A Benchmark for Studying Diabetic Retinopathy: Segmentation, Grading, and Transferability

People with diabetes are at risk of developing an eye disease called diabetic retinopathy (DR). This occurs when high blood glucose levels cause damage to vessels in the retina. Computer-aided DR diagnosis is a promising tool for early detection and severity grading, due great success deep learning. However, most current systems do not achieve satisfactory performance or interpretability ophthalmologists, lack training data consistent fine-grained annotations. To address this problem, we construct large annotated dataset containing 2,842 images (FGADR). has 1,842 pixel-level DR-related lesion annotations, 1,000 image-level labels graded by six board-certified ophthalmologists intra-rater consistency. The proposed will enable extensive studies on diagnosis. We set up three benchmark tasks evaluation: 1. segmentation; 2. grading joint classification 3. Transfer learning ocular multi-disease identification. Moreover, novel inductive transfer method introduced third task. Extensive experiments using different state-of-the-art methods conducted our FGADR dataset, which can serve as baselines future research.