Deep learning-based binocular system for automated diabetic retinopathy grading with prior clinical knowledge integration

Diabetic retinopathy (DR), a complication of diabetes mellitus, poses a significant risk to vision by damaging retinal blood vessels over time. Traditional methods for identifying DR rely on skilled ophthalmologists to analyze digital color retinal fundus images, leading to time-consuming and labor-intensive diagnoses. To overcome these challenges, our paper presents a deep learning-based approach inspired by real-world diagnostic practices to automatically classify DR into five stages based on color retinal fundus images. Our proposed Deep DR (2DR) system leverages the correlation in DR severity between a patient’s two eyes and integrates prior clinical knowledge of disease manifestations, such as microangiomas around the macular area, fibrovascular proliferation, and neovascularization around the optic disk. The 2DR system comprises a binocular network that evaluates disease severity in one eye while incorporating valuable information from the opposite eye to address binocular differences in grading. Furthermore, informed by prior knowledge of DR manifestation, the system extracts features from the optic disk and macular regions, along with global features from the patient’s fundus images, to enhance decision-making. Additionally, we introduce a difference-weighted mean squared error with cross-entropy as a loss function to penalize discrepancies between predicted and actual classes. Comprehensive evaluations on both a private-DR dataset and the public Kaggle-DR dataset validate the effectiveness of our approach, achieving higher accuracies of 87.3% and 86.5%, respectively, compared to state-of-the-art methods.