Voxel-Level Brain States Prediction Using Swin Transformer

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Understanding brain dynamics is crucial for neuroscience and mental health.
• Functional magnetic resonance imaging (fMRI) measures neural activity via blood-oxygen-level-dependent (BOLD) signals, reflecting brain states.

Purpose of the Study:

• To predict future human resting brain states using fMRI data.
• To develop a novel deep learning architecture for accurate spatio-temporal fMRI analysis.

Main Methods:

• A novel architecture combining a 4D Shifted Window (Swin) Transformer encoder and a convolutional decoder was proposed.
• The model was trained and tested on fMRI data from 100 unrelated subjects from the Human Connectome Project (HCP).

Main Results:

• The model achieved high accuracy in predicting 7.2s of resting-state brain activity from a prior 23.04s fMRI time series.
• The predicted brain states closely matched the BOLD contrast and dynamics of actual brain activity.

Conclusions:

• The Swin Transformer model effectively learns the spatiotemporal organization of human brain activity from fMRI data at high resolution.
• This approach shows potential for reducing fMRI scan duration and advancing brain-computer interfaces.