Interpretable Model for Clinical Use in Left Atrial Appendage Segmentation via an Optimised Deformable-Attention U-Net With Spatial-Channel Fusion.

Accurate segmentation of the left atrial appendage (LAA) is essential for device occlusion planning in atrial fibrillation patients who cannot receive anticoagulation. Yet 3D echocardiography suffers from low signal-to-noise ratio, anisotropy and marked morphological variability, increasing overfitting risk and reliance on operator-dependent post-processing. Existing U-Net variants capture local detail but often miss long-range dependencies, while transformers improve context at the cost of boundary precision; semi-automated pipelines still require experts. We present a fully automated, AERO-optimised DAT-DAD U-Net with SE-augmented skip fusion. Deformable attention transformers (DAT) provide content-adaptive global context, while a dual attention with deformable convolution (DAD) block refines rims and addresses shape irregularities; spatial-channel squeeze-and-excitation improves multi-scale fusion. Hyperparameters are selected by AERO, a surrogate- and multi-fidelity-driven optimiser balancing exploration and exploitation under limited data in practice. Validation used a 22-patient 3D echocardiography cohort from King's College Hospital. Volumes were reformatted into axial 2D slices, trained with on-the-fly anatomy-preserving augmentation and evaluated using strict patient-wise splits. The model achieved Dice 0.8925 ± 0.0144, IoU 0.8026 ± 0.0156 and HD95 9.14 ± 1.96 mm. Ablations confirmed additive gains from DAT, DAD and SE, with faster convergence and a lower error floor, supporting operator-light, time-sensitive LAA workflows.