Interpretable Deep Learning for Early Detection of Diabetic Retinopathy using Lightweight Vision Transformers

An interpretable and computationally efficient deep learning framework for the early screening of Diabetic Retinopathy (DR) that exploits a Lightweight Vision Transformer (L ViT) has been proposed in this research. The model combines distillation and explanation consistency loss to balance accuracy, interpretability and efficiency. In addition, a teacher–student network framework is employed to distil semantic and spatial knowledge from a huge transformer to the small student model, which is beneficial for real time and edge applications. In addition, an explanation head incorporates attention rollout with gradient based saliency maps to enhance clinically significant retinal lesions (microaneurysms, hemorrhages, and exudates). Experiments on EyePACS, Messidor 2 and IDRiD datasets show that the proposed L ViT outperforms CNN and the existing ViT baselines with 95.6% accuracy and 0.982 AUC at a reduced computational cost. An interpretable and reliable diagnostic model, which is a stepping stone toward affordable AI based retinal screening in resource constrained environments has been developed.