Retinal Vessel-Driven Deep Learning Framework for Glaucoma Classification Using Segmentation Method
Year: 2026
Type: conference-paper
Abstract: Structural and geometric changes in the retinal vasculature are increasingly recognized as important biomarkers for detecting and monitoring glaucoma, yet most deep learning approaches still rely on full-fundus or disc-centric representations rather than explicitly modeling retinal vessels. This study proposes a vessel-based deep learning framework for glaucoma detection that operates on binary vessel maps derived from color fundus images. Green-channel fundus images from the public FIVES and RIM-ONE datasets are preprocessed using CLAHE and heuristic global thresholding to obtain retinal vessel masks, which are replicated to three channels and used as inputs to ImageNet-pretrain EfficientNet-B0, ResNet50, and MobileNetV3-Small models fine-tuned for binary (glaucoma vs. normal) classification. Performance is evaluated with stratified 5-fold cross-validation using accuracy, sensitivity, specificity, F1-score, and AUC. On FIVES, all backbones achieve accuracy 0.925, F1-score 0.9302, and AUC values up to 0.9831, while on RIM-ONE they reach accuracies between 0.8235 and 0.8676 and AUC values up to 0.9463. Grad-CAM visualizations on vessel-based inputs show that high-activation regions concentrate along major retinal vessels and altered vascular patterns in glaucomatous eyes, supporting the clinical plausibility of the learned decision behavior. These results indicate that vessel-centric fundus representations combined with modern transfer learning architectures provide a promising basis for accurate and explainable glaucoma screening.
Language: English
Topic: Retinal Imaging and Analysis
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W7204828867