2009 Winter Project Week Tractography Segmentation

Return to Project Week Main Page

Key Investigators

• Xiaogang Wang, MIT
• Carl-Fredrik Westin, BWH

Objective

We are developing a nonparametric Bayesian approach to cluster white matter fiber tracts to bundles. These bundles correspond to white matter anatomy. This approach is used to cluster fibers across multiple subjects and compare their anatomical structures. The number of clusters is learnt from data instead of being manually specified.

Approach

Our approach models bundles as multinomial distributions over voxels and orientations. Under a hierarchical Bayesian model, Dirichlet process is used as a prior distribution to learn the number of clusters. Gibbs sampling is used for inference.

Our plan is to integrate this approach into 3D-slicer, use 3D-slicer to do experimental evaluation (including registering multiple subjects, generating fibers and clustering fibers), and compare with approach proposed by O'Donnell and Westin, whose use the spectral clustering and the mean of closest distances.

Progress

We have the code of this approach written in C++ and matlab. In this project week, we need to figure out how to integrate our code into 3D-slicer, how to use 3D-slicer to run registration, tractography and O' Donnell's approach.

References

[1] Brun A, Knutsson H, Park HJ, Shenton ME, Westin CF. Clustering fiber tracts using normalized cuts. In Seventh International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI'04), Lecture Notes in Computer Science. Rennes - Saint Malo, France, 2004;368-375.

[2] O'Donnell LJ, Westin CF. [http://lmi.bwh.harvard.edu/papers/papers/odonnellTMI07.html Automatic tractography segmentation using a high-dimensional white matter atlas.} IEEE Transactions on Medical Imaging 2007;26(11):1562-1575.