SGL-UNet: Pushing the Boundary of Extreme Lightweight Segmentation for Point-of-Care Biliary Atresia Screening

Early ultrasound screening for Biliary Atresia (BA) is hindered by lesion atrophy and low echogenicity contrast, rendering automatic segmentation an ill-posed problem. Compounding this challenge, the strict computational constraints of Point-of-Care Ultrasound (POCUS) devices preclude the deployment of existing high-accuracy networks. This paper investigates the unexplored accuracy upper bound of BA segmentation under an extreme parameter budget (<3M). We propose SGL-UNet, a framework engineered via three constraint-aware principles: 1) A DWSE encoder that prioritizes spatial gradient extraction to amplify boundary cues while minimizing redundancy; 2) A Coarse-Grained Global Transformer that leverages spatial reduction attention to capture global anatomical context with O(N²/r²) complexity; and 3) A LightFusion decoder that enforces high-frequency detail preservation via residual modulation. Experiments on PBUS and BUSI datasets demonstrate that SGL-UNet achieves 84.08% IoU with merely 2.62M parameters (approximately 7.6% of U-Net), outperforming state-of-the-art methods and advancing the Pareto frontier for mobile healthcare applications.