Semantic-ICP: Iterative Closest Point for Non-rigid Multi-Organ Point Cloud Registration

TL;DR

This work introduces Semantic-ICP (SemICP), a non-rigid point-cloud registration framework that leverages semantic organ labels and linear elastic regularization to address generalizability and explainability gaps in learning-based CAI methods. The method uses a rigid initialization with semantic-consistent matching, followed by a non-rigid deformation represented on a control-point grid and regularized by elastic, magnitude, and gradient terms. Across four datasets, SemICP achieves lower surface distances (HD95 and MSD) than state-of-the-art baselines and remains effective when integrated with learned segmentation labels, highlighting its potential for robust US–MR and multi-organ registrations in clinical workflows. The approach offers a controllable, physics-informed alternative to purely data-driven methods, enabling plausible deformations with practical runtimes suitable for intraoperative use.

Abstract

Point cloud registration is important in computer-aided interventions (CAI). While learning-based point cloud registration methods have been developed, their clinical application is hampered by issues of generalizability and explainability. Therefore, classical point cloud registration methods, including Iterative Closest Point (ICP), are still widely applied in CAI. ICP methods fail to consider that: (1) the points have well-defined semantic meaning, in that each point can be related to a specific anatomical label; (2) the deformation required for registration needs to follow biomechanical energy constraints. In this paper, we present a novel non-rigid semantic ICP (SemICP) method that handles multiple point labels and uses linear elastic energy regularization. We use semantic labels to improve the robustness of closest point matching and propose a novel point cloud deformation representation that incorporates explicit biomechanical energy regularization. Our experiments on four datasets show that our method significantly improves the Hausdorff distance and mean surface distance compared with other point cloud registration methods. We also demonstrate that integrating deep learning segmentation models with our registration pipeline enables effective alignment of US and MR point clouds.

Figures (7)

• Figure 1: The workflow of our proposed Sem-ICP. The method includes a rigid initialization, and then the deformation is estimated iteratively.
• Figure 2: Visualization of the registration results on one example US-CT pair in TORS dataset (posterior-anterior view). The medial-lateral view is shown in supplemental material Fig. S1.
• Figure 3: Visualization of the registration results on one example CT-MR pair in the AbdominalMRCT dataset (posterior-anterior view). Another viewpoint showing the left-to-right perspective is shown in the supplemental material Fig. S2.
• Figure 4: Visualization of the registration results on one example MR-MR pair in HippocampusMR dataset (superior-inferior view). Another viewpoint showing the left-to-right perspective is shown in the supplemental material Fig. S3.
• Figure 5: The proof-of-concept integration with learned labels study.