Automated Tool Support for Glaucoma Identification With Explainability Using Fundus Images

Glaucoma is a progressive eye condition that causes irreversible vision loss due to damage to the optic nerve. Recent developments in deep learning and the accessibility of computing resources have provided tool support for automated glaucoma diagnosis. Despite deep learning’s advances in disease diagnosis using medical images, generic convolutional neural networks are still not widely used in medical practices due to the limited trustworthiness of these models. Although deep learning-based glaucoma classification has gained popularity in recent years, only a few of them have addressed the explainability and interpretability of the models, which increases confidence in using such applications. This study presents state-of-the-art deep learning techniques to segment and classify fundus images to predict glaucoma conditions and applies visualization techniques to explain the results to ease understandability. Our predictions are based on U-Net with attention mechanisms with ResNet50 for the segmentation process and a modified Inception V3 architecture for the classification. Attention U-Net with modified ResNet50 backbone obtained 99.58% and 98.05% accuracies for optic disc segmentation and optic cup segmentation, respectively for the RIM-ONE dataset. Additionally, we generate heatmaps that highlight the regions that impacted the glaucoma diagnosis using both Gradient-weighted Class Activation Mapping (Grad-CAM) and Grad-CAM++. Our model that classifies the segmented images achieves accuracy, sensitivity, and specificity values of 98.97%, 99.42%, and 95.59%, respectively, with the RIM-ONE dataset. This model can be used as a support tool for automated glaucoma identification using fundus images.