A deep learning model for data-driven discovery of functional connectivity

Usman Mahmood, Zening Fu, Vince D. Calhoun, Sergey M. Plis

Abstract

Functional connectivity (FC) studies have demonstrated the overarching value of studying the brain and its disorders through the undirected weighted graph of fMRI correlation matrices. Most of the work with the FC, however, depends on the way the connectivity is computed and further depends on the manual post-hoc analysis of the FC matrices. In this work, we propose a deep learning architecture, BrainGNN, that learns the connectivity structure as part of learning to classify subjects. It simultaneously applies a graphical neural network to this learned graph and learns to select a sparse subset of brain regions important to the prediction task. We demonstrate the model's state-of-the-art classification performance on a schizophrenia fMRI dataset and demonstrate how introspection leads to disorder-relevant findings. The graphs learned by the model exhibit strong class discrimination, and the sparse subset of relevant regions are consistent with the schizophrenia literature.

Introduction

In neuroimaging, Functional Connectivity (FC) describes the temporal interactions between distinct brain regions. Traditionally, researchers calculate FC using simple cross-correlation (Pearson correlation) between regions of interest (ROIs). While useful, this standard approach assumes a rigid, universal formula for connectivity that ignores inter-subject variations and often fails to distinguish between task-relevant regions and background noise.

BrainGNN addresses these challenges by moving away from hand-crafted metrics. Instead, it employs an end-to-end framework to discover functional connectivity patterns while performing disease classification, identifying the specific neural "highways" most relevant to disorders such as schizophrenia.

Methodology: The BrainGNN Architecture

The BrainGNN framework consists of three integrated stages designed to transform raw fMRI data into interpretable diagnostic insights:

  1. CNN Encoder: Raw fMRI time-series data from 116 brain regions (defined by the AAL atlas) are processed by a 1D Convolutional Neural Network (CNN). This stage generates a compact embedding for each region, capturing its unique temporal signature.
  2. Self-Attention Module: Unlike traditional methods that use fixed correlation matrices, BrainGNN utilizes a multi-head self-attention mechanism to estimate connection weights between regions. This allows the model to learn a directed graph where edge importance is optimized for the specific predictive task.
  3. GNN with Top-K Pooling: A 6-layer Graph Neural Network (GNN) processes the learned graph. To enhance interpretability and focus on biomarkers, the architecture incorporates Top-K pooling layers. These layers prune the graph, discarding less informative nodes and narrowing the focus to a sparse subset of critical regions.
The BrainGNN architecture overview, illustrating the pipeline from raw fMRI signals to binary classification.

Experimental Results

The model was evaluated using the Function Biomedical Informatics Research Network (FBIRN) dataset, comprising resting-state fMRI data from 311 subjects (151 healthy controls and 160 schizophrenia patients).

Classification Performance

BrainGNN achieved a mean Area Under the Curve (AUC) of 0.79. While this performance is comparable to classical methods like Support Vector Machines (SVM), BrainGNN offers the distinct advantage of automated feature discovery and significantly higher interpretability.

Receiver Operating Characteristic (ROC) curve for BrainGNN, showing a mean AUC of 0.79 ± 0.11 across cross-validation folds.

Biomarker Discovery

A key strength of BrainGNN is its ability to identify relevant brain regions through its pooling mechanism. The model successfully reduced the initial 116 regions to a sparse set of 23 key regions (approximately 20% of the brain). These selected regions—including the cerebellum, temporal lobe, caudate, and Supplementary Motor Area (SMA)—align closely with established medical literature on the pathophysiology of schizophrenia.

Discussion and Insights

The study revealed that connectivity graphs for schizophrenia patients were significantly sparser than those of healthy controls, suggesting that a reduction in functional pathways is a hallmark of the disorder. Furthermore, by learning directed edges, BrainGNN captures the asymmetric influence regions exert on one another, a nuance lost in standard undirected correlation analysis.

Despite its successes, the model currently relies on predefined anatomical atlases. Future research aims to extend this architecture to the voxel level, potentially allowing the model to define its own functional units without being constrained by traditional anatomical boundaries.

Conclusion

BrainGNN represents a significant step toward data-driven neuroimaging analysis. By integrating graph construction, region selection, and classification into a single differentiable pipeline, it provides a powerful tool for both diagnosing psychiatric disorders and discovering the underlying neural biomarkers that define them.

Summary updated

Abstract

Functional connectivity (FC) studies have demonstrated the overarching value of studying the brain and its disorders through the undirected weighted graph of fMRI correlation matrix. Most of the work with the FC, however, depends on the way the connectivity is computed, and further depends on the manual post-hoc analysis of the FC matrices. In this work we propose a deep learning architecture BrainGNN that learns the connectivity structure as part of learning to classify subjects. It simultaneously applies a graphical neural network to this learned graph and learns to select a sparse subset of brain regions important to the prediction task. We demonstrate the model's state-of-the-art classification performance on a schizophrenia fMRI dataset and demonstrate how introspection leads to disorder relevant findings. The graphs learned by the model exhibit strong class discrimination and the sparse subset of relevant regions are consistent with the schizophrenia literature.
BibTeX
@article{mahmood2021deep,
  author = {Usman Mahmood and Zening Fu and Vince D. Calhoun and Sergey M. Plis},
  title = {{A deep learning model for data-driven discovery of functional connectivity}},
  journal = {Algorithms},
  year = {2021},
  doi = {10.3390/a14030075},
}
Figure