A MULTI-TASK EXPLAINABLE AI FRAMEWORK FOR PAPILLEDEMA DETECTION AND SYSTEMIC DISEASE RISK PREDICTION USING RETINAL FUNDUS IMAGES

Abstract

Papilledema is optic-disc swelling associated with raised intracranial pressure and requires timely recognition. This work develops a multi-task explainable artificial intelligence framework that classifies retinal fundus images while extracting vessel and optic-disc information for systemic-risk analysis. The dataset preparation pipeline converts the original Normal, Pseudopapilledema, and Papilledema organization into a binary Normal, Papilledema task and applies an 80:20 training validation split. Six architectures—ResNet50, DenseNet201, ViT-base, ViT-CapsuleNetwork, Xception, and ConvNeXt Tiny—are evaluated using accuracy, precision, recall, F1-score, AUC, confusion matrices, and ROC curves. Xception provides the strongest verified result with 0.996 accuracy, precision, recall, and F1-score and 1.000 AUC. The structural analysis branch enhances the green channel, generates vessel masks and skeletons, estimates vessel diameter, detects the optic-disc edema region, derives artery and vein diameters and the artery-to-vein ratio, and demonstrates EfficientUNet optic-disc segmentation. Grad-CAM explains the classifier, while a Flask and SQLite interface supports authenticated image upload and integrated result presentation. The system is intended as a research decision-support prototype rather than an autonomous clinical diagnosis tool.