Motion-guided scar detection from static left atrial MRI via deformable registration to a healthy atlas

Accurate identification and localization of left atrial (LA) scar tissue provides critical information for guiding catheter ablation therapy in atrial fibrillation (Afib). However, the assessment of LA scar from motion remains challenging due to the lack of standardized imaging protocols and motion-resolved data, compounded by the thin atrial wall, variability across patients, and limitations in spatial resolution of MRI. To overcome these limitations, we propose a framework that infers scar regions from static late gadolinium-enhanced (LGE)-MRI by analyzing abnormal motion patterns derived from deformable registration to a healthy motion atlas. Our method combines 3D segmentation method with deformable registration to a cine derived LA atlas constructed from healthy subjects. While direct scar segmentation from LGE-MRI is often unreliable due to poor contrast and heterogeneity, LA segmentation is highly consistent and reproducible, serving as a robust anatomical prior. The segmented LA masks enable accurate shape-based alignment and facilitate the extraction of deformation vector fields (DVFs). We compute the voxel-wise Mahalanobis distance between the patient DVF magnitude and healthy atlas statistics within a morphologically defined left atrial mask. This helps to quantify deviations from normal motion, and voxels exceeding an adaptive threshold are labeled as regions of abnormal motion. These motion outliers, constrained by the wall, indicate functionally stiff atrial tissue regions that often extend beyond the LGE-defined scar core, serving as biomarkers of diffuse tissue remodeling. Unlike traditional methods that require full-cycle cine data or motion-resolved contrast sequences, our framework supports scar detection and localization from static LGE-MRI alone. We implemented two 3D segmentation frameworks, nnUNet and nnFormer, and evaluated three deformable registration strategies—SimpleITK-DDVF, ANTs, and BSpline and demonstrated that our approach achieves reliable performance without the need for highly accurate scar segmentation. These results demonstrate that motion-based markers can complement conventional intensity-based methods and and have the potential to improve ablation therapy planning by identifying scarred areas that should be avoided during treatment.