Clinically Informed Optic Disc and Optic Cup Segmentation for Automated Glaucoma Assessment Using Adaptive Feature Fusion Multi-Task U-Net++

Background/Objectives: Glaucoma is considered one of the most serious eye diseases leading to permanent vision loss worldwide. Early diagnosis and regular follow-up are crucial for slowing disease progression and preserving visual function. Accurate segmentation of the optic disc (OD) and optic cup (OC) from fundus images enables reliable estimation of clinically important biomarkers such as the cup-to-disc ratio (CDR). However, the low contrast of OC boundaries, vascular occlusions, and variations in image acquisition conditions continue to limit the performance of automated segmentation methods. This study proposes a novel deep learning framework, Adaptive Feature Fusion Multi-Task U-Net++ (AFF-MT U-Net++), for accurate OD and OC segmentation. Methods: The proposed architecture incorporates an EfficientNet-B5-based encoder, an Adaptive Feature Fusion (AFF) module for refining feature representations, a boundary-aware learning strategy for improving anatomical boundary localization, and a clinical feature learning branch that models structural indicators related to glaucoma assessment. The proposed method was evaluated on 1200 eligible annotated images retained from the REFUGE2 dataset after data screening using a source-stratified, image-level five-fold cross-validation protocol on the pooled REFUGE2 data and compared with U-Net, U-Net++, nnU-Net, and PVTv2 models. Results: Experimental results demonstrated that the proposed method achieved a Dice score of 0.8910 ± 0.0023 and an Intersection over Union (IoU) of 0.8035 ± 0.0037 for OC segmentation, while obtaining a Dice score of 0.9602 ± 0.0016 and an IoU of 0.9235 ± 0.0029 for OD segmentation. In addition, the model achieved a vertical cup-to-disc ratio error (δvCDR) of 0.0433 ± 0.0011. Paired statistical comparisons indicated higher Dice and IoU scores and lower δvCDR errors than the evaluated baseline models, although improvements were not observed for every metric. Conclusions: AFF-MT U-Net++ provided reliable and stable OD/OC segmentation performance on the REFUGE2 dataset through adaptive multi-scale feature fusion, boundary-aware learning, and clinical feature modeling. The proposed framework produced anatomically consistent segmentation results and demonstrated its potential to support computer-aided glaucoma assessment from fundus images.