Selective Boundary-aware Skip Connections for High-precision Ophthalmic Image Segmentation

Background: Ophthalmic image segmentation plays a key role in diagnosing retinal conditions. U-Net and its variants, though effective, often propagate redundant features that impair boundary accuracy. Objective: To propose a novel segmentation framework using a Boundary-Aware Gating Mechanism (BAGM) to enhance precision in retinal boundary delineation. Methods: The architecture integrates depth-wise separable convolutions, boundary-guided spatial attention, and a hybrid loss function (Dice + boundary-aware focal loss). The model was validated on OCT-500, DRIVE, and REFUGE datasets. Results: Our model achieved superior Dice score (0.923 ± 0.012), improved boundary F-score, and reduced Hausdorff distance, outperforming Attention U-Net and TransUNet while maintaining computational efficiency. Conclusion: The proposed dynamic boundary-gated architecture is both accurate and lightweight, suitable for real-time clinical settings such as OCT and fundus imaging.