Adaptive Control Point B-Spline Method for Improving Accuracy and Deformation Plausibility in Deep Learning Brain...

Deformable brain image registration plays a crucial role in aligning and jointly analyzing multi-phase and multi-modal neuroimaging data. However, existing unsupervised deep learning-based registration methods typically predict dense displacement fields directly at the voxel level, which often leads to deformation noise and optimization difficulties due to redundant degrees of freedom. Furthermore, these methods are prone to topological distortions when noise or irrelevant information is present. To address these challenges, we propose a robust B-spline registration post-processing framework with adaptive control point density, which can be seamlessly integrated into mainstream registration architectures for brain analysis. Specifically, we design a Control Point Spacing Predictor (CPSPredictor) that adaptively adjusts the distribution density of control points based on the input images, enabling effective regulation of registration flexibility. By modeling displacements with sparse control points and generating continuous and smooth deformation fields via locally supported B-spline basis functions, the proposed method significantly suppresses folding artifacts and preserves anatomical consistency. Extensive experiments on multiple publicly available brain MRI datasets demonstrate that our approach can be seamlessly integrated into existing mainstream registration frameworks, significantly improving both registration accuracy and the physical plausibility of the transformation. For instance, on the IXI dataset, our method boosts the VoxelMorph baseline's DSC by 2.63% and reduces the number of folding points nearly 69-fold. The resulting registration outputs better align with clinical diagnostic requirements, highlighting the method's robustness and practical application potential.