An interpretable diagnosis of retinal diseases using vision transformer and Grad-CAM

Abstract

Timely identification of retinal disorders is essential to mitigate the risk of vision impairment or total blindness. This study introduces an innovative and interpretable diagnostic framework that integrates the capabilities of a CNN-based architecture and a Transformer model, followed by visualization using Grad-CAM, to tackle the complexities of multi-label classification in retinal disease analysis. By utilizing OCT imagery, we have developed a specialized hybrid system that synergizes deep learning techniques from convolutional and transformer-based networks to enhance diagnostic accuracy. Beyond classification, this approach provides translucent insights into the decision-making process of the model through activation-based visual explanations, improving trust and interpretability in medical AI applications. The CNN-based model achieved an accuracy rate of 88.88%, while the Transformer-based approach reached 91.39%. However, our optimized hybrid configuration significantly outperformed these standalone models. With an impressive accuracy of 98.8%, this advanced system demonstrates its strong potential in revolutionizing retinal disease detection and assisting healthcare professionals in early intervention.