Benchmarking complete-to-partial point cloud registration techniques for laparoscopic surgery.

Area of Science:

• Medical Imaging
• Computer-Assisted Surgery
• Machine Learning

Background:

• Accurate 3D organ model registration with intraoperative video is crucial for surgical augmented reality.
• Existing deep learning point-cloud registration methods need evaluation for surgical scenarios.

Purpose of the Study:

• To benchmark state-of-the-art deep learning point-cloud registration methods.
• To evaluate their generalizability in surgical contexts.
• To provide guidelines for developing advanced non-rigid registration algorithms.

Main Methods:

• Systematic evaluation of traditional and deep learning registration approaches (GMM-based, correspondence-based, etc.) on surgical datasets (IRCAD, DePoll).
• Proposal of a novel complete-to-partial point cloud registration framework using keypoint extraction, overlap estimation, and a Transformer architecture.

Main Results:

• Deep learning methods achieved good registration performance (TRE<10 mm) on the IRCAD liver dataset with minor deformations.
• Performance significantly dropped on the DePoll dataset due to large deformations.

Conclusions:

• Deep learning rigid registration is reliable for small deformations but lacks accuracy with severe non-rigid changes.
• Future research should focus on non-rigid registration architectures enhancing correspondence estimation for laparoscopic surgery.