 97%, sensitivity 96.1%, and specificity 98%. In addition to that, this model is also capable of analysing similar features of different diseases; hence, it can distinguish both DR and RVO irrespective of their common lesions. For this particular experiment, the images have been mixed and matched from multiple databases. The DR images (both NPDR and PDR) are Swinburne University of Technology Sarawak Campus | Abstract ii THESIS – DOCTOR OF PHILOSOPHY (PHD) collected from STARE, DRIVE and Messidor database, the RVO images (both BRVO and CRVO) from STARE and Retinal Image Bank, and Normal images from STARE, Messidor, DRIVE, and Dr. Hossain Rabbani). For the 3-class classification of DR, RVO and normal image, the proposed model has obtained 98.8% accuracy, 100% sensitivity, and 98.3% specificity. This dissertation has made an attempt to fulfil the gap in the research for automatic diagnosis of Retinal Vein Occlusion and detect all its variants, viz., Central Retinal Vein Occlusion (CRVO), Branch Retinal Vein Occlusion (BRVO), and Hemiretinal Vein Occlusion (HRVO). A Cascaded Convolutional Neural Network (CCNN) has been proposed, which is a chain of three CNNs of same proposed configuration. This special novel deep architecture carefully analyses the ambiguous features of HRVO. In this chain of three CNNs, every two CNNs carefully investigate the features of each type of RVO. The Cascaded Convolutional Neural Network (CCNN) takes a final decision on RVO types based on the result of two internal CNNs. There is no such algorithm proposed till date to detect all three types of RVO. Therefore, this proposed method is one of its first kinds to detect all three types of RVO. For the experiment, the RVO images and normal images have been collected from multiple databases, viz., DRIVE, Messidor, STARE, and Dr. Hossain Rabbani. The proposed Cascaded Convolutional Neural Network can successfully classify BRVO, CRVO, HRVO, and normal images with an accuracy of 96%, sensitivity of 97%, and specificity of 95.2%. The proposed deep learning methodologies have efficiently overcome the multiple levels of challenges in the field of retinal image analysis; computer-aided automated diagnosis methods, and deep learning. Various experiments have been carried out on multiple publicly available databases, and both of the proposed deep learning based methods have performed outstandingly in each individual task. Therefore, the proposed model is a potential model for diagnosing retinal blood vascular diseases, and can help the ophthalmologists to detect especially DR and RVO at the early stage, thus, can prevent the total vision loss of the patients. It would be a cost effective method and patients at distance areas would be able to diagnose remotely. Swinburne University of Technology Sarawak Campus | Abstract iii THESIS – DOCTOR OF PHILOSOPHY (PHD) Acknowledgement First and foremost, I would like to express my sincere gratitude to my principal-supervisor, Prof. Patrick HH Then for his continuous support, motivation, enthusiasm, guidance and sharing his invaluable and vast knowledge in the field of research. His profound insights and attention to details have been a true inspiration to my research. I could not have imagined of having better advisor and mentor for my PhD research. I would also like to thank my co-supervisor Dr. Valliappan Raman for his technical knowledge sharing and constructive criticism during the entire span of my PhD research. His insightful comments have helped me a lot in shaping this dissertation. My sincere thanks and gratitude goes to my previous co-supervisor Dr. Manas Kumar Haldar and my external supervisor Dr. Biju Issac of Northumbria University, UK, and for their continuous encouragement and valuable advices to widen my research ideas from various perspectives. I would also like to take this opportunity to thank all my friends and colleagues Brian Loh, Wan Tze Vong, Mohmd. Yuzrie, Michelle Gian, Clement Ting, Emily Rogos and Akshay Kakar for their continuous help, encouragement and unforgettable moments that we had in the last four years. Last but not the least; I would like to thank my parents and elder brother for all their love, blessings, encouragement and support. Especially I would like to mention my Father who trusted me and supported me throughout all of my pursuits. Bismita Choudhury Swinburne University of Technology,

Load more