Optimized left atrial appendage segmentation via hierarchical cross-scale consistency with gated routing for echocardiography

Accurate left atrial appendage (LAA) delineation is vital for atrial fibrillation-related left atrial appendage occlusion (LAAO) planning and device sizing; yet transthoracic echocardiography (TTE) segmentation is challenged by speckle noise, acoustic shadowing, low or variable contrast, pronounced anatomical variability, and limited annotated data. We propose a lightweight, plug-in module, the Hierarchical Cross-Scale Consistency with Gated Routing (H-CSCA), which preserves the U-Net topology and can be integrated into existing U-Net decoders as a drop-in replacement for skip connections with minimal architectural changes and parameter overhead. Our key contribution is to couple cross-scale cycle consistency with learnable gated routing within the U-Net decoder, enabling controllable fusion for echo boundary refinement. Our model aligns the encoder and decoder features across neighboring scales and selectively routes information between them to reinforce fine boundary cues without relying on transformer-heavy backbones. Unlike prior multi-scale U-Net variants that rely on fixed skip and fusion pathways, H-CSCA explicitly enforces cross-scale agreement and adaptively selects which scales contribute to refinement via learnable routing. We evaluated our method on an anonymized dataset of LAA TTE views annotated by cardiologists, with a strict patient-level split (16 training, 5 validation, 5 test; held-out, locked test set). Compared to the MSCCA-UNet baseline, H-CSCA improved the Dice coefficient from 0.9490 ± 0.0030 to 0.9598 ± 0.0026 and IoU from 0.9040 ± 0.0040 to 0.9232 ± 0.0037, with only a marginal parameter increase (15.45M to 15.71M). Boundary quality also improved, reducing mean absolute distance from 0.8223 ± 0.0791 to 0.7918 ± 0.0613 and reducing worst-case boundary deviation (Hausdorff) by approximately 2%. Overall, H-CSCA yields sharper, more stable contours at minimal computational cost, supporting reliable LAAO planning and guidance in data-limited echocardiography workflows.