Learning Macroscopic Brain Connectomes via Group-Sparse Factorization

Farzane Aminmansour, Andrew Patterson, Lei Le, Yisu Peng, Daniel Mitchell, Franco Pestilli, Cesar F. Caiafa, Russell Greiner, Martha WhiteNeurIPS 2019 · OpenReview

Mapping the "wiring diagram" of the human brain—the macroscopic connectome—is one of the most daunting challenges in neuroscience. Traditionally, this process relies on expert-heavy, rule-based algorithms that interpret diffusion-weighted MRI (dMRI) data. However, these methods are notoriously sensitive to parameter tweaks, often producing vastly different results from the same scan.

In this work, researchers propose a shift from manual rules to unsupervised learning. By framing tractography as a group-sparse tensor factorization problem, they demonstrate a way to automatically extract biologically plausible brain structures directly from raw data.

TL;DR

The authors present a fully unsupervised framework that learns brain connectomes from dMRI data by optimizing a convex objective with a novel group-sparse regularizer. To handle the massive dimensionality of the problem, they developed a specialized "Orientation Greedy" screening algorithm that outperforms standard methods like Orthogonal Matching Pursuit in capturing the smooth, contiguous nature of neural fascicles.

What Problem Does This Paper Address?

To understand how the brain functions, we need to see the fascicles—bundles of axons that connect different brain regions. We "see" these through dMRI, which measures how water molecules diffuse in the brain; water moves more easily along the length of a fiber than across it.

Traditional tractography relies on rule-based models and stopping criteria, which can yield varying results.

The gap lies in tractography: the process of turning those diffusion measurements into a 3D map of fibers. Current "best practices" involve choosing a specific mathematical model and a set of heuristic rules (e.g., "stop if the turn is sharper than 30 degrees"). Because these rules are hand-tuned, different experts can get different results from the same brain. Supervised learning could help, but it requires "ground truth" labels, which we don't truly have for living human brains. This paper seeks a fully unsupervised, data-driven way to learn these structures.

Methods

The researchers build upon a tensor encoding called ENCODE, which represents the dMRI signal as a factorization of a sparse connectome tensor ($\Phi$) and a dictionary of diffusion signals ($D$).

The ENCODE tensor representation and the proposed block regularization strategy.

1. The Group-Sparse Regularizer

A key contribution is the design of a regularizer that encodes biological "common sense." Brain fibers aren't random; they are contiguous (they don't just disappear) and smooth (they don't make 90-degree jagged turns). The authors designed a regularizer that groups neighboring voxels and similar orientations, encouraging the model to activate "blocks" of related parameters. This ensures that the resulting connectome consists of smooth, flowing paths rather than scattered noise.

2. The "Orientation Greedy" Algorithm

The optimization problem is gargantuan—billions of potential parameters. To make it tractable, the authors use a "screening" step to pick the most likely fiber orientations in each voxel before running the full optimization.

  • The Innovation: Standard greedy algorithms like Orthogonal Matching Pursuit (OMP) try to pick orientations that are as different as possible to reduce error quickly.
  • The Fix: In the brain, fibers often overlap in similar directions. The authors' "Orientation Greedy" criterion uses a submodular objective that balances reconstruction accuracy with the individual utility of each orientation, allowing the model to pick the "similar" directions actually present in the tissue.

Datasets

The authors evaluated their method using data generated from an expert connectome solution (focusing on the Arcuate Fasciculus and SLF1 structures). By using a "ground truth" $\Phi$ generated by an expert model, they could objectively measure how well their unsupervised learning algorithm recovered the known underlying structure.

Key Insights and Ideas

  • Convexity is King: The authors prove their objective is convex and has a unique solution. This is a big deal for reproducibility; unlike deep learning models that might get stuck in different local minima, this method ensures an identifiable connectome for an individual.
  • OMP is the Wrong Tool: The "aha moment" regarding screening is that OMP’s preference for orthogonality is actually a hindrance in neuroscience, where biological structures are inherently correlated and redundant.
  • Biological Plausibility via Math: By grouping spatial and directional parameters, the math naturally "prefers" the long, winding structures we see in real brains, effectively replacing manual "stopping criteria" with a global optimization constraint.

Key Results

  • Superior Screening: The proposed greedy algorithm significantly reduced the "angular distance" to the ground truth compared to OMP. It picked the right directions more often and with higher precision.
  • Lower Bias: In optimization tests, the method reached a much lower reconstruction error several orders of magnitude faster than OMP-based initialization.
  • Visual Fidelity: Qualitatively, the recovered fascicles were smooth and contiguous. In some cases, the learned model produced even more plausible-looking fibers than the "ground truth" tractography, which sometimes contained jagged, unlikely artifacts.

Limitations and Future Work

While a major step forward, the paper notes a few areas for growth:

  • Fixed Assignments: The current experiments focused on evaluating orientation selection while keeping the assignment of fascicles to voxels relatively fixed.
  • Complexity of Real Data: The method was tested on specific major tracts. Scaling this to a full-brain connectome with hundreds of crossing fiber populations remains a significant computational challenge.
  • Future Directions: The authors suggest incorporating even more complex regularizers and testing the framework across diverse, real-world clinical datasets to see how it handles noise and pathology.

Summary updated

Abstract

Mapping structural brain connectomes for living human brains typically requires expert analysis and rule-based models on diffusion-weighted magnetic resonance imaging. A data-driven approach, however, could overcome limitations in such rule-based approaches and improve precision mappings for individuals. In this work, we explore a framework that facilitates applying learning algorithms to automatically extract brain connectomes. Using a tensor encoding, we design an objective with a group-regularizer that prefers biologically plausible fascicle structure. We show that the objective is convex and has unique solutions, ensuring identifiable connectomes for an individual. We develop an efficient optimization strategy for this extremely high-dimensional sparse problem, by reducing the number of parameters using a greedy algorithm designed specifically for the problem. We show that this greedy algorithm significantly improves on a standard greedy algorithm, called Orthogonal Matching Pursuit. We conclude with an analysis of the solutions found by our method, showing we can accurately reconstruct the diffusion information while maintaining contiguous fascicles with smooth direction changes.
BibTeX
@inproceedings{aminmansour2019learning,
  author = {Farzane Aminmansour and Andrew Patterson and Lei Le and Yisu Peng and Daniel Mitchell and Franco Pestilli and Cesar F. Caiafa and Russell Greiner and Martha White},
  title = {{Learning Macroscopic Brain Connectomes via Group-Sparse Factorization}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2019},
  url = {https://openreview.net/forum?id=A88hOFkuMW},
}
Figure