Interpretable Glaucoma Classification and Severity Assessment Using Deep Learning Models

Abstract

Globally, glaucoma is a major cause of permanent blindness and early and accurate diagnosis is essential to saving sight and enhancing patient outcomes. Retinal fundus imaging can be used for non-invasive glaucoma screening. However, regular manual evaluation is time consuming and requires a lot of clinical expertise and is prone to inter-observer variability. The categorisation of glaucoma was done using a publicly available dataset of retinal fundus images from multiple ophthalmic sources. The photos were acquired during the past few years containing normal and glaucomatous retinal samples. The designed preparation pipeline aims to increase the resilience and generalisation of the models by image scaling, normalisation, contrast enhancement, data augmentation, and dataset balancing . Various DL models such as ResNet50, DenseNet201, ViT, ViT-CapsNet, Xception and ConvNeXtTiny were compared to classify glaucoma. The task of segmentation of optic disc and cup was done by EfficientUNet to estimate CDR and the task of glaucoma localisation was done by the YOLO based models. Visual Explanation for Interpretability Improvement: Grad-CAM and Grad-CAM++. The accuracy, precision, recall, F1- score, Dice Score, IoU and mAP served as performance metrics. DenseNet201 attained the highest classification accuracy of 89.60%, while EfficientUNet performed best at 95/82.40% Dice Scores for optic disc and optic cup segmentation, respectively, and YOLOv5s6 had [email protected] of 97.50%. To improve computer-aided ocular screening, the integrated framework offers the following: Accurate diagnosis, interpretable prediction, anatomical analysis and clinically useful glaucoma assessment.Index Terms: Glaucoma classification, Vision Transformer, Capsule Network, retinal fundus images, explainable artificial intelligence, deep learning, medical image analysis.