Download files
Abstract
Diabetic retinopathy (DR) is a leading cause of preventable blindness among individuals of working-age in developed nations. While early detection and treatment of DR is effective in preventing visual loss, the provision of cost-effective regular eye tests for DR has become a critical challenge. To meet it, regular screening programs are available. However, they need to be simultaneously practical and reliable involving suitable personnel while also reaching an appropriate level of sensitivity.
DR screening has been performed using various methods. This research focuses on the new ultra-wide field (UWF) retinal images. UWF images allow central pole-to-periphery views of the retina, compared to conventional retinal cameras which show only the central retina. Sensitivity of screening may be improved with evaluation of the peripheral retina. As far as is known no studies on automatic detection of DR in UWF retinal images have been published. In this research, a unique dataset of UWF retinal images was collected, and customized software developed to enable dataset annotation.
This thesis proposes an automated solution for DR lesion detection in both central and peripheral retina using UWF retinal images. A number of approaches were devised and tested for the purpose. A hand-crafted feature set was designed to perform pixel-level classification of bright and dark DR lesions. The applicability of machine learning algorithms on a unique dataset such as UWF retinal image data was established. A convolutional neural network for detection of bright and two subtypes of dark DR lesions was then designed, trained and validated. The utility of convolutional neural networks for classification of UWF retinal image data based on limited training data was established empirically. Finally, a transfer learning approach for detection of four subtypes of DR lesions in UWF retinal images was built, which is novel in the field of retinal image analysis.
A new automated system for DR lesion detection in the peripheral retina with overall AUC of 80% has been developed and presented for the first time. The proposed system can be used to facilitate DR diagnosis at early stages and assist in grading of DR.