Voxel-wise Brain State Prediction Using Swin Transformer

Motivation: Understanding brain dynamics is important for neuroscience and mental health. fMRI enables the measurement of neural activities, which can be regarded as brain states. Goal(s): We aim to predict the human resting-state brain state represented as voxel-wise fMRI data. Approach: We used 100 HCP subjects resting-state fMRI data. To predict the brain state, we adapted a transformer for the 4D fMRI by adding a convolutional decoder. Results: Our model achieved a mean squared error of 0.069 when predicting the single brain states based on only 23s of prior fMRI time series. The predicted brain states resemble the real fMRI data. Impact: This work provides critical insights into the temporal organization of the human brain in healthy individuals, showing the potential of swin transformer in predicting brain states on the voxel level, which may reduce the fMRI scan time in the future.