[Paper Review] FBNETGEN: Task-aware GNN-based fMRI Analysis via Functional Brain Network Generation

Xuan Kan, Hejie Cui+3 more|PubMed|May 25, 2022

Functional Brain Connectivity Studies42 references29 citations

TL;DR

FBNetGen introduces a task-aware, end-to-end framework that learns to generate brain networks from fMRI time series and uses a GNN for downstream prediction, achieving state-of-the-art gender prediction on ABCD and PNC with interpretable, task-focused graphs.

ABSTRACT

Functional magnetic resonance imaging (fMRI) is one of the most common imaging modalities to investigate brain functions. Recent studies in neuroscience stress the great potential of functional brain networks constructed from fMRI data for clinical predictions. Traditional functional brain networks, however, are noisy and unaware of downstream prediction tasks, while also incompatible with the deep graph neural network (GNN) models. In order to fully unleash the power of GNNs in network-based fMRI analysis, we develop FBNETGEN, a task-aware and interpretable fMRI analysis framework via deep brain network generation. In particular, we formulate (1) prominent region of interest (ROI) features extraction, (2) brain networks generation, and (3) clinical predictions with GNNs, in an end-to-end trainable model under the guidance of particular prediction tasks. Along with the process, the key novel component is the graph generator which learns to transform raw time-series features into task-oriented brain networks. Our learnable graphs also provide unique interpretations by highlighting prediction-related brain regions. Comprehensive experiments on two datasets, i.e., the recently released and currently largest publicly available fMRI dataset Adolescent Brain Cognitive Development (ABCD), and the widely-used fMRI dataset PNC, prove the superior effectiveness and interpretability of FBNETGEN. The implementation is available at https://github.com/Wayfear/FBNETGEN.

Motivation & Objective

• Motivate and enable end-to-end analysis of fMRI data with task-aware brain network generation.
• Generate positive, task-specific functional brain graphs compatible with GNNs for improved prediction.
• Provide interpretable brain connectivity patterns that highlight prediction-related regions.
• Demonstrate superior predictive performance over time-series baselines and traditional graph methods.

Proposed method

• Time-series encoder (1D-CNN or bi-GRU) to obtain reduced, denoised ROI features h_e from BOLD signals.
• Graph generator computes A = h_A h_A^T with h_A = softmax(h_e) to produce a task-oriented brain network with positive weights.
• Graph predictor uses a GNN (K-layer) on nodes initialized with Pearson-based features F and adjacency A to predict labels.
• End-to-end training optimizing L = L_ce + alpha L_intra + beta L_inter + gamma L_sparsity, with three task-aligned regularizers.
• Regularizers include group intra loss (within-class consistency), group inter loss (between-class separation), and sparsity loss (edge sparsity).
• Ablation analysis shows the impact of removing regularizers and the benefits of the end-to-end, task-aware design.

Experimental results

Research questions

• RQ1 Does FBNetGen improve prediction performance over time-series baselines and traditional graph construction methods across datasets?
• RQ2 Are the learned, task-oriented brain graphs interpretable and aligned with known neurobiological findings (e.g., DMN, gender-related connectivity)?
• RQ3 How do the proposed regularizers (intra/group and sparsity) influence predictive accuracy and interpretability?
• RQ4 Do the learned graphs provide consistent, reproducible insights across datasets (PNC and ABCD) and brain atlases?

Key findings

• FBNetGen outperforms all baselines, including time-series encoders and traditional/learnable graph methods, on both PNC and ABCD for AUROC and accuracy.
• Learned task-oriented graphs emphasize prediction-relevant connections and positively weighted edges compatible with GNNs, enabling better discrimination.
• Ablation studies show that including all regularizers (CE + GL + SL) yields more stable and higher performance than ablated variants.
• Interpretability analysis reveals that the learned graphs highlight the default mode network and other modules in line with neurobiological findings on gender differences.
• Across datasets, the method demonstrates consistent interpretability and predictive gains, validating its end-to-end, task-aware design.