Resolving the Ambiguity of Complete-to-Partial Point Cloud Registration for Image-Guided Liver Surgery with Patches-to-Partial Matching

TL;DR

The paper tackles the challenge of automatic initial rigid registration between a complete preoperative liver surface and a partially visible intraoperative surface, where complete-to-partial ambiguity is prevalent in laparoscopic surgery. It introduces a patches-to-partial (P2P) matching module that plugs into existing learning-based point cloud registration pipelines, transforming a single ambiguous registration into multiple patch-based hypotheses without altering end-to-end learning. The authors validate the approach on a large in silico liver dataset with deformations and a phantom in vitro dataset, showing significant improvements in registration accuracy and robustness under low visibility, noise, and deformation, with statistically significant gains. The work also provides a public benchmark, discusses limitations of current baselines and datasets, and emphasizes clinical relevance by enabling more reliable, automatic initialization for image-guided liver surgery workflows and potentially extending to other organs.

Abstract

In image-guided liver surgery, the initial rigid alignment between preoperative and intraoperative data, often represented as point clouds, is crucial for providing sub-surface information from preoperative CT/MRI images to the surgeon during the procedure. Currently, this alignment is typically performed using semi-automatic methods, which, while effective to some extent, are prone to errors that demand manual correction. Point cloud correspondence-based registration methods are promising to serve as a fully automatic solution. However, they may struggle in scenarios with limited intraoperative surface visibility, a common challenge in liver surgery, particularly in laparoscopic procedures, which we refer to as complete-to-partial ambiguity. We first illustrate this ambiguity by evaluating the performance of state-of-the-art learning-based point cloud registration methods on our carefully constructed in silico and in vitro datasets. Then, we propose a patches-to-partial matching strategy as a plug-and-play module to resolve the ambiguity, which can be seamlessly integrated into learning-based registration methods without disrupting their end-to-end structure. It has proven effective and efficient in improving registration performance for cases with limited intraoperative visibility. The constructed benchmark and the proposed module establish a solid foundation for advancing applications of point cloud correspondence-based registration methods in image-guided liver surgery.

Figures (10)

• Figure 1: Comparison of widely used public point registration datasets in computer vision (3DMatch, 4DMatch, ModelNet40) and our liver registration dataset. In each example, the source point clouds (blue) are aligned with the target point clouds (red). 3DMatch, 4DMatch, and ModelNet40 focus on partial-to-partial cases. The liver registration dataset is distinct from these datasets, designed under the complete-to-partial scenarios.
• Figure 2: Illustration of the general paradigm of learning-based correspondence registration methods (top) and our plug-and-play P2P module (bottom). The input source and target point clouds are voxelized to maintain similar densities, ensuring a uniform and continuous representation of the liver surface regions (see §\ref{['subsec:Preprocessing']} and §\ref{['sec:dicsB']}). Our proposed module generates candidate patches, samples their point-wise features from a learning-based correspondence registration method, and performs feature matching and rigid transformation estimation for each patch with the target point cloud. Finally, a candidate selection rule determines the optimal rigid transformation. We present the case where the patch number $k$ is set to 5. The patch consists of the same number of points as the target point cloud (see §\ref{['subsec:patches']} and §\ref{['subsec:p2p3']}).
• Figure 3: Illustration of the in silico phantom generation process. Source and target point clouds/ meshes are shown in blue and red, respectively.
• Figure 4: Properties of the testing set from the in silico phantom dataset: (a) Distribution of RMS-TRE across all sample pairs after rigid alignment using volumetric vertices correspondences to remove rigid components. (b) Distribution of max-TRE per sample pair after rigid alignment. (c) Initial RMS-TRE across samples after applying random rigid transformations, used to evaluate rigid registration methods. (d) Visibility ratio of the target point cloud across samples after random cropping.
• Figure 5: Qualitative comparison of registration results on the in silico phantom dataset. The source and target point clouds are shown in blue and red, respectively.