Deep learning in resting-state fMRI.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Human brain functional magnetic resonance imaging (fMRI) data presents complex spatiotemporal variations.
• Current analysis often relies on regional and connection-level interpretations, or pre-engineered temporal transformations.
• Direct spatiotemporal learning in 4D fMRI voxel-time space may yield superior brain representations.

Purpose of the Study:

• To develop and evaluate a 4D deep learning (DL) model for direct spatiotemporal analysis of fMRI data.
• To compare the performance of this 4D DL model against standard machine learning (SML) and existing DL methods using fMRI features.
• To investigate the potential of DL models trained at the voxel level for enhanced brain representation.

Main Methods:

• Extension of a structural MRI (sMRI) DL pipeline to incorporate temporal variations.
• End-to-end training of a 4D DL model on preprocessed fMRI data.
• Comparison of DL model performance against SML and other DL approaches on voxel, region, and connection-level fMRI features.

Main Results:

• The 4D DL model generated highly discriminative encodings, outperforming SML and other DL methods on most fMRI features.
• The model's performance surpassed traditional methods, except for the simple temporal mean of fMRI data.
• Specific fMRI features were identified where DL significantly outperformed SML for voxel-level analysis.

Conclusions:

• Deep learning models trained directly on voxel-level fMRI data demonstrate significant efficiency and potential for enhanced brain representation.
• Spatiotemporal learning in the 4D fMRI voxel-time space offers advantages over pre-engineered features and traditional analysis levels.
• Further development of auxiliary tools is crucial for interpreting complex DL models in neuroimaging analysis.