Consensus approach to shape regularization in cross-modality cardiac segmentation

Adapting segmentation models to new imaging modalities without original training data is known as source-free domain adaptation (SFDA). SFDA is often unstable and limited to two-dimensional slices, causing inconsistencies and errors between adjacent slices. This problem is apparent in cardiac segmentation when adapting from computed tomography (CT) to magnetic resonance imaging (MRI), as these inconsistencies lead to three-dimensional anatomical distortions. In this study, we analyzed a postprocessing framework for SFDA-based cardiac segmentation using a probabilistic consensus map (PCMap). Inspired by traditional probabilistic atlases, PCMap is an average anatomical map generated from all initial segmentations, intended to enforce structural consistency. This framework was evaluated on CT and MRI volumes from the publicly available multi-modality whole-heart segmentation dataset. PCMap-guided refinement increased shape regularity and reduced interslice discontinuities but did not consistently improve voxel-wise accuracy. A correlation of 0.803 between reference quality and refinement performance was found, showing that the choice of reference volume strongly affects refinement performance.