Hybrid Multimodal Deformable Registration with a Data-Driven Deformation Prior

Deformable registration for images with different contrast-enhancement and hence different structure appearance is extremely challenging due to the ill-posed nature of the problem. Utilizing prior anatomical knowledge is thus necessary to eliminate implausible deformations. Landmark constraints and statistically constrained models have shown encouraging results. However, these methods do not utilize the segmentation information that may be readily available. In this paper, we explore the possibility of utilizing such information. We propose to generate an anatomical correlation-regularized deformation field prior by registration of point sets using mixture of Gaussians based on a thin-plate spline parametric model. The point sets are extracted from the segmented object surface and no explicit landmark matching is required. The prior is then incorporated with an intensity-based similarity measure in the deformable registration process using the variational framework. The proposed prior does not require any training data set thus excluding any inter-subject variations compared to learning-based methods. In the experiments, we show that our method increases the registration robustness and accuracy on 12 sets of TAVI patient data, 8 myocardial perfusion MRI sequences, and one simulated pre-and post-tumor resection MRI.