Advancing Glaucoma detection: A novel framework with Kolmogorov Arnold-Vision Transformer and Grad-CAM explainability technique

Abstract Glaucoma is a leading cause of preventable blindness requiring early detection to prevent irreversible vision loss. The conventional diagnosis approach involves a detailed structural and functional assessment of the visual system, with the optic cup to disc ratio (CDR) as a key biomarker. Clinical CDR estimation suffers from significant interobserver variability limiting its diagnostic consistency. In this work, we propose a computer aided automated framework for glaucoma detection that limits dependence on subjective CDR measurements. The method first extracts the optic disc (OD) Region of Interest (ROI) from retinal fundus images, analyzed using the proposed Kolmogorov Arnold-Vision Transformer (KA-ViT) framework. KA-ViT integrates a novel Group Adaptive Rational-Kolmogorov Arnold Network (GAR-KAN) in place of the MLPs present in conventional ViT for improved feature learning and classification efficiency. Results on the publicly available ACRIMA, Drishti-GS, and RIM-ONE datasets demonstrate the efficacy of the proposed KA-ViT model. The explainability of the proposed model is analysed using Grad-CAM and we introduced two novel metrics, namely Anatomical Localization Score and Spatial Consistency Score for the quantitative analysis of Grad-CAM heatmaps. The activation maps obtained consistently highlight the optic disc and cup regions, confirming the model aligns with established clinical practices used in glaucoma assessment.