Welikala, Roshan Alex (2014) Automated detection of proliferative diabetic retinopathy from retinal images. (PhD thesis), Kingston University, .
Abstract
Diabetic retinopathy (DR) is a retinal vascular disease associated with diabetes and it is one of the most common causes of blindness worldwide. Diabetic patients regularly attend retinal screening in which digital retinal images are captured. These images undergo thorough analysis by trained individuals, which can be a very time consuming and costly task due to the large diabetic population. Therefore, this is a field that would greatly benefit from the introduction of automated detection systems. This project aims to automatically detect proliferative diabetic retinopathy (PDR), which is the most advanced stage of the disease and poses a high risk of severe visual impairment. The hallmark of PDR is neovascularisation, the growth of abnormal new vessels. Their tortuous, convoluted and obscure appearance can make them difficult to detect. In this thesis, we present a methodology based on the novel approach of creating two different segmented vessel maps. Segmentation methods include a standard line operator approach and a novel modified line operator approach. The former targets the accurate segmentation of new vessels and the latter targets the reduction of false responses to non-vessel edges. Both generated binary vessel maps hold vital information which is processed separately using a dual classification framework. Features are measured from each binary vessel map to produce two separate feature sets. Independent classification is performed for each feature set using a support vector machine (SVM) classifier. The system then combines these individual classification outcomes to produce a final decision. The proposed methodology, using a dataset of 60 images, achieves a sensitivity of 100.00% and a specificity of 92.50% on a per image basis and a sensitivity of 87.93% and a specificity of 94.40% on a per patch basis. The thesis also presents an investigation into the search for the most suitable features for the classification of PDR. This entails the expansion of the feature vector, followed by feature selection using a genetic algorithm based approach. This provides an improvement in results, which now stand at a sensitivity and specificity 3 of 100.00% and 97.50% respectively on a per image basis and 91.38% and 96.00% respectively on a per patch basis. A final extension to the project sees the framework of dual classification further explored, by comparing the results of dual SVM classification with dual ensemble classification. The results of the dual ensemble approach are deemed inferior, achieving a sensitivity and specificity of 100.00% and 95.00% respectively on a per image basis and 81.03% and 95.20% respectively on a per patch basis.
Actions (Repository Editors)
 |
Item Control Page |