Structural Brain Imaging Predicts Individual-Level Task Activation Maps Using Deep Learning.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomarker Discovery

Background:

• Accurate individual functional mapping of brain activity is crucial for biomarker discovery and clinical applications.
• Structural neuroimaging typically does not directly map task-related brain activation.

Purpose of the Study:

• To investigate if structural neuroimaging data can predict inter-subject variations in functional magnetic resonance imaging (fMRI) task activation patterns.
• To identify which structural features are most informative for predicting functional activation.

Main Methods:

• A convolutional neural network (U-Net model) was trained using multimodal structural MRI data (T1-weighted, T2-weighted, diffusion tensor imaging) from 591 subjects.
• The model predicted 47 different fMRI task activation volumes across seven task domains.
• An ablation study was performed to assess the contribution of different structural components and imaging modalities.

Main Results:

• The model successfully predicted individual task activation maps, showing stronger correlations with actual maps than with maps from other subjects.
• Cortical and subcortical shape information independently predicted activation differences, but whole-brain shape was less effective.
• T2-weighted and diffusion tensor imaging provided additional predictive information beyond T1-weighted imaging.

Conclusions:

• Structural neuroimaging data contains predictive information about inter-subject variability in task-based brain activation.
• Cortical folding patterns and subcortical microstructural features are key components linking brain structure to function.
• This approach holds potential for non-invasive biomarker discovery and personalized neurological care.