A Deep Learning Model for Data-Driven Discovery of Functional Connectivity

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

This paper presents BrainGNN, an end-to-end deep learning framework designed to discover functional connectivity patterns in the brain while simultaneously performing disease classification. By moving away from fixed, hand-crafted connectivity metrics, the authors demonstrate a data-driven way to identify the specific brain regions and neural "highways" most relevant to disorders like schizophrenia.

TL;DR

BrainGNN is a novel architecture that learns to construct brain connectivity graphs directly from fMRI data using self-attention, rather than relying on predefined correlation formulas. It achieves state-of-the-art classification of schizophrenia while automatically identifying a sparse subset of "biomarker" brain regions that align with clinical literature.

What Problem Does This Paper Address?

In neuroimaging, Functional Connectivity (FC) describes how different brain regions interact over time. Traditionally, researchers calculate FC using simple cross-correlation (Pearson correlation) between regions of interest (ROIs).

However, this standard approach has two major flaws:

  1. Rigidity: It assumes the "formula" for connectivity is the same for everyone, ignoring inter-subject variations caused by disease.
  2. Noise: It treats all brain regions as equally important, even though a specific disorder might only affect a small subset of the brain.

The authors sought to create a model that doesn't just use a graph, but learns the graph structure and prunes away irrelevant data to find the most informative neural biomarkers.

Methods

The BrainGNN architecture consists of three distinct stages:

  1. CNN Encoder: The model takes raw fMRI time-series data from 116 brain regions (defined by the AAL atlas). A 1D Convolutional Neural Network (CNN) processes these sequences to create a compact "embedding" (a mathematical representation) for each region.
  2. Self-Attention Module: This is the core innovation. Instead of using a fixed correlation matrix, the model uses a multi-head self-attention mechanism to estimate the weights (edges) between regions. This allows the model to learn a directed graph where the importance of a connection is determined by its predictive value for the task.
  3. GNN with Top-K Pooling: A 6-layer Graph Neural Network (GNN) processes the learned graph. To improve interpretability, the authors use Top-K pooling layers. These layers act like a filter, discarding the least important regions and keeping only the "top" nodes. By the final layer, the model has narrowed the brain down to a sparse subset of critical regions.

Datasets

The study utilized the Function Biomedical Informatics Research Network (FBIRN) dataset:

  • Subjects: 311 total (151 Healthy Controls and 160 Schizophrenia patients).
  • Data Type: Resting-state fMRI.
  • Preprocessing: Standard pipeline using SPM12, including motion correction and mapping to the Montreal Neurological Institute (MNI) space.

Key Insights and Ideas

  • Sparsity as a Diagnostic Tool: One of the most striking findings was that the connectivity graphs learned for schizophrenia patients were significantly sparser than those for healthy controls. This suggests that the model identifies a "reduction" in functional pathways as a key feature of the disorder.
  • Directed vs. Undirected: Unlike standard Pearson correlation (which is undirected), BrainGNN learns directed edges. This captures the inherent asymmetry in how brain regions influence one another.
  • End-to-End Discovery: Because the graph construction and classification are trained together, the model discovers connectivity patterns that are specifically optimized to distinguish between health and disease, rather than just measuring general signal similarity.

Key Results

  • Classification Performance: BrainGNN achieved a mean AUC (Area Under the Curve) of 0.79. This is comparable to state-of-the-art classical methods like Support Vector Machines (SVM) but provides significantly more interpretability.
  • Region Selection: The model successfully reduced the 116 initial brain regions down to just 23 key regions (roughly 20% of the brain).
  • Clinical Validation: The regions selected by the model—including the cerebellum, temporal lobe, caudate, and Supplementary Motor Area (SMA)—are highly consistent with existing medical literature regarding the pathophysiology of schizophrenia.

Limitations and Future Work

  • Atlas Dependency: While the model learns the connections, it still relies on a predefined anatomical atlas (AAL) to define the initial "nodes."
  • Dataset Size: Like many deep learning studies in neuroimaging, the sample size (311 subjects) is relatively small. Larger datasets would be needed to ensure the learned graphs generalize across diverse populations.
  • Future Direction: The authors envision an extension of this model that works at the voxel-level. This would allow the AI to define its own "regions" from scratch, potentially discovering functional units that don't perfectly align with traditional anatomical boundaries.

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 functional magnetic resonance imaging (fMRI) correlation matrix. However, most of the work with the FC depends on the way the connectivity is computed, and it 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 that 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 that are 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},
}

Loading related content...

Figure