Advancing diabetic retinopathy diagnosis with fundus imaging: A comprehensive survey of computer-aided detection, grading and classification methods

Abstract

The incidence of diabetic retinopathy globally calls for advanced and more universally applicable computer-aided diagnosis (CAD) systems. This survey paper explores the current state of vision-based CAD techniques for the detection and classification of diabetic retinopathy, a diabetes-induced eye disorder that can lead to severe visual impairment or blindness. Characterized by a variety of manifestations including microaneurysms, exudates, hemorrhages, and macular detachment, diabetic retinopathy presents substantial challenges for automated detection. This is primarily due to the heterogeneity of retinal fundus images, which display diverse spatiotextural features and intricate vascular structures. Our exhaustive review indicates that most existing methodologies predominantly concentrate on isolated diabetic retinopathy types, employing localized spatiotextural feature analysis for classification. Such specificity often results in limited accuracy and generalizability, restricting practical real-world application. Furthermore, contemporary leading methods generally focus on single retinal characteristics, necessitating patients to undergo multiple CAD procedures, thereby increasing time, costs, and possibly intensifying retinal complexities. To overcome these obstacles, we propose the adoption of multi-trait-driven CAD solutions. Utilizing the potent capabilities of deep learning, these solutions could employ high-dimensional, multi-cue sensitive feature extraction and ensemble learning for classification. This approach is designed to improve the generalizability and dependability of CAD systems, offering a holistic solution capable of effectively managing the diverse manifestations of diabetic retinopathy. Our study highlights the need for a fundamental transformation in diabetic retinopathy CAD systems, motivating further research towards robust, multi-modal methods to enhance detection, classification, and grading of this widespread ailment.