GeoCorNet: Geometry-Aware Boundary Supervision for Accurate Corneal Segmentation in Anterior Segment OCT

Background and Objectives: Corneal opacity is a major cause of preventable blindness, and accurate delineation of the corneal contour in anterior segment optical coherence tomography (AS-OCT) is essential for opacity grading and disease staging. The cornea's thin, continuously curved, low-contrast geometry exposes a limitation shared by most segmentation pipelines: region-overlap objectives such as Dice are dominated by the correctly classified bulk interior and under-penalize boundary-level error, regardless of encoder capacity. A systematic comparison across CNN, hybrid, and transformer encoder families confirms this directly: boundary error diverges sharply between architectures while region-overlap accuracy remains comparatively stable, indicating that the supervision objective, not feature extraction, is the binding constraint for thin anatomical structures. This motivates supervision that explicitly encodes boundary location and continuous distance-to-boundary as geometric constraints on the learned solution, rather than as auxiliary signals subordinate to region overlap.Methods: We propose GeoCorNet, a geometry-aware, boundary-supervised segmentation network built around an EfficientNet-B4 encoder and a dual-pathway decoder, in which a dedicated boundary decoding pathway is fused into the segmentation pathway via Boundary Attention Gates at every decoder stage, refined by a Boundary Refinement Block, and jointly optimized with an auxiliary Signed Distance Map (SDM) regression head through a hybrid multi-task loss.Results: The resulting supervision strategy is evaluated on the public MCOA benchmark (2,992 AS-OCT images), exceeding the best previously reported Dice under an identical five-fold protocol and substantially reducing boundary error (HD95, ASD). A replicated ablation across MCOA and the independently collected AIDK dataset shows that individual geometric constraints contribute in a statistically validated, complementary, dataset-dependent manner, and external validation under zero-shot and fully supervised transfer confirms the strategy generalizes beyond a single dataset.Conclusions: Explicit geometric supervision of boundary location and distance-to-boundary yields consistent gains in boundary-level accuracy for thin, low-contrast anatomical structures, supporting a supervision paradigm applicable beyond the cornea.