A 3D-2D Rigid Liver Registration Method Using Pre-Training and Transfer Learning With Staged Alignment of Anatomical Landmarks

Augmented reality navigation in laparoscopic liver resection can integrate surgical planning information such as liver resection lines, blood vessels, and tumors to enhance surgical safety. However, the 3D-2D registration still faces challenges, including long registration time and manual initialization. Preoperative 3D liver point cloud and intraoperative laparoscopic image data are pre-trained to generate a patient-specific initial pose. A staged fine registration strategy targeting local anatomical landmarks is employed, involving normalization of the distance loss between the projection points of various anatomical landmarks in the preoperative 3D model and the corresponding ground truth landmarks in the intraoperative 2D laparoscopic images. The proposed method was evaluated using pixel-wise reprojection error (RPE) and target registration error (TRE). The results demonstrate that the method achieves superior registration accuracy compared to existing rigid registration methods. Deep learning integrated into 3D-2D rigid registration achieved full automation and sped up the computation.