Abstract
Functional brain connectivity analysis often suffers from alignment errors and high dimensionality. This paper presents a distributional representation that captures voxel-level interactions through joint and conditional distributions of correlation and spatial distance. This coordinate-free approach enables robust outlier detection and effective classification of ADHD subtypes, outperforming traditional alignment-dependent methods.
Introduction
The analysis of functional brain connectivity typically involves mapping 4D fMRI data onto a common spatial template. This process assumes that voxels at the same coordinates across different subjects represent the same functional units. In practice, perfect alignment is unattainable, and the resulting "alignment errors" can introduce significant noise. Furthermore, the high dimensionality of voxel-level data often forces researchers to aggregate data into coarse regions of interest (ROIs), potentially losing critical fine-grained information.
To address these challenges, we propose a distributional representation that bypasses absolute coordinates. By focusing on the relative relationships—specifically the correlation and Euclidean distance between all pairs of voxels—we can create a low-dimensional summary of a brain's functional architecture that is invariant to spatial shifts and rotations.
Methodology: From Joint to Conditional Distributions
The core of our approach is the Correlation-Distance Joint Distribution. For every pair of voxels in a scan, we calculate the Pearson correlation coefficient of their time courses and the Euclidean distance between their spatial locations. A 2D histogram of these values serves as a global representation of the brain's connectivity.
However, the spatial distribution of voxels in the brain is non-uniform; most voxel pairs are separated by intermediate distances. This creates a natural bias in the joint distribution. To isolate the functional signal from this spatial artifact, we employ a Conditional Distribution—the probability of a correlation value given a specific distance.
We define group "centroids" using Accumulative Conditional Histograms (ACH), which average the distributions across a cohort. Individual deviations from these centroids are then measured using Kullback-Leibler (KL) divergence, providing a metric for classification and outlier detection.
Visual Outlier Detection
The 2D distributional representation acts as a unique "fingerprint" for each scan. This allows for intuitive visual quality control. While standard preprocessing pipelines may miss subtle data artifacts, anomalies often manifest as distinct deformations in the joint distribution shape.
Results and Discussion
We evaluated our method on the NYU site sample of the ADHD-200 dataset, comprising 216 subjects. The goal was to classify subjects as "Typically Developing" or "ADHD."
The ACH (Accumulative Conditional Histogram) method emerged as the superior representation, achieving an accuracy of 86% for the typical group and 72% for the ADHD group. This outperformed the joint distribution (AJH) and Gaussian-fitting (EGD/ECD) methods, confirming that conditioning on distance is crucial for removing spatial bias.
Interestingly, analysis of KL divergence revealed an "orbit" phenomenon: individual brains do not cluster at a single point but rather distribute in a ring-like fashion around the group centroid. This suggests that "normal" brain connectivity encompasses a range of variations that are equidistant from the categorical average in the distributional space.
Conclusion
Distributional representations offer a robust, task-free alternative to traditional alignment-based fMRI analysis. By capturing the statistical properties of voxel interactions, we can detect outliers and classify clinical populations without the need for a common coordinate space. Future work will explore more sophisticated distribution-fitting techniques and the application of embedding methods from natural language processing to further refine these brain "fingerprints."
Summary updated