Retinal Twins: Leveraging Binocular Symmetry with Siamese Networks for Enhanced Diabetic Retinopathy Detection

Abstract

Diabetic Retinopathy (DR), a critical ocular condition affecting retinal vasculature, can lead to visual deterioration or complete blindness if not identified early. The disorder is classified into various stages, encompassing normal, abnormal, mild, moderate, severe, proliferative diabetic retinopathy (PDR), and non-proliferative diabetic retinopathy (NPDR). The prompt and accurate detection of DR continues to pose a substantial challenge in the field of ophthalmology. This study introduces an innovative hybrid binocular Siamese deep learning (DL) framework for precise DR image classification. The methodology integrates advanced noise reduction techniques, robust feature extraction utilizing hybrid metaheuristic algorithms, and blood vessel (BV) and optic disc (OD) segmentation through an open-closed watershed management approach. Subsequently, an optimized boosting classifier is applied to categorize the segmented images, ensuring high accuracy and resilience. The system's performance is evaluated using DB0 and DB1 datasets, demonstrating exceptional results across various metrics, including accuracy, precision, F1-score, and MCC. Notably, the proposed system achieves an impressive 99% accuracy, setting a new standard in DR classification. This research highlights the potential of the proposed system to transform DR diagnosis and treatment approaches.