Robust clustering of massive tractography datasets

Pamela Guevara, Cyril Poupon, Denis Rivière, Yann Cointepas, Maxime Descoteaux, Bertrand Thirion, Jean-François Mangin

TL;DR

This paper introduces a novel, robust hierarchical clustering method designed to efficiently organize the massive datasets generated by diffusion MRI tractography. By effectively compressing millions of individual fiber tracts into a manageable number of bundles, this approach paves the way for more sophisticated large-scale analyses of brain white matter structure and its variations across individuals.

What Problem Does This Paper Address?

The human brain's white matter is a complex network of nerve fibers that transmit signals between different brain regions. Diffusion MRI tractography is a powerful technique that allows us to visualize these pathways in vivo. However, tractography algorithms often generate millions of individual fiber trajectories, creating a "spaghetti plate" of data that is computationally challenging to analyze. Existing methods struggle with this scale, often requiring manual intervention or being limited to smaller datasets. This paper tackles the problem of efficiently and robustly clustering these massive tractography datasets into meaningful "fiber bundles" that represent coherent white matter pathways, enabling further analysis and comparison across individuals.

Methods

The authors propose a multi-step hierarchical clustering approach that breaks down the complex problem into manageable stages:

  • Hierarchical Decomposition: The entire dataset is first split into four main subsets: left hemisphere, right hemisphere, inter-hemispheric, and cerebellum tracts. This initial step helps manage the complexity by processing smaller, more localized regions.
  • Length-Based Segmentation: Within each subset, tracts are further grouped based on their length. This is a crucial step for robustness, as it helps to separate overlapping bundles that might otherwise be incorrectly merged.
  • Voxel-Based Clustering: This is a key innovation. Instead of directly clustering millions of fibers, the method clusters white matter voxels based on their connectivity. This is achieved by creating parcels (small groups of voxels) and then using a hierarchical clustering algorithm on these parcels, leveraging the number of tracts passing through them. This significantly reduces the computational load.
  • Extremity-Based Clustering: Once fiber clusters are identified, this step focuses on the ends of the fibers (extremities) to further subdivide them into more regular bundles called "fascicles." This is done using a watershed algorithm on the density of fiber extremities.
  • Fascicle Merge: Finally, a centroid tract is computed for each fascicle, and these centroids are clustered to merge fascicles that have similar geometries. This step aims to overcome potential over-segmentation from previous steps.

The method is designed to be robust and scalable, handling millions of tracts without requiring extensive parameter tuning for most steps.

Datasets

The method was validated using several datasets:

  • Adult HARDI datasets: Data from twelve adult subjects from the NMR public database, acquired with high angular resolution diffusion imaging (HARDI).
  • Child DTI datasets: Data from two children, acquired using diffusion tensor imaging (DTI).
  • Phantom datasets: A physical MR phantom designed to simulate realistic crossing, kissing, splitting, and bending fiber configurations, serving as a ground truth.

Key Insights and Ideas

  • Shifting from Fiber to Voxel Clustering: The most significant insight is the move from clustering individual fibers to clustering white matter voxels. This dramatically improves scalability and computational efficiency, allowing the analysis of massive datasets.
  • Hierarchical Approach for Robustness: The multi-step hierarchical decomposition is crucial for handling the inherent complexity and potential ambiguities in tractography data, leading to more reliable bundle identification.
  • Length as a Primary Separator: Grouping fibers by length early in the process is a surprisingly effective heuristic for separating overlapping bundles, contributing significantly to the method's robustness.
  • Extremities as Discriminators: Focusing on fiber extremities in a later stage is a clever way to refine bundle segmentation, as the ends of fibers often provide clearer distinctions between pathways than their entire trajectories.

Key Results

  • The method successfully compressed datasets of over a million fiber tracts into approximately two thousand fiber bundles.
  • Validations using simulated data and a physical phantom demonstrated the method's robustness and accuracy in identifying bundles, even in the presence of noise and complex fiber configurations.
  • The method showed good scalability, with computational time significantly lower than brute-force approaches. For instance, a brute-force clustering took over 6 hours, while their method completed in 15 minutes for a dataset of ~24,000 fibers.
  • The method was able to identify subdivisions of white matter pathways and potential U-fiber bundles, suggesting its utility for building detailed brain atlases.

Limitations and Future Work

  • Dependence on Tractography Quality: The method's results are inherently dependent on the quality of the initial tractography data. Bundles that are not tracked will not be detected.
  • Spurious Bundles: While the method aims to filter out spurious bundles, some may still persist, especially in cases of similar configurations of bundle crossings across subjects or errors in the propagation mask.
  • Future Work: The authors envision their method as a crucial preprocessing step for more sophisticated analyses. They highlight its potential for building brain fiber bundle models from inter-subject analysis, inferring subdivisions of major white matter pathways, and improving the understanding of folding variability and its relation to fiber bundles. They also suggest that automatic inference of such models could be performed by clustering bundle centroids at the population level.

Summary updated

Abstract

This paper presents a clustering method that detects the fiber bundles embedded in any MR-diffusion based tractography dataset. Our method can be seen as a compressing operation, capturing the most meaningful information enclosed in the fiber dataset. For the sake of efficiency, part of the analysis is based on clustering the white matter (WM) voxels rather than the fibers. The resulting regions of interest are used to define subset of fibers that are subdivided further into consistent bundles using a clustering of the fiber extremities. The dataset is reduced from more than one million fiber tracts to about two thousand fiber bundles. Validations are provided using simulated data and a physical phantom. We see our approach as a crucial preprocessing step before further analysis of huge fiber datasets. An important application will be the inference of detailed models of the subdivisions of white matter pathways and the mapping of the main U-fiber bundles. Research Highlights ►We propose a robust fiber clustering method inferring the main fiber bundles present in any diffusion-based tract dataset. ►The method can be seen as a compressing operation, capturing the most meaningful information enclosed in the fiber dataset. ►Is the first method capable to handle a very big number of fibers. ►Results can be used for the inference of a brain fiber bundle model from an intersubject analysis. Previous article in issue Next article in issue
BibTeX
@article{guevara2011robust,
  author = {Pamela Guevara and Cyril Poupon and Denis Rivière and Yann Cointepas and Maxime Descoteaux and Bertrand Thirion and Jean-François Mangin},
  title = {{Robust clustering of massive tractography datasets}},
  journal = {NeuroImage},
  year = {2011},
  doi = {10.1016/j.neuroimage.2010.10.028},
}
Figure