Automatic Intraoperative CT-CBCT Registration For Image-Guided Pelvic Fracture Reduction

Pelvic fractures are highly complex and threaten the stability of the pelvic ring. Accurate surgical navigation in closed fracture reduction necessitates precise preoperative to intraoperative image registration, typically involving iterative closest point (ICP) matching between CT and CBCT models. The automation of this process is currently limited, with frequent need for trial-and-error adjustment and compromised precision due to the noise and artifacts that affect segmentation quality. In this study, we develop a deep learning-based intraoperative CT-CBCT registration pipeline to enhance both the efficiency and reliability of this process. A unique cross-modality image annotation scheme and transfer learning from CT data are used to train a CBCT segmentation network amid the challenge of limited CBCT data. A lightweight landmark detection network is incorporated to facilitate initial alignment before ICP matching. Our method took 10 seconds to execute and achieved a Dice of 0.94 in hipbone segmentation and 0.97 mm error in final registration, demonstrating significant improvements over the traditional approach.