MedViT-XAI: Explainable Vision Transformer for Automated Diabetic Retinopathy Grading

Abstract

Diabetic retinopathy (DR) is among the top causes of preventable blindness in the world especially in working-age groups. To deliver clinical interventions in time and manage the disease, prevention and correct identification of the severity of the DR is a critical condition. Nevertheless, the procedure of retinal fundus image screening by manual screening and ophthalmologists is challenging, subjective, and laborious in large population-based screening programs. In order to eliminate these weaknesses, in this paper, a novel explainable Vision Transformer network named MedViT-XAI is proposed that can be used to classify retinal fundus images with diabetic retinopathy among various conditions. The new model incorporates a hierarchical patch embedding, transformer-based global feature extraction, and explainable artificial intelligence (XAI) mechanism aiming to improve the diagnostic performance and the interpretability of the model. Normalization, contrast enhancement and resizing are initially used to preprocess retinal images in order to standardize the input data and eliminate noise. The images are subsequently broken up into fixed-size patches which are fed through multi-head self attention layers that take into account spatial relationships over a large distance in retinal structures like microaneurysms, hemorrhages and exudates. Moreover, explainability tools, e.g. attention visualization and gradient based attribution, are included to indicate clinically relevant retinal areas that affect model predictions.