GAB-Seg: a task-specific boundary-aware framework for irregular multi-scale spot-like target segmentation in anterior segment images
The automated segmentation of localized spot-like anomalies on the bulbar conjunctiva presents a specific image processing challenge. While anterior segment images generally maintain high overall clarity without severe lighting interference, the visual data in this domain is strictly characterized by scattered, irregularly shaped targets with extreme scale variations and ambiguous boundary transitions. Compounded by data scarcity and severe foreground-background class imbalance, we propose GAB-Seg, a task-tailored generative boundary-aware segmentation framework exclusively designed for these morphological attributes. To overcome limited training samples, conditional generative adversarial networks provide data augmentation. Crucially, a Cross-Domain Gradient Orthogonal Decoupling (CGOD) module resolves optimization conflicts, preventing synthesized images from disrupting the learning of real edge gradients. To tackle boundary ambiguity, a Multi-Contour Predictive Map Auxiliary (MC-PMA) module constructs structural priors via polygon approximation, guiding the network to capture core target trajectories. Finally, a newly designed Multi-scale Smooth Distance-aware Synergistic Loss (MSDS-Loss) mitigates missed detections intrinsic to long-tailed distributions while suppressing local high-frequency noise. Experimental results on our internal test set suggest that GAB-Seg provides an effective, task-tailored engineering solution for this specific segmentation scenario.
Authors
- Houpan Zhou
- Le Luo
- Zhaocheng Fu
Institutions
- Hangzhou Dianzi University (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-13
- DOI
- https://doi.org/10.1038/s41598-026-79971-4
- Primary Topic
- Medical Image Segmentation Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00