Diabetic Retinopathy Detection Using Deep Learning

In this research, we present DRNet, a deep learning model for automated detection and grading of diabetic retinopathy (DR) using fundus images. DR is a leading cause of vision loss in diabetic individuals, emphasizing the importance of early detection. Our model combines the ResNet architecture for feature extraction with the U-Net architecture for lesion segmentation. Trained on a large dataset of 666,383 fundus images from 121,342 diabetic patients, DRNet incorporates realtime image quality assessment, lesion detection, and grading of DR severity. Evaluation on both local and external datasets, totaling 409,455 images, demonstrates the effectiveness of DRNet, achieving an area under the receiver operating characteristic curves (AUC) of 0.90-0.936 for lesion detection and 0.96-0.98 for DR severity grading. Our results suggest that DRNet has the potential to enhance DR screening programs by improving accuracy and efficiency in diagnosis, thereby benefiting patient outcomes and healthcare systems.