Knee Joint Registration Based on Chart and Learnable Superpoint Match

Registration between the segmented and captured point clouds of the knee is the basic for the alignment of preoperative images with intraoperative patient in total knee arthroplasty navigation. The state-of-the-art point cloud registration methods uniformly or randomly select superpoints to employ a coarse-to-fine registration strategy. Due to the uneven density of shape features in different parts for the knee joint, this selection cannot fully utilize the topological information of bone point cloud. Moreover, the relatively low overlap between the preoperative complete point cloud and the intraoperative partial point cloud makes the accurate registration hard. A point cloud registration method based on chart and learnable superpoint match is proposed in this paper. The chart-based topological information module is used to optimize the selection of superpoints, and a learnable superpoint match module is introduced to address the low overlap and shape distribution variations. Comparative experiments were conducted on multiple methods using two knee joint datasets, and the results showed the superiority of the proposed method.