Tractography segmentation using a hierarchical Dirichlet processes mixture model.

Area of Science:

• Neuroscience
• Computational Biology
• Medical Imaging

Background:

• White matter fiber tract clustering is crucial for understanding brain connectivity.
• Existing methods often require manual specification of the number of clusters or computationally intensive pairwise comparisons.

Purpose of the Study:

• To develop a new nonparametric Bayesian framework for unsupervised clustering of white matter fiber tracts.
• To enable automatic learning of bundle structures and facilitate cross-subject comparisons.

Main Methods:

• Utilized a hierarchical Dirichlet processes mixture (HDPM) model for nonparametric Bayesian clustering.
• Employed a Dirichlet process (DP) prior for automatic determination of the number of clusters.
• Developed a method that does not require pairwise fiber distance computations.

Main Results:

• Successfully clustered white matter fiber tracts into bundles without supervision.
• Demonstrated the ability to use learned bundle models as priors for new subject data.
• Showcased the creation of new clusters for previously unobserved structures.
• Validated the approach on datasets with up to 120,000 fibers, highlighting scalability.

Conclusions:

• The proposed HDPM framework offers an effective and scalable solution for white matter tract clustering.
• This approach facilitates unsupervised learning and robust cross-subject analysis of brain connectomes.
• The method's ability to adapt to new data and structures enhances its utility in neuroscience research.