Abstract
The integration of multi-site datasets in neuroimaging is essential for increasing statistical power, yet it introduces significant variability—known as "site effects"—that stems from differences in scanner hardware and protocols rather than biological factors. This paper demonstrates that functional connectivity (FC) matrices derived from resting-state fMRI contain substantial imaging-site bias. To address this, we leverage the mathematical property that FC matrices reside on the manifold of symmetric positive-definite (SPD) matrices. We propose Rigid Log-Euclidean Translation (RLET), a geometry-aware harmonization approach that respects the underlying Riemannian structure of the data. When applied to the ADHD-200 dataset, RLET successfully removed site-related clustering and enhanced the detection of clinical anomalies, aligning results with established ADHD pathophysiology. This work suggests that geometric harmonization is a critical step for robust multi-site functional neuroimaging analysis.
Introduction
In neuroimaging research, pooling data from multiple centers is a common strategy to achieve the large sample sizes required for reproducible science. However, the "site effect" remains a primary obstacle. Variations in scanner manufacturers, magnetic field strengths, and acquisition sequences introduce non-biological noise that can confound clinical findings. If left uncorrected, machine learning models may learn to distinguish sites rather than disease states.
Traditional harmonization methods, such as ComBat, typically treat connectivity data as standard Euclidean vectors. However, functional connectivity matrices—which represent the correlation of activity between brain regions—are Symmetric Positive-Definite (SPD). These matrices do not exist in a flat, Euclidean space but rather on a curved mathematical manifold. Applying standard linear operations (like subtraction or addition) to these matrices can result in mathematically impossible structures, such as matrices with negative eigenvalues. This study introduces a framework to harmonize data while strictly adhering to the geometric constraints of the SPD manifold.
Methodology: Rigid Log-Euclidean Translation (RLET)
To preserve the mathematical integrity of the connectivity data, the authors utilized Riemannian geometry. The proposed method, Rigid Log-Euclidean Translation (RLET), operates through a multi-step process:
- The Log-Euclidean Framework: The researchers employed the Log-Euclidean Riemannian Metric (LERM) to map the curved SPD matrices into a "flat" tangent space using the matrix logarithm. In this tangent space, standard linear algebra can be performed without violating the manifold's constraints.
- Mean Alignment: For each imaging site, the "average" connectivity (the Fréchet mean) was calculated. RLET then translates the entire cluster of matrices from a specific site so that their local mean aligns with the global mean of the entire multi-site dataset.
- Manifold Projection: Once translated, the matrices are projected back onto the SPD manifold using the matrix exponential.
Unlike more aggressive methods like Matrix Whitening—which moves all site means to the Identity matrix—RLET is a "rigid" translation. It shifts the data to a common center without altering the relative distances between individual subjects within a site. This ensures that the biological variability (the signal) is preserved while the site-specific bias (the noise) is mitigated.
Experimental Design and Results
The framework was validated using a subsample of the ADHD-200 dataset, comprising 160 subjects (80 ADHD patients and 80 healthy controls) across four sites: Kennedy Krieger Institute (KKI), NYU Child Study Center, Peking University (PKG), and the NeuroIMAGE sample (NIM). Data were preprocessed using the Athena pipeline and the CC200 parcellation, resulting in 190x190 functional connectivity matrices.
Removal of Site Bias
Prior to harmonization, low-dimensional visualizations (t-SNE) revealed that subjects clustered primarily by their imaging site rather than their clinical diagnosis. Following the application of RLET, these site-specific clusters disappeared, indicating that the geometric translation successfully neutralized the scanner-induced variance.
Enhanced Clinical Sensitivity
The impact of RLET was most evident in the statistical analysis of ADHD-related anomalies. Using a geometry-aware permutation testing framework, the researchers compared the connectivity of ADHD patients against healthy controls:
- Non-harmonized data: Identified only 21 significant connectivity anomalies.
- RLET-harmonized data: Identified 33 significant anomalies.
Crucially, the anomalies identified after RLET harmonization showed a remarkable correlation with known ADHD pathophysiology. The analysis highlighted disruptions in frontostriatal pathways, fronto-cerebellar circuits, and the dorsal stream of the visual network. The non-harmonized analysis failed to detect the cerebellar and posterior parietal involvement, both of which are well-documented in ADHD literature. This suggests that site effects in the raw data were actively masking true biological signals.
Discussion and Future Directions
While RLET represents a significant advancement in neuroimaging harmonization, it is primarily a "first-order" correction. It aligns the centers (means) of the data but does not yet account for differences in the "spread" or dispersion of data across different sites. The authors suggest that future iterations of the method could include a site-dependent rescaling parameter to normalize this dispersion.
In conclusion, this study provides strong evidence that functional connectivity harmonization must respect the geometric properties of the SPD manifold. By using Rigid Log-Euclidean Translation, researchers can pool multi-site data more effectively, uncovering clinical insights that might otherwise remain hidden beneath the noise of multi-site variability.
Summary updated