Efficient Deep Learning Models for Predicting Individualized Task Activation From Resting‐State Functional Connectivity
6 affiliations
- 1Department of Psychiatry Stanford University Stanford California USA
- 2Department of Psychiatry University of California Los Angeles California USA
- 3Department of Psychology Stanford University Stanford California USA
- 4Department of Psychology Washington University in St. Louis St. Louis Missouri USA
- 5Department of Pediatrics University of Minnesota Minneapolis Minnesota USA
- 6Department of Radiology Weill Cornell Medicine New York New York USA
Abstract
Deep learning models have demonstrated the potential to predict task-evoked brain activation from resting-state functional magnetic resonance imaging, offering a pathway toward individualized brain mapping without requiring task-based data. In this study, we systematically evaluate architectural strategies for improving the efficiency and scalability of such models. Using data from the Human Connectome Project, we replicate the BrainSurfCNN framework and introduce two extensions: BrainSERF, which incorporates channel-wise attention through squeeze-and-excitation modules, and BrainSurfGCN, a graph-based model that leverages cortical mesh topology for efficient message passing. Across multiple evaluation metrics, including spatial correlation, Dice score, Dice AUC, and subject identification accuracy, all models achieve comparable predictive performance. Despite similar accuracy, the proposed models offer distinct advantages. BrainSERF provides modest improvements in capturing individual-specific features, while BrainSurfGCN achieves substantial reductions in model size and training time, highlighting a favorable trade-off between performance and computational efficiency. Beyond architectural comparisons, we investigate factors driving variability in prediction accuracy. We find that behavioral task performance, resting-state data quality, and inter-subject variability in task activation jointly constrain prediction fidelity. In particular, contrasts with lower signal reliability and higher variability exhibit reduced predictability across all models. Together, these findings demonstrate that incorporating topological and functional structural priors can improve the efficiency of deep learning models without sacrificing accuracy, while also emphasizing that prediction performance is fundamentally limited by the reliability of the underlying neural signals.
Related
- Mapping dynamics in neuronal activation and molecular connectivity with functional PET
- Cognitive Control as a Brake on the Depression–Internet Addiction Link: Task-Based fMRI Evidence from Activation and Task-Modulated Functional Connectivity
- Predicting alcohol consumption from resting-state functional connectivity: Feature relevance and the potential of meta-matching.
- Resting-state functional connectivity of hippocampal subregions in predicting escitalopram treatment outcome in major depressive disorder
- Moral and perceptual conflict processing are implemented by distinct brain activation and functional connectivity patterns
BibTeX
@article{madsen2026efficient,
title = {{Efficient Deep Learning Models for Predicting Individualized Task Activation From Resting‐State Functional Connectivity}},
author = {Soren J. Madsen and Young-Eun Lee and Shaun K.L. Quah and Lucina Q. Uddin and Jeanette A. Mumford and Deanna M. Barch and Damien A. Fair and Ian H. Gotlib and Russell Alan Poldrack and Amy F. Kuceyeski and Manish Saggar},
journal = {Human Brain Mapping},
year = {2026},
month = jun,
doi = {10.1002/hbm.70557},
url = {https://doi.org/10.1002/hbm.70557},
note = {CC BY},
}
RIS
TY - JOUR
TI - Efficient Deep Learning Models for Predicting Individualized Task Activation From Resting‐State Functional Connectivity
AU - Soren J. Madsen
AU - Young-Eun Lee
AU - Shaun K.L. Quah
AU - Lucina Q. Uddin
AU - Jeanette A. Mumford
AU - Deanna M. Barch
AU - Damien A. Fair
AU - Ian H. Gotlib
AU - Russell Alan Poldrack
AU - Amy F. Kuceyeski
AU - Manish Saggar
PY - 2026
DA - 2026/06/15
JO - Human Brain Mapping
DO - 10.1002/hbm.70557
UR - https://doi.org/10.1002/hbm.70557
AB - Deep learning models have demonstrated the potential to predict task-evoked brain activation from resting-state functional magnetic resonance imaging, offering a pathway toward individualized brain mapping without requiring task-based data. In this study, we systematically evaluate architectural strategies for improving the efficiency and scalability of such models. Using data from the Human Connectome Project, we replicate the BrainSurfCNN framework and introduce two extensions: BrainSERF, which incorporates channel-wise attention through squeeze-and-excitation modules, and BrainSurfGCN, a graph-based model that leverages cortical mesh topology for efficient message passing. Across multiple evaluation metrics, including spatial correlation, Dice score, Dice AUC, and subject identification accuracy, all models achieve comparable predictive performance. Despite similar accuracy, the proposed models offer distinct advantages. BrainSERF provides modest improvements in capturing individual-specific features, while BrainSurfGCN achieves substantial reductions in model size and training time, highlighting a favorable trade-off between performance and computational efficiency. Beyond architectural comparisons, we investigate factors driving variability in prediction accuracy. We find that behavioral task performance, resting-state data quality, and inter-subject variability in task activation jointly constrain prediction fidelity. In particular, contrasts with lower signal reliability and higher variability exhibit reduced predictability across all models. Together, these findings demonstrate that incorporating topological and functional structural priors can improve the efficiency of deep learning models without sacrificing accuracy, while also emphasizing that prediction performance is fundamentally limited by the reliability of the underlying neural signals.
SN - 1065-9471
SN - 1097-0193
ER -