Interpretable Model for Clinical Use in Left Atrial Appendage Segmentation via an Optimised Deformable-Attention...

Main Methods:

• Development of a novel DAT-DAD U-Net architecture incorporating deformable attention transformers (DAT) for global context and a dual attention with deformable convolution (DAD) block for boundary refinement.
• Integration of spatial-channel squeeze-and-excitation (SE) augmentation for improved multi-scale feature fusion.
• Hyperparameter optimization using AERO, a surrogate- and multi-fidelity-driven approach, applied to a 22-patient 3D echocardiography cohort with on-the-fly anatomy-preserving augmentation.

Main Results:

• The proposed DAT-DAD U-Net achieved high segmentation accuracy, with a Dice score of 0.8925 ± 0.0144, IoU of 0.8026 ± 0.0156, and HD95 of 9.14 ± 1.96 mm.
• Ablation studies confirmed the significant additive contributions of DAT, DAD, and SE components to model performance.
• The model demonstrated faster convergence and a lower error floor compared to baseline methods, indicating improved efficiency and robustness.

Conclusions:

• The AERO-optimised DAT-DAD U-Net offers a fully automated and accurate solution for LAA segmentation from 3D echocardiography.
• The model's ability to capture global context and refine boundaries addresses key challenges in segmenting the morphologically variable LAA.
• This approach supports operator-light, time-sensitive workflows for LAA device occlusion planning in AF patients.