Abstract
In neuroimaging, the brain is frequently modeled as a connectome—a graph where nodes represent anatomical regions and edges represent structural or functional connections. Analyzing the "natural frequencies" of these networks through harmonic waves (eigenvectors of the graph Laplacian) offers deep insights into neural signal propagation. However, standard group-level analyses typically employ Euclidean averaging, which violates the orthogonality constraints of these waves and introduces significant noise. This paper presents a manifold learning approach that treats harmonic waves as points on a Stiefel manifold. By optimizing for a common population mean using the Alternating Direction Method of Multipliers (ADMM), we establish a mathematically rigorous reference for cross-subject comparison. Our results on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset demonstrate that this manifold-aware approach enhances replicability and identifies specific frequency-based biomarkers associated with neurodegeneration, particularly within the Default Mode Network.
Introduction
The human brain's structural architecture dictates the pathways through which information flows. To quantify this, researchers utilize the graph Laplacian matrix, whose eigenvectors—termed harmonic waves—represent the fundamental vibration patterns of the brain's physical wiring. These waves are essential for understanding how biological signals, such as cortical thickness or amyloid-beta deposition, are distributed across the network.
A persistent challenge in population-based studies is the "irregularity" of individual brain graphs. When researchers attempt to find a "representative" set of harmonic waves for a group, they often resort to simple Euclidean averaging of individual matrices. This approach is fundamentally flawed because the space of orthogonal matrices is non-Euclidean. Averaging in a flat space leads to "mathematical drift," where the resulting mean no longer possesses the critical property of orthogonality, effectively destroying the physical validity of the harmonic waves.
Methodology
To address these limitations, we propose a framework that respects the underlying geometry of the data. Instead of viewing a set of $p$ harmonic waves for $n$ brain regions as a simple $n \times p$ matrix, we treat it as a single point on a Stiefel manifold $\mathcal{V}(n, p)$. The Stiefel manifold is defined as the set of all $n \times p$ matrices with orthonormal columns.
The core of our method involves finding the Fréchet mean—the point on the manifold that minimizes the sum of squared distances to all individual subjects' wave sets. We utilize an iterative optimization scheme based on the Alternating Direction Method of Multipliers (ADMM). This process alternates between two primary steps:
- Native Estimation: Refining individual harmonic waves to align more closely with the group mean while maintaining their fidelity to the subject's unique anatomy.
- Mean Update: Calculating the center of the manifold using the Weiszfeld algorithm. This involves projecting data onto a local tangent space (a flat approximation of the manifold), averaging there, and then mapping the result back to the manifold via an exponential map.
Once a common reference $\Psi$ is established, we calculate harmonic energy. This metric quantifies the alignment of a biological signal (e.g., cortical thickness) with specific network frequencies, allowing us to pinpoint which "vibration modes" are affected by disease.
Datasets and Evaluation
The framework was validated using both synthetic and clinical data. Synthetic tests involving 3D rotation matrices (quaternions) confirmed that manifold averaging preserves geometric constraints that Euclidean methods violate, which often produce invalid, non-orthogonal results.
For clinical validation, we utilized data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI):
- Training Set: 94 subjects (including Healthy Controls, Mild Cognitive Impairment, and Alzheimer’s patients) were used to learn the common harmonic waves.
- Testing Set: 50 Healthy Controls (CN) and 47 Alzheimer’s (AD) subjects were used to evaluate the diagnostic utility of the learned waves.
Key Results
Our manifold-based approach demonstrated superior replicability compared to traditional methods. In resampling experiments, the learned common waves remained significantly more consistent across different data subsets than those produced by "pseudo" manifold or Euclidean methods.
Clinically, the method revealed that Alzheimer’s disease significantly alters the brain's harmonic profile. Key findings include:
- Reduced Harmonic Energy in Cortical Thickness: The AD group showed significantly lower total energy ($15.9 \pm 4.6$) compared to the CN group ($19.6 \pm 4.4$), suggesting a loss of structural integrity across specific network frequencies.
- Increased Amyloid Energy: Conversely, amyloid deposition energy was higher in AD subjects ($4.41 \pm 1.85$ vs $3.19 \pm 1.40$), indicating a pathological "overloading" of certain network modes.
- DMN Localization: The boundaries (zero-crossings) of the most significant harmonic 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.
Discussion and Conclusion
By treating the brain's "physics" as a problem of manifold geometry rather than simple linear algebra, we can extract more precise biomarkers for neurodegeneration. This approach identifies subtle patterns of disruption that are often masked by the noise of Euclidean averaging.
While computationally more intensive, the use of generalized power iterations helps mitigate costs. Future work will focus on applying this spectral analysis to other "disconnection syndromes," such as schizophrenia and frontotemporal dementia, and developing data-driven methods for optimal spectral resolution selection.
Summary updated