Source-linked AI summary
Spatio-Temporal Graph Convolution for Resting-State fMRI Analysis
Soham Gadgil, Qingyu Zhao, Adolf Pfefferbaum, Edith V. Sullivan, Ehsan Adeli, Kilian M. Pohl
TL;DR
Existing rs-fMRI deep-learning methods do not jointly preserve spatial functional dependencies and temporal BOLD dynamics. The paper represents functional connectivity as spatio-temporal graphs, trains ST-GCN on short BOLD sub-sequences, and learns edge importance; it reports higher age and sex prediction accuracy than common approaches across large NCANDA and HCP datasets.
Problem
Existing rs-fMRI methods often fail to consider spatial functional dependencies and temporal BOLD dynamics simultaneously.
Method
The paper constructs spatio-temporal graphs from ROI BOLD signals, trains ST-GCN on short sub-sequences, and learns an edge-importance matrix.
Results
ST-GCN achieved higher accuracy than LSTM-based methods across window sizes and reached 77.7% age and 79.8% sex accuracy in the NCANDA no-to-low drinking cohort.
Takeaways & Limitations
Similar optimal sub-sequence windows across datasets support using short BOLD sequences to model dynamic functional connectivity.
Abstract
from arXiv · showhide
The Blood-Oxygen-Level-Dependent (BOLD) signal of resting-state fMRI (rs-fMRI) records the temporal dynamics of intrinsic functional networks in the brain. However, existing deep learning methods applied to rs-fMRI either neglect the functional dependency between different brain regions in a network or discard the information in the temporal dynamics of brain activity. To overcome those shortcomings, we propose to formulate functional connectivity networks within the context of spatio-temporal graphs. We train a spatio-temporal graph convolutional network (ST-GCN) on short sub-sequences of the BOLD time series to model the non-stationary nature of functional connectivity. Simultaneously, the model learns the importance of graph edges within ST-GCN to gain insight into the functional connectivities contributing to the prediction. In analyzing the rs-fMRI of the Human Connectome Project (HCP, N=1,091) and the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA, N=773), ST-GCN is significantly more accurate than common approaches in predicting gender and age based on BOLD signals. Furthermore, the brain regions and functional connections significantly contributing to the predictions of our model are important markers according to the neuroscience literature.
1 Introduction
The paper frames rs-fMRI as structured spatio-temporal data and introduces ST-GCN to jointly model functional dependencies between brain regions and temporal BOLD dynamics.
- Spatio-temporal graphs represent temporal activity dynamics at each brain region alongside spatial functional interactions between regions.
- Existing rs-fMRI deep-learning methods often model spatial or temporal information separately rather than considering both aspects simultaneously.
- ST-GCN automatically extracts features that jointly characterize functional connectivity patterns and their temporal dynamics within BOLD series.
- Training on short BOLD sub-sequences is intended to model the non-stationary nature of functional connectivity.
- Learning an edge-importance matrix identifies selective functional connections contributing significantly to predictions.
2 ST-GCN for rs-fMRI Analysis
The method constructs spatio-temporal graphs from ROI BOLD signals, applies graph and temporal convolutions in ST-GCN layers, and learns edge importance for interpretability.
- Representing Functional Networks as Spatio-Temporal Graphs: The graph contains ROI-time nodes, temporal edges linking each ROI across successive frames, and spatial edges weighted by functional affinity.
- Spatio-Temporal Graph Convolution (ST-GC): ST-GC neighborhoods combine spatial and temporal reach, controlled by spatial kernel size K and temporal kernel size Γ.
- Spatio-Temporal Graph Convolution (ST-GC): The convolution decomposes into a spatial graph kernel and a temporal convolutional kernel applied to intermediate node features.
- Classifying BOLD Time Series by ST-GCN: The classifier uses three ST-GC layers, global average pooling, and a sigmoid-based fully connected output for class probabilities.
- Classifying BOLD Time Series by ST-GCN: A shared positive symmetric edge-importance matrix rescales neighboring-node contributions during spatial convolution.
- Training ST-GCN: Training and testing use short BOLD sub-sequences, with testing predictions combined by voting across sampled windows.
3 Experiments
The experiments evaluate age and sex prediction on large NCANDA and HCP rs-fMRI cohorts, including window-size comparisons and controls for drinking-related confounding.
- NCANDA: NCANDA includes 773 adolescents aged 12–21 years, with balanced age and sex groups and 638 participants meeting no-to-low alcohol drinking criteria.
- HCP: HCP includes 1,091 young adults aged 22–35 years after excluding five scans, with 498 females and 593 males analyzed.
- Experimental Setup: NCANDA experiments use five-fold cross-validation across sub-sequence windows from T′ = 16 through the full sequence, with S = 64 voting windows except for full-sequence analysis.
- Baselines: Sex classification is repeated on both NCANDA and HCP, alongside comparisons with MLP and LSTM-based baselines.
4 Results and Analysis
Across NCANDA and HCP, ST-GCN outperformed comparison methods for age and sex prediction using short BOLD subsequences, while its learned edge importance highlighted neurobiologically relevant regions and connections.
- NCANDA: p<.0001: ST-GCN achieved significantly higher accuracy than two LSTM-based methods across NCANDA window sizes.
- NCANDA: 77.7%: ST-GCN’s highest NCANDA age-prediction accuracy occurred at T ′ = 24, while sex prediction reached 79.8% at T ′ = 32.Sex accuracy remained relatively stable for T ′ = [24, 40].
- NCANDA: ST-GCN identified the inferior temporal lobe and frontal-posterior-cingulate connection as important for NCANDA sex prediction.The connection defines the default mode network and aligns with prior resting-state and structural MRI findings.
- NCANDA: ST-GCN identified supramarginal and par opercularis regions and their connection as important for NCANDA age prediction.That connection defines the inferior temporo-parieto-frontal network, reported to decrease in older NCANDA adolescents.
- HCP: Static correlation coefficients produced suboptimal performance relative to deep-learning methods trained on subsequences of T ′ ∈[32, 128].The authors interpret this as evidence that functional-interaction dynamics require spatio-temporal modeling.
- HCP: 83.7%: ST-GCN achieved the highest HCP sex-prediction accuracy at T ′ = 128 (92.16 s), outperforming both LSTM-based methods across window sizes.Prior studies used different cohorts and processing pipelines, so their results were not strictly comparable.
5 Conclusion
The paper introduced an ST-GCN framework for rs-fMRI and showed that short BOLD sequences can support dynamic functional-connectivity modeling across two datasets.
- The framework analyzes rs-fMRI with spatio-temporal graph convolution networks.
- Similar optimal subsequence window sizes across the two datasets supported using short BOLD sequences to model dynamic functional connectivity.
- Future work includes dynamic graph structures, automatic window-size selection, finer brain parcellations, and neuropsychiatric biomarkers.