Deep Learning for Diabetic Retinopathy Early Detection and Severity Grading

Early detection and severity grading of Diabetic Retinopathy (DR) are crucial for preventing vision loss. This study investigates the efficacy of deep learning models in achieving this goal using a large, pre-processed dataset of Gaussian-filtered retinal scans. The dataset, combining images from APTOS 2019 Blindness Detection and DRD, underwent merging, augmentation, resizing, cleaning, balancing, and Gaussian filtering. We evaluated various models, including ConvXLarge (89%/83.32% train/test accuracy), ResNetRS101 (98.9%/93% train/test accuracy), ResNetRS270 (97.76%/89.14% train/test accuracy), and ResNetRS420 (98.24%/87.59% train/test accuracy). This study demonstrates the potential of deep learning, particularly ResNetRS models, using pre-processed retinal scans for accurate early detection and severity grading of DR, potentially contributing to improved patient outcomes.