Related Experiment Video
Updated: Jun 20, 2026
07:56
Resting-State Connectivity and Neuroimaging of Prefrontal Cortex Activity During a Block-Design Yoga Asana Practice Using fNIRS
Published on: June 24, 2025
Efficient Deep Learning Models for Predicting Individualized Task Activation From Resting-State Functional
Soren J Madsen1, Young-Eun Lee1, Shaun K L Quah1
1Department of Psychiatry, Stanford University, Stanford, California, USA.
Human Brain Mapping
|June 18, 2026
Summary
Deep learning models can predict brain activity from resting-state fMRI, enabling personalized brain mapping. New models improve efficiency without sacrificing accuracy, though prediction is limited by signal reliability.
Area of Science:
- Neuroimaging
- Machine Learning
- Computational Neuroscience
Background:
- Predicting task-evoked brain activation from resting-state fMRI (rs-fMRI) offers a non-invasive approach to individualized brain mapping.
- Deep learning models show promise in this area, but efficiency and scalability require further investigation.
Purpose of the Study:
- To systematically evaluate and enhance deep learning architectures for predicting brain activation from rs-fMRI.
- To introduce and assess novel models, BrainSERF and BrainSurfGCN, focusing on efficiency and performance.
Main Methods:
- Replication of the BrainSurfCNN framework using Human Connectome Project data.
- Development of BrainSERF with channel-wise attention and BrainSurfGCN utilizing graph convolutional networks.
- Evaluation using spatial correlation, Dice score, Dice AUC, and subject identification accuracy.
Main Results:
- All evaluated models demonstrated comparable predictive performance.
- BrainSERF offered marginal improvements in individual-specific feature capture.
- BrainSurfGCN significantly reduced model size and training time, demonstrating improved computational efficiency.
- Prediction accuracy was found to be constrained by resting-state data quality, behavioral task performance, and inter-subject variability.
Conclusions:
- Incorporating topological and functional priors enhances deep learning model efficiency for brain activation prediction without compromising accuracy.
- The reliability of neural signals fundamentally limits prediction performance.
- Graph-based approaches like BrainSurfGCN offer a promising balance of accuracy and computational efficiency.
