Iterative Similarity Perturbation Point Cloud Registration Based on Deformation-Resistant Region Detection

Image-guided surgery has been widely applied in skull base surgery, where precise registration between preoperative images and the intraoperative patient space is essential. However, facial skin often undergoes local non-rigid deformation during surgery due to anesthesia and surgical manipulation, severely impairing the accuracy of point cloud registration. To address this issue, this study proposes an iterative optimization approach for point cloud registration via second-order similarity perturbation, guided by deformation-resistant regions identified through incremental geometric descriptors. A teacher-student structured network for detecting deformation-resistant regions is introduced, employing a shared architecture and incremental training strategy across different data domains to effectively identify intraoperative non-rigid deformation regions. Based on this, a deformation-score-guided registration mechanism is designed, where initial pose is generated using second-order similarity perturbation, and registration robustness is enhanced through multiple local optimization iterations. A high-quality facial point cloud registration dataset encompassing various facial deformations, occlusions, and pose variations was constructed for training. Experimental results demonstrate that the proposed method outperforms existing methods in both registration accuracy and stability. The presented framework not only achieves high-precision registration under non-rigid disturbances but also provides new insights for facial point cloud tasks.