Density-adaptive registration of pointclouds based on Dirichlet Process Gaussian Mixture Models.

Area of Science:

• Medical Imaging
• Computer Vision
• Robotic Surgery

Background:

• Accurate registration of pre- and intra-operative patient anatomy is crucial for augmented reality (AR) systems in minimally invasive surgery.
• Existing methods struggle with variations in point density and spatial overlap between pre- and intra-operative pointclouds.

Purpose of the Study:

• To propose a novel algorithm for robust rigid registration of patient anatomy pointclouds.
• To improve the accuracy and efficiency of AR-guided surgical interventions.

Main Methods:

• A probabilistic approach using a Dirichlet Process Gaussian Mixture Model (DPGMM) to represent transformed pointclouds.
• Minimizing Kullback-Leibler divergence within a variational Bayesian inference framework.
• Utilizing KDTrees for coarse-to-fine data and model expansion, with neighborhood-based scanning weights for robustness.

Main Results:

• The proposed algorithm demonstrates comparable accuracy to existing Gaussian Mixture Model (GMM) methods.
• Achieves higher computational efficiency compared to GMM methods, which are sensitive to the number of model components.
• Exhibits robustness to variations in point density, noise, outliers, and spatial overlap.

Conclusions:

• The DPGMM-based registration algorithm offers an efficient and robust solution for pre- and intra-operative pointcloud alignment.
• This method facilitates the development of more reliable AR systems for surgical guidance.
• The algorithm's ability to infer the optimal number of mixture components enhances its adaptability to diverse datasets.