Identification And Risk Level Prediction Of Diabetic Retinopathy Using Transfer Learning With Novel Vision Transformer And Grad-Cam Explainable Ai

Abstract

Diabetic retinopathy (DR) is one of the leading causes of visual impairment that requires correct prediction and early detection. The system suggested herein introduces a hybrid deep learning model that combines statistical feature analysis with a modified version of the Vision Transformer (ViT) architecture to be used in strong classification. Models that had been pre-trained to identify Diabetic Retinopathy were tested and the model with highest performance was integrated into a ViT-inspired architecture to achieve prediction. The improved version comprises of an optimized multi-head attention block and an enhanced transformer block to achieve a high-quality feature extraction and classification accuracy. used explainable AIs, such as a variant of Gradient-weighted Class Activation Mapping (Grad-CAM) tailored to Vision Transformers, to facilitate transparency and make informed decisions. Key performance measures were compared and the assessment revealed an evaluation of 83 % accuracy.