Persistent Homology in Sparse Regression and Its Application to Brain Morphometry

Moo K. Chung, Jamie L. Hanson, Jieping Ye, Richard J. Davidson, Seth D. Pollak

Introduction

In neuroimaging, mapping structural differences between groups often relies on Tensor-Based Morphometry (TBM). However, these studies frequently encounter the "small-n, large-p" problem, where the number of voxels (p) vastly exceeds the number of subjects (n). To address this, researchers employ sparse models that penalize complex connections to highlight the most significant structural relationships. This paper introduces a novel framework that integrates persistent homology—a tool from topological data analysis—with sparse regression to analyze brain connectivity more robustly and efficiently.

The Challenge of Sparse Networks

Traditional sparse models depend heavily on a tuning parameter ($\lambda$) that dictates the network's sparsity. Selecting an optimal $\lambda$ is often arbitrary and computationally intensive, requiring iterative numerical optimization. If the parameter is poorly chosen, statistical results may become unstable or vanish. Furthermore, calculating these networks for large-scale brain data can take dozens of hours, limiting the feasibility of high-resolution analysis.

Methodology

Sparse Pearson Correlation and Soft-Thresholding

The authors propose a new sparse model specifically for Pearson correlations. A key contribution is Theorem 1, which proves that the optimal solution for this model can be obtained via a simple soft-thresholding rule. This analytical approach replaces hours of iterative optimization with a calculation that takes only seconds, representing a massive leap in computational efficiency.

Graph Filtration and Persistent Homology

Rather than selecting a single fixed $\lambda$, the framework considers the entire range of possible values. As $\lambda$ increases, the brain network becomes increasingly sparse as edges are removed. This sequence of nested graphs is known as a "filtration."

To characterize the evolving topology of these networks, the authors use Betti numbers, specifically $\beta_0$, which counts the number of connected components. By plotting $\beta_0$ against the filtration value, they generate a "Betti plot" that serves as a global topological signature of the brain's structural organization.

Illustration of graph filtration and the resulting Betti plot (β₀) as edges are removed based on a filtration value.

Clinical Application and Results

Dataset and Modalities

The method was validated using simulated data and a clinical dataset of 54 subjects: 23 children who experienced severe early-life stress (maltreatment) and 31 age-matched controls. The study utilized T1-weighted MRI to measure local tissue growth/atrophy and Diffusion Tensor Imaging (DTI) to assess white matter integrity (Fractional Anisotropy).

Anatomical Homogeneity in Maltreated Children

The Betti plots revealed a significant statistical difference ($p < 0.001$) between the two groups across both imaging modalities. A clear "horizontal gap" in the plots indicated that networks in maltreated children remain connected (lower $\beta_0$) at higher thresholds compared to controls.

This suggests a state of anatomical homogeneity: maltreated children showed higher correlations across white matter regions, implying a lack of specialized structural differentiation. This diffuse organization may be a hallmark of developmental trauma.

Robustness and Performance

The framework demonstrated exceptional robustness:

  • Resolution Independence: Results remained consistent when increasing the number of brain nodes from 548 to 1856.
  • Registration Tolerance: The findings held even with up to 4mm of intentional image registration error.
  • Computational Speed: The soft-thresholding discovery reduced computation time from approximately 56 hours to just a few seconds.

Conclusion and Future Work

By shifting the focus from static network "snapshots" to the persistence of topological features, this framework provides a more stable and efficient way to analyze brain connectivity. While the current study primarily utilizes the zeroth Betti number ($\beta_0$), future work may incorporate higher-order features like $\beta_1$ (cycles) or extend the analysis to node-level data to identify specific brain regions driving these topological differences.

Summary updated

BibTeX
@article{chung2015persistent,
  author = {Moo K. Chung and Jamie L. Hanson and Jieping Ye and Richard J. Davidson and Seth D. Pollak},
  title = {{Persistent Homology in Sparse Regression and Its Application to Brain Morphometry}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2015},
  url = {https://openreview.net/forum?id=gFIUOK6NZDZ},
}
Figure