Non-rigid Registration with Missing Correspondences in Preoperative and Postresection Brain Images

Registration of preoperative and postresection images is often needed to evaluate the effectiveness of treatment. While several non-rigid registration methods exist, most would be unable to accurately align these types of datasets due to the absence of tissue in one image. Here we present a joint registration and segmentation algorithm which handles the missing correspondence problem. An intensity-based prior is used to aid in the segmentation of the resection region from voxels with valid correspondences in the two images. The problem is posed in a maximum a posteriori (MAP) framework and optimized using the expectation-maximization (EM) algorithm. Results on both synthetic and real data show our method improved image alignment compared to a traditional non-rigid registration algorithm as well as a method using a robust error kernel in the registration similarity metric.