NeuralBoneReg: An instance-specific label-free point cloud-based method for multi-modal bone surface registration.

Area of Science:

• Medical imaging
• Computer-assisted surgery
• Orthopedic surgery

Background:

• Accurate registration of preoperative and intraoperative data is crucial for computer- and robot-assisted orthopedic surgery (CAOS).
• Heterogeneity in imaging modalities and devices presents challenges, leading to registration inaccuracies.
• Robust, automatic, and modality-agnostic bone surface registration methods are needed to improve surgical accuracy.

Purpose of the Study:

• To introduce NeuralBoneReg, a novel self-supervised framework for instance-specific bone surface registration.
• To enable accurate cross-modal alignment of bone surfaces in CAOS.
• To overcome limitations of existing supervised and modality-dependent registration techniques.

Main Methods:

• NeuralBoneReg utilizes a surface-based, self-supervised framework with 3D point clouds.
• It incorporates an implicit neural unsigned distance field (UDF) module for bone model representation.
• A multilayer perceptron (MLP)-based registration module handles global initialization and local refinement using a coarse-to-fine strategy.

Main Results:

• NeuralBoneReg demonstrated competitive performance across various anatomies and modalities on public datasets (UltraBones100k, SpineDepth).
• The method achieved stable performance on a new cadaveric dataset (UltraBonesHip), outperforming others where they degraded.
• Quantitative metrics indicate accuracy close to pseudo ground truth across all evaluated datasets.

Conclusions:

• NeuralBoneReg provides robust, accurate, and modality-agnostic registration of bone surfaces.
• It presents a promising solution for reliable cross-modal alignment in CAOS.
• The self-supervised, instance-specific approach reduces reliance on large, labeled training datasets.