NeuralBoneReg: A Novel Self-Supervised Method for Robust and Accurate Multi-Modal Bone Surface Registration

TL;DR

NeuralBoneReg addresses cross-modal bone surface registration in CAOS by learning a neural unsigned distance field (UDF) from preoperative bone surfaces (NeuralUDF) and performing parallel SE(3) hypothesis-based registration (NeuralReg) on intraoperative data. The framework is self-supervised and modality-agnostic, avoiding large paired datasets and ground-truth transformations, while shifting heavy computation to the preoperative phase. Evaluations on UltraBones100k, UltraBones-Hip, and SpineDepth show mean rotation errors around $1.68$–$3.79$ degrees and translation errors around $1.86$–$2.45$ mm, often surpassing traditional and some supervised baselines, with strong generalization across anatomies and modalities. The approach offers a practical CAOS pipeline with robust registration under multimodal imaging, and releases new data and code to support further research and clinical translation.

Abstract

In computer- and robot-assisted orthopedic surgery (CAOS), patient-specific surgical plans derived from preoperative imaging define target locations and implant trajectories. During surgery, these plans must be accurately transferred, relying on precise cross-registration between preoperative and intraoperative data. However, substantial modality heterogeneity across imaging modalities makes this registration challenging and error-prone. Robust, automatic, and modality-agnostic bone surface registration is therefore clinically important. We propose NeuralBoneReg, a self-supervised, surface-based framework that registers bone surfaces using 3D point clouds as a modality-agnostic representation. NeuralBoneReg includes two modules: an implicit neural unsigned distance field (UDF) that learns the preoperative bone model, and an MLP-based registration module that performs global initialization and local refinement by generating transformation hypotheses to align the intraoperative point cloud with the neural UDF. Unlike SOTA supervised methods, NeuralBoneReg operates in a self-supervised manner, without requiring inter-subject training data. We evaluated NeuralBoneReg against baseline methods on two publicly available multi-modal datasets: a CT-ultrasound dataset of the fibula and tibia (UltraBones100k) and a CT-RGB-D dataset of spinal vertebrae (SpineDepth). The evaluation also includes a newly introduced CT--ultrasound dataset of cadaveric subjects containing femur and pelvis (UltraBones-Hip), which will be made publicly available. NeuralBoneReg matches or surpasses existing methods across all datasets, achieving mean RRE/RTE of 1.68°/1.86 mm on UltraBones100k, 1.88°/1.89 mm on UltraBones-Hip, and 3.79°/2.45 mm on SpineDepth. These results demonstrate strong generalizability across anatomies and modalities, providing robust and accurate cross-modal alignment for CAOS.

Figures (8)

• Figure 1: Network architecture of the two-staged approach of NeuralBoneReg. In the preoperative stage, the NeuralUDF module learns an implicit neural representation of the target bone surface from the preoperative data $\mathcal{C}$. During the intraoperative stage, the NeuralReg module leverages multiple lightweight neural heads built on a randomly initialized shared backbone to efficiently generate registration hypotheses aligning the intraoperative data $\mathcal{U}$ with the preoperative data $\mathcal{C}$. FC($\text{dim}_\text{in}$, $\text{dim}_\text{out}$) denotes a fully connected layer with input dimension $\text{dim}_\text{in}$ and output dimension $\text{dim}_\text{out}$, followed by a Tanh activation.
• Figure 2: Recall–rate curves for UltraBones100k (first column), UltraBones-Hip (second column), and SpineDepth (last column). $x$ denotes the threshold and $y$ denotes $\text{RR}(x)$. The top row shows the full range, while the bottom row focuses on $0 \le x \le 30$.
• Figure 3: Exemplary qualitative results on UltraBones100k. The preoperative point cloud $\mathcal{C}$ is visualized in yellow, while the intraoperative point cloud $\mathcal{U}$ is visualized in blue. The mean relative translation error (mm) and rotation error (°) are reported below each figure. Sever errors with RRE$>$10 ° or RTE$>$10 mm are highlighted in red.
• Figure 4: Exemplary qualitative results on UltraBones-Hip. The preoperative point cloud $\mathcal{C}$ is visualized in yellow, while the intraoperative point cloud $\mathcal{U}$ is visualized in blue. The mean relative translation error (mm) and rotation error (°) are reported below each figure. Sever errors with RRE$>$10 ° or RTE$>$10 mm are highlighted in red.
• Figure 5: Exemplary qualitative results on SpineDepth. The preoperative point cloud $\mathcal{C}$ is visualized in yellow, while the intraoperative point cloud $\mathcal{U}$ is visualized in blue. The mean relative translation error (mm) and rotation error (°) are reported below each figure. Sever errors with RRE$>$10 ° or RTE$>$10 mm are highlighted in red.