Severity Grading and Lesion Segmentation in Diabetic Retinopathy

Diabetic retinopathy (DR) is the most frequent cause of avoidable diabetes blindness and that is why it has to be identified and treated early. Computer-aided diagnosis must be automated because screening of retinal fundus image by hand is tedious and subjective. The paper categorizes and quantifies diabetic retinopathy lesions with the help of deep learning and retinal fundus images. In order to enhance image quality and model robustness, IdriD dataset scales, normalizes and optimizes retinal images. U-Net can identify imbalance in classes, hard exudates and hemorrhages accurately. The classification models based on ResNet-50 classify DR severity as Mild, Moderate, Severe, Proliferative, or No DR. Severity-classification and segmentation were effective with trials being at 98.27% pixel accuracy. Grad-CAM visualization is further improved by removing clinically significant areas, thereby improving the interpretability of the model. The suggested procedure is effective and is explainable in automated diabetic retinopathy screening.