Laparoscopic liver point cloud registration via hyperbolic-topology interaction.

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Geometric Deep Learning

Background:

• Precise registration of 3D models to intraoperative data is crucial for laparoscopic liver resection.
• Challenges include liver deformation, respiratory motion, and limited visibility, hindering accurate alignment.

Purpose of the Study:

• To develop a robust registration method for stable alignment of preoperative 3D models and intraoperative point clouds.
• To address challenges of low-overlap and non-rigid conditions in laparoscopic liver resection.

Main Methods:

• A novel framework utilizing a Hyperbolic-Topology Interaction Module.
• Mapping rotation-invariant features into hyperbolic space to amplify geometric differences.
• Constructing a topological graph for enhanced feature consistency and robust correspondence estimation via coarse-to-fine matching.

Main Results:

• Achieved state-of-the-art results on simulated and real datasets.
• Lowest surface target registration error (TRE) of 6.2 mm and internal TRE of 7.0 mm on the DePOLL dataset.
• Demonstrated strong generalization capability under varying deformation conditions.

Conclusions:

• The method combines local geometric and global topological reasoning for improved accuracy, robustness, and efficiency.
• Provides reliable rigid initialization for augmented reality (AR)-guided resection navigation.
• Shows strong potential for clinical application in liver surgery.