Diabetic Retinopathy Detection Using Deep Learning on Fundus Images

Diabetic Retinopathy (DR) is primary cause of blindness in diabetic patients if left untreated, it happens with blockage of blood vessels due to high blood sugar levels. The need is to develop hybrid deep learning framework to early detect and classify fundus image into five class. The aim of this research is to evaluate high accuracy using deep learning on multiple datasets. This research has introduced EVH1 hybrid model of EfficientNetB0 with Vision Transformer (ViT) with CLAHE and Gaussian blur to enhance Fundus Images (FI) contrast on DRH1 combination of six publicly available F1 datasets. Literature was reviewed from Google Schooler, Science Direct and IEEE Xplore. The EVH1 achieve an overall training accuracy of 99.83%, precision 99.86%, recall 99.86%, AUC 100% and testing accuracy of 96% that illustrates model’s ability to demonstrate better performance on unseen dataset. This study also includes Grad-CAM and SHAP for interpretability for clinical settings. The EVH1 stoned new markup in comparison with previous DR prediction models.