Learning Common Harmonic Waves on Stiefel Manifold - A New Mathematical Approach for Brain Network Analyses

Jiazhou Chen, Guoqiang Han, Hongmin Cai, Defu Yang, Paul J. Laurienti, Martin Styner, Guorong Wu

TL;DR

Researchers have developed a novel mathematical framework that treats brain network structures as points on a Stiefel manifold rather than simple data arrays. By using manifold optimization to find a "common" set of harmonic waves across individuals, this approach provides a more accurate and reproducible way to detect how neurodegenerative diseases like Alzheimer’s disrupt large-scale brain communication.

What Problem Does This Paper Address?

In neuroimaging, we often represent the brain as a "connectome"—a complex graph where nodes are brain regions and edges are the physical connections between them. To understand how signals (or pathologies) spread through this network, researchers use harmonic waves, which are the eigenvectors of the network's Laplacian matrix. Think of these as the "natural frequencies" or vibration patterns of the brain's architecture.

The challenge arises when we want to compare these patterns across a group. Standard methods typically use Euclidean averaging—simply adding up matrices and dividing by the number of subjects. However, brain networks are "irregular" data structures. Euclidean averaging destroys the very properties that make harmonic waves useful, such as their orthogonality (the property that each wave represents a unique, independent pattern). This "mathematical drift" leads to noisy results and a lack of a reliable reference point for clinical diagnosis.

Methods

The authors propose a manifold learning method to discover the unbiased population mean of individual brain Eigen-systems.

  1. The Stiefel Manifold: Instead of treating a set of harmonic waves as a flat matrix, they treat it as a single point on a Stiefel manifold—a high-dimensional space specifically defined for sets of ordered orthonormal vectors.
  2. Manifold Optimization: They use an iterative scheme based on the Alternating Direction Method of Multipliers (ADMM) to solve two sub-problems:
  3. Native Estimation: Adjusting an individual's harmonic waves so they are "attracted" toward the group mean while still respecting their own unique anatomy.
  4. Mean Update: Finding the Fréchet mean (the "center" of the manifold) using the Weiszfeld algorithm. This involves projecting data onto a "tangent space" (a flat approximation of the manifold), averaging there, and mapping it back.
  5. Harmonic Energy: Once they have a common reference, they calculate "harmonic energy"—a measure of how much a biological signal (like cortical thickness) aligns with specific network frequencies.

Datasets

The study utilized two primary data sources:

  • Synthetic Data: 3D rotation matrices represented as quaternions were used to prove that manifold averaging stays on the "surface" of the data's geometry, while Euclidean averaging falls "inside" (producing invalid, non-orthogonal results).
  • ADNI (Alzheimer’s Disease Neuroimaging Initiative):
    • Training Set: 94 subjects (including Healthy Controls, Mild Cognitive Impairment, and Alzheimer’s patients) to learn the common harmonic waves.
  • Testing Set: 50 Healthy Controls (CN) and 47 Alzheimer’s (AD) subjects to evaluate the diagnostic power of the learned waves.

Key Insights and Ideas

  • Respecting Geometry Matters: The "Aha!" moment is that the brain's network topology isn't just a collection of numbers; it has a rigid geometric structure. By using manifold algebra, the researchers ensure that the "average" brain they calculate is still a mathematically valid brain network.
  • Frequency-Based Biomarkers: The paper shows that Alzheimer’s isn't just about "losing brain matter" in random spots. Instead, the disease disrupts specific "frequencies" of the brain network.
  • The DMN Connection: They found that the "zero-crossings" (the boundaries of the harmonic waves) in significant waves were heavily concentrated in the Default Mode Network (DMN). This provides a mathematical link between the brain's physical "wiring" and the functional networks known to fail in Alzheimer’s.

Key Results

  • Superior Replicability: In resampling tests, the Stiefel manifold approach produced significantly more consistent results across different subsets of data compared to "pseudo" manifold methods.
  • Energy Differences: The AD group showed significantly lower total harmonic energy in cortical thickness ($15.9 \pm 4.6$ vs $19.6 \pm 4.4$) and significantly higher energy in amyloid deposition ($4.41 \pm 1.85$ vs $3.19 \pm 1.40$).
  • Sensitivity: The manifold approach discovered 16 significant harmonic waves for cortical thickness and 15 for amyloid burden, identifying patterns that Euclidean methods missed.

Limitations and Future Work

While the results are promising for Alzheimer’s, the authors acknowledge that this is a first step.

  • Computational Cost: Manifold optimization is more computationally intensive than simple averaging, though the authors used a "generalized power iteration" to mitigate this.
  • Generalization: The current study focused on Alzheimer’s. The authors plan to apply this framework to other "disconnection syndromes" like frontotemporal dementia and schizophrenia to see if they exhibit unique "frequency signatures."
  • Dimension Selection: Choosing the number of harmonic waves ($p$) is currently done empirically; a more automated, data-driven way to select the optimal "spectral resolution" would be a valuable improvement.

Summary updated

Abstract

Converging evidence shows that disease-relevant brain alterations do not appear in random brain locations, instead, their spatial patterns follow large-scale brain networks. In this context, a powerful network analysis approach with a mathematical foundation is indispensable to understand the mechanisms of neuropathological events as they spread through the brain. Indeed, the topology of each brain network is governed by its native harmonic waves, which are a set of orthogonal bases derived from the Eigen-system of the underlying Laplacian matrix. To that end, we propose a novel connectome harmonic analysis framework that provides enhanced mathematical insights by detecting frequency-based alterations relevant to brain disorders. The backbone of our framework is a novel manifold algebra appropriate for inference across harmonic waves. This algebra overcomes the limitations of using classic Euclidean operations on irregular data structures. The individual harmonic differences are measured by a set of common harmonic waves learned from a population of individual Eigen-systems, where each native Eigen-system is regarded as a sample drawn from the Stiefel manifold. Specifically, a manifold optimization scheme is tailored to find the common harmonic waves, which reside at the center of the Stiefel manifold. To that end, the common harmonic waves constitute a new set of neurobiological bases to understand disease progression. Each harmonic wave exhibits a unique propagation pattern of neuropathological burden spreading across brain networks. The statistical power of our novel connectome harmonic analysis approach is evaluated by identifying frequency-based alterations relevant to Alzheimer's disease, where our learning-based manifold approach discovers more significant and reproducible network dysfunction patterns than Euclidean methods.
BibTeX
@article{chen2021learning,
  author = {Jiazhou Chen and Guoqiang Han and Hongmin Cai and Defu Yang and Paul J. Laurienti and Martin Styner and Guorong Wu},
  title = {{Learning Common Harmonic Waves on Stiefel Manifold - A New Mathematical Approach for Brain Network Analyses}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2021},
  url = {https://openreview.net/forum?id=AagYRHkoZA0},
}
Figure