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

Figures

• 

  FIGURE 1. Training‐time augmentation policy. The far‐left panel shows the original image; subsequent panels illustrate small in‐plane rotations, isotropic scaling, elastic warps, brightness/contrast shifts and region‐aware random crops used to increase robustness. Flips and 90° rotations are intentionally excluded to preserve anatomical plausibility.

• 

  FIGURE 2. Structural overview of the proposed DAT–DAD U‐Net with SE‐augmented skip connections.

• 

  FIGURE 3. The architecture of the deformable attention module, showing the sequential process of embedding, layer normalisation, deformable multi‐head attention, residual connections and multi‐layer perceptron (MLP) layers for feature refinement.

• 

  FIGURE 4. The schematic diagram of the deformable attention architecture.

• 

  FIGURE 5. Structure of the dual attention with deformable convolution block, illustrating the integration of position and channel attention modules along with a deformable convolution layer to enhance feature extraction and adaptability in capturing spatial and channel‐specific details.

• 

  FIGURE 6. Diagram depicting the architecture of the channel and spatial squeeze‐and‐excitation with deformable convolution unit, illustrating its configuration for enhancing feature selection and integration across spatial and channel dimensions.

• 

  FIGURE 7. Qualitative segmentation on the held‐out test set. Columns (left→right): (1) original 2D echocardiography slice, (2) U‐Net prediction, (3) Attention U‐Net prediction and (4) proposed DAT–DAD U‐Net (AERO‐selected) prediction. All masks are binarised with the same threshold and post‐processing.

• 

  FIGURE 8. Top: Dice; Bottom: IoU/Jaccard—box‐and‐whisker plots on the held‐out, patient‐wise test set comparing U‐Net, Attention U‐Net and the proposed DAT–DAD U‐Net (AERO‐selected). The proposed model shows higher medians and tighter IQRs than the baselines, indicating better overlap and more stable performance.

• 

  FIGURE 9. Ablation‐based analysis of architectural component contributions supporting robustness across anatomical variability, shown in terms of Dice and HD95.

• 

  FIGURE 10. AERO operating‐point trade‐off analysis in terms of Dice and HD95 across deployment‐oriented configurations.

• 

  FIGURE 11. Patient‐wise edge consistency (HD95, mm) for U‐Net, Attention U‐Net and the proposed DAT–DAD U‐Net (AERO‐selected) across samples S1–S5 (top) and S6–S10 (bottom).

• 

  FIGURE 12. Convergence of hyperparameter optimisers for the DAT–DAD U‐Net. (a) Single training run: AERO (black) drives the mean‐square error down fastest and reaches the lowest floor versus Random Search, BayesOpt (GP) and TPE. (b) Mean ± SD over three runs shows AERO's consistent advantage and reduced variance.