Aims: Accurate identification of left atrial (LA) scar from static late gadolinium-enhanced (LGE)-MRI remains challenging due to low contrast, thin atrial walls, and patient variability. Precise scar localization is critical for planning and guiding catheter ablation therapy. This study aims to develop a motion-based framework that infers scar regions from static imaging by analyzing abnormal motion patterns leveraging deformable registration to a healthy atlas, reducing dependence on accurate scar segmentation or cine data.
Methods: LA cavity and scar segmentation were obtained using 3D nnUNet and nnFormer, where segmentation was used primarily for anatomical alignment. Patient volumes were resampled and pre-processed before deformably registered to a healthy cine-derived LA atlas using three approaches: SimpleITK dense displacement vector field (DDVF), ANTs SyN, and BSpline free-form deformation (FFD). Voxel-wise deformation magnitudes were compared to healthy atlas statistics, and abnormal motion was quantified using a Mahalanobis distance-based metric within a morphologically defined LA wall mask. Voxels exceeding an adaptive threshold (μ + 3σ) were classified as abnormal regions. We used the LAScarQS 2022 and the Aladdin CMR LA datasets for segmentation and healthy atlas generation, respectively.
Results: The LA segmentation was robust (Dice 0.91/0.92) using nnUNet and nnFormer, while scar segmentation remained modest (Dice 0.49–0.51), supporting the need for motion-based inference. SimpleITK-DDVF produced the most stable deformation fields, with mean Jacobian values closest to 1 and no negative Jacobians, and consistently detected abnormal motion across all evaluated patients. ANTs and BSpline showed more variable registration performance with some detection failures.
Conclusion: Motion-derived abnormalities provide a robust alternative for scar localization from static LGE-MRI. By leveraging atlas-based deformation analysis, the proposed framework enables clinically meaningful identification of scar-related regions without requiring cine imaging or highly accurate scar segmentation, supporting improved ablation planning.