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

Abstract

Objective: Registering a preoperative 3D model of an organ with its actual anatomy viewed from an intraoperative video is a fundamental challenge in computer-assisted surgery, especially for surgical augmented reality. To address this, we present a benchmark of state-of-the-art deep learning point-cloud registration methods, offering a transparent evaluation of their generalizability to surgical scenarios and establishing a robust guideline for developing advanced non-rigid algorithms. Methods: We systematically evaluate traditional and deep learning GMM-based, correspondence-based, correspondence-free, matching-based, and liver-specific point cloud registration approaches on two surgical datasets: a deformed IRCAD liver set and DePoll dataset. We also propose our complete-to-partial point cloud registration framework that leverages keypoint extraction, overlap estimation, and a Transformer-based architecture, culminating in competitive registration results. Results: Experimental evaluations on deformed IRCAD tests reveal that most deep learning methods achieve good registration performances with TRE<10 mm, MAE(R) < 4 and MAE(t)<5 mm. On DePoll, however, performance drops dramatically due to the large deformations. Conclusion: In conclusion, deep-learning rigid registration methods remain reliable under small deformations and varying partiality but lose accuracy when faced with severe non-rigid changes. To overcome this, future work should focus on building non-rigid registration architectures that preserve the strengths of self-, cross-attention and overlap modules while enhancing correspondence estimation to handle large deformations in laparoscopic surgery.