PaperlayerView exact revision on arXiv

Source-linked AI summary

FBNETGEN: Task-aware GNN-based fMRI Analysis via Functional Brain Network Generation

Xuan Kan, Hejie Cui, Joshua Lukemire, Ying Guo, Carl Yang

arXiv:2205.12465v2cs.LGcs.NEeess.IVq-bio.NC

TL;DR

Traditional fMRI functional networks can be noisy, task-unaware, and incompatible with GNNs. FBNETGEN addresses these issues with an end-to-end encoder, task-oriented graph generator, and GNN predictor, and experiments report consistently stronger performance with neurobiologically aligned interpretations.

  • Problem

    Traditional fMRI functional brain networks are noisy, unaware of downstream prediction tasks, and incompatible with deep GNN models.

  • Method

    FBNETGEN jointly encodes BOLD time series, generates task-oriented brain networks, and performs clinical prediction with a GNN in an end-to-end trainable model.

  • Results

    FBNetGen consistently outperforms four types of baselines and produces interpretations aligned with neurobiological findings.

  • Takeaways & Limitations

    The generated graphs provide prediction-specific brain-region interpretations alongside clinical predictions.

Abstract

from arXiv · show

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.

1. Introduction

FBNETGEN addresses noisy, task-agnostic brain networks and their incompatibility with GNNs by learning task-oriented functional networks directly from fMRI time series. Experiments report stronger prediction and interpretable brain-region findings.

  • Motivation: The motivation includes improving interpretation of neural origins underlying observed sex differences in cognition.The paper frames explicit interpretation of these differences as important for biomedical neuroscience.
  • Existing approach: Traditional fMRI network analysis extracts ROI time series, computes pairwise connectivity, and then uses the resulting networks for clinical prediction.Common connectivity measures include Pearson and partial correlation.
  • Limitations: Correlation-based networks ignore temporal order and may contain negative edges that are incompatible with standard GNN message passing.Shuffling time steps does not change correlation results, and existing GNN applications therefore require additional handling of negative weights.
  • FBNETGEN: The proposed end-to-end pipeline encodes BOLD time series, generates task-oriented brain networks, and predicts clinical outcomes with a GNN.The graph generator is designed to produce networks compatible with GNNs and interpretable for brain-region connectivity.
  • Results: FBNETGEN consistently achieves better gender-prediction accuracy than four types of baselines while identifying regions aligned with neurobiological findings.The experiments use real-world fMRI datasets and compare against time-series, traditional graph, learnable graph, and other deep-model baselines.

2. FBNetGen

FBNetGen combines temporal feature encoding, task-oriented graph generation, and GNN prediction in an end-to-end framework. Its generator uses class-aware regularization and sparsity to produce interpretable, GNN-compatible brain networks.

  • 2.2. Time-series Encoder: Bi-GRU and 1D-CNN encoders capture nonlinear or temporal information that PCA and ICA-based approaches ignore.The paper selects bi-GRU over LSTM because it can achieve similar performance with fewer parameters, while empirical results attribute gains to the generated graphs.
  • 2.3. Graph Generator: FBNetGen generates learnable ROI connectivity matrices from encoded time-series features instead of using traditional Pearson correlations.The generated matrix stores pairwise ROI connectivity strengths and is used by the subsequent predictor.
  • 2.3. Graph Generator: Limited supervision motivates exterior regularization because the encoder can contain many parameters relative to the available training samples.The paper gives the PNC training set as 353 samples and notes that a 1D-CNN encoder can reach about 20k parameters.
  • 2.3. Graph Generator: Group losses encourage within-class graph consistency and between-class graph differences, while sparsity highlights contributory task-specific ROI connections.The inter-group loss maximizes differences across groups, and the sparsity loss mitigates excessive graph-value deviations.
  • 2.4. Graph Predictor: The GNN propagates node features and graph structure, pools node embeddings into a graph representation, and applies an MLP for class prediction.Initial node features are Pearson-correlation vectors, while the connectivity matrix comes from the graph generator.
  • 2.1. Overview: The model combines a time-series encoder, graph generator, and GNN predictor into an end-to-end pipeline from BOLD signals to labels and task-oriented graphs.The final objective combines cross-entropy with intra-group, inter-group, and sparsity regularizers.

3. Experiments

Experiments evaluate FBNetGen’s predictive effectiveness and interpretability on two fMRI datasets, comparing it with multiple baseline families and analyzing its components. FBNetGen achieves stronger prediction performance and produces task-relevant brain-network patterns aligned with neuroscience findings.

  • 3.1. Experimental Settings: The evaluation uses gender prediction on PNC and ABCD, reporting AUROC and accuracy averaged over five test-set runs.Both datasets are treated as balanced binary-classification benchmarks.
  • 3.2. Performance Comparison and Ablation Studies: FBNetGen consistently surpasses time-series, traditional graph, learnable graph, and other deep-model baselines in gender prediction.The comparison uses four baseline types, with grid search applied to baselines for fairness.
  • 3.2. Performance Comparison and Ablation Studies: On both PNC and ABCD, the full model achieves the highest ablation performance and improves more stably than versions removing Group Loss or Sparsity Loss.CE+SL and CE+GL remain close to optimal, supporting the usefulness of exterior regularization.
  • 3.3. Interpretability Analysis: Learnable graphs consistently highlight the default mode network across PNC and ABCD, whereas Pearson graphs emphasize within-module connections less specifically tied to gender prediction.The default mode network alignment remains across datasets and brain atlases.
  • 3.3. Interpretability Analysis: Top-ranked discriminative modules include memory retrieval, default mode, and ventral attention networks in PNC, and default mode and ventral salience networks in ABCD.These rankings align with reported literature on regions showing sex differences more closely than Pearson graphs.

4. Conclusion

The paper presents FBNetGen as a task-aware GNN framework that jointly generates functional brain networks and predicts clinical outcomes from fMRI BOLD signals. Experiments show consistent baseline improvements and interpretations aligned with neurobiological findings.

  • 4. Conclusion: FBNetGen jointly generates brain connectivity matrices and predicts clinical outcomes from fMRI BOLD signal series.The framework is designed as a task-aware GNN-based approach for functional brain-network generation.
  • 4. Conclusion: FBNetGen consistently outperforms four baseline types and provides interpretation results aligned with neurobiological findings.The conclusion summarizes both predictive effectiveness and interpretability evidence.

A.1. fMRI-based Brain Network Analysis

fMRI brain-network analysis uses functional connectivity to study brain organization, but existing approaches and GNN applications leave task-aware, interpretable network generation insufficiently addressed. FBNetGen is positioned against these limitations.

  • A.1. fMRI-based Brain Network Analysis: Functional connectivity links fMRI-derived brain networks to intrinsic neural processing and cognitive, emotional, visual, and motor functions.These networks represent spatially disjoint regions with coherent BOLD temporal dynamics.
  • A.1. fMRI-based Brain Network Analysis: FCNet generates functional connectivity matrices but requires separate feature selection and SVM prediction, so its networks are not prediction-task-specific.The passage contrasts this pipeline with an end-to-end task-aware approach.
  • A.2. Graph Neural Networks: GNNs combine node features and graph structures end-to-end for downstream prediction tasks, motivating their use in brain-network analysis.The cited examples include graph convolution, attention, and graph-isomorphism architectures.
  • A.1. fMRI-based Brain Network Analysis: Prior work generated brain networks without downstream tasks, and their interpretability for those tasks remained unexplored.The paper identifies task-oriented interpretability as a critical need in neuroimaging research.

B.1. Dataset

The experiments use two real-world fMRI cohorts and evaluate gender prediction as a balanced binary-classification task. AUROC is the primary comparison metric, with accuracy added for practical prediction assessment.

  • B.1. Dataset: The study evaluates FBNetGen on PNC and ABCD real-world fMRI datasets.PNC contributes 503 subjects aged 8–21, while ABCD is a large longitudinal study of children recruited across 21 U.S. sites.
  • B.1. Dataset: Gender prediction is selected as the downstream evaluation task using labels available in both datasets.The task is binary classification, and both datasets are balanced across classes.
  • B.1. Dataset: AUROC is the primary metric for fair comparison, while accuracy measures practical prediction performance.Reported performances are averaged over five test-set runs.

B.3. Implementation details

Implementation uses configurable 1D-CNN or bi-GRU encoders to transform ROI BOLD sequences into embeddings before graph generation, with hyperparameters tuned for the datasets.

  • B.3. Implementation details: The implementation fixes the CNN depth at 3 layers, uses 4 GRU layers, and searches embedding size d over {4, 8, 12}.Window sizes are tuned over {4, 6, 8} for PNC and {8, 16, 32} for ABCD.
  • B.3. Implementation details: FBNetGen’s feature encoder, graph generator, and graph predictor have complexities O(µvt), O(v^2), and O(kv^2), respectively.The overall complexity is reported as O(v(v + t)).
  • B.3. Implementation details: The feature encoder produces d-dimensional embeddings for v ROIs from BOLD sequences using either a u-layer 1D-CNN or a bi-GRU.For the CNN, embeddings are obtained through convolution, max pooling, and an MLP; the bi-GRU processes segments before the final MLP projection.

Appendix D. Influence of Hyper-parameters

Hyperparameter analysis varies window size and embedding size for FBNetGen and LDS encoders, showing stable FBNetGen performance and broad superiority over baselines.

  • Appendix D. Influence of Hyper-parameters: FBNetGen consistently outperforms LDS-CNN and LDS-GRU across large ranges of window sizes and embedding sizes.The comparison covers both PNC and ABCD datasets.
  • Appendix D. Influence of Hyper-parameters: Increasing window size or embedding size does not necessarily improve FBNetGen’s overall performance, indicating stable behavior across these settings.The analysis varies both hyperparameters in the 1D-CNN and GRU feature encoders.
  • Appendix D. Influence of Hyper-parameters: With more training data, FBNetGen exceeds the baseline methods by a large margin.This pattern is reported in the hyperparameter experiments.

Appendix E. Influence of Pooling Strategies

The pooling-strategy ablation compares sum and concat pooling, with concat reported as the consistently stronger strategy across datasets.

  • Appendix E. Influence of Pooling Strategies: Concat pooling consistently outperforms sum pooling across all datasets.The comparison is reported using AUROC performance in Table 4.
  • Appendix E. Influence of Pooling Strategies: The ablation evaluates the original model with different pooling strategies, specifically sum and concat.Each pooling variant is compared by its observed performance.

Appendix F. Training Curves of FBNetGen Variants

Training curves compare FBNetGen variants across two datasets, showing benefits from Group Loss and best performance when both regularizers accompany supervised cross-entropy.

  • Appendix F. Training Curves of FBNetGen Variants: Using Cross Entropy, Group Loss, and Sparsity Loss together consistently achieves the best performance across the evaluated variants.The result is reported for both datasets through the variant training-curve analysis.
  • Appendix F. Training Curves of FBNetGen Variants: Group Loss produces pronounced performance improvement in the FBNetGen variants.The training curves are reported to display similar patterns across datasets.
  • Appendix F. Training Curves of FBNetGen Variants: The training curves of different FBNetGen variants display similar patterns across datasets.The supplied passages identify the compared variants but do not provide curve values.

H.2. Difference Score T of Functional Modules on Learnable Graph and Pearson Graph

FBNETGEN’s task-aware learnable graphs better highlight gender-differentiated functional modules than Pearson graphs on PNC, while results are similar on ABCD. The learned emphasis also aligns with neurobiological findings, including auditory-related Somatomotor regions.

  • FBNETGEN better highlights gender-differentiated functional modules than Pearson graphs on PNC, whereas the two graph types perform similarly on ABCD.The ABCD similarity is attributed to its having fewer functional modules than PNC.
  • The learnable graphs emphasize Somatomotor on ABCD, whose auditory-related ROIs are highly differentiated between genders.
  • Difference-score comparisons indicate that FBNETGEN graphs are task-oriented and capture more authentic gender differences than Pearson graphs.
Loading 2205.12465v2…