A Computational Framework for Diabetic Retinopathy Severity Grading Categorization using Ophthalmic Image Processing

Abstract

Diabetic Retinopathy (DR) is a retinal vascular disease characterized by prolonged diabetic complication leading to severe blindness. The eye related impediment arises due to progressive deterioration of retinal blood vessels and can be distinguished by the appearance of different types of clinical lesions.Regular screening and diagnosis can reduce the chances of vision loss up to a large extent. Early stage prognosis of DR requires regular eye examination and ophthalmologists rely on retinal fundus segmentations for the treatment of DR abnormalities.Automated detection, segmentation and classification approaches have become an eminent research area for effective DR diagnosis and treatment of severe eye diseases preventing visual impairment. Accurate segmentation of retinal vasculatures distinguishing between the anatomy and pathology of retinal fundus images is significant for precise prediction of the disease. Detection and analysis of different DR lesion as well as disease severity grades help the ophthalmic experts in analyzing the variations in the fundus images and taking necessary action before disease progression. Diagnostic relevance of DR prediction and grading to aid the ophthalmologists in regular screening has led to the expansion of automated DR severity systems. In this thesis, DR diagnosis is addressed by proposing a retinal anatomical structure segmentation approach to reduce fallacious lesion segmentation.Physiology identification and detection are accomplished by developing an effective lesion discrimination approach to provide a robust DR lesion detection solution and optimal categorization capability. A robust framework is proposed for DR categorization that offers a generalized approach for DR severity grading.A computation transfer learning framework has been proposed in this thesis, to provide a Deep Learning based mass screening solution for DR classification problem. All the proposed approaches were tested on designated benchmark datasets defined in the literature for DR diagnostic task.