Abstract

Glaucoma is a leading cause of blindness, and early detection is critical. Convolutional neural networks (CNNs) have shown impressive performance in glaucoma diagnosis, but their black-box nature remains a barrier to clinical use. Existing explainable AI (XAI) methods such as Grad-CAM have limitations in identifying and quantifying subtle regional features. In this study, we propose a method to clarify what CNNs focus on by analyzing how model performance changes under localized adversarial noise. Using VGG16 for glaucoma classification, we applied noise generated by the Fast Gradient Sign Method (FGSM) to the whole fundus image and to specific subregions, then compared the impact on classification performance. Results showed that perturbations to the optic disc, especially its outer margin, had the greatest effect on model performance. This suggests that the CNN captures fine anatomical features such as optic disc cupping and neuroretinal rim thinning, which aligns with what ophthalmologists typically look for. At the same time, perturbations in the macula and perivascular regions also affected performance, indicating gaps between current clinical diagnostic criteria and the CNN’s decision-making process. This approach can help establish the clinical reliability of CNNs and may also reveal features that have not been recognized in conventional clinical practice.

IPC Classification

Keywords