st-DenseViT: A Weakly Supervised Spatiotemporal Vision Transformer for Dense Prediction of Dynamic Brain Networks.

Area of Science:

• Computational Neuroscience
• Neuroimaging Analysis
• Computer Vision for Brain Mapping

Background:

• Current computational neuroscience models struggle to capture the full spatiotemporal dynamics of brain networks.
• A more granular representation of brain activity over time is needed for precise tracking of neural fluctuations.

Purpose of the Study:

• To develop a novel weakly supervised spatiotemporal dense prediction model for generating personalized 4D dynamic brain networks from fMRI data.
• To capture and represent complex spatiotemporal dynamics within brain networks more effectively.

Main Methods:

• Utilized a vision transformer (ViT) backbone for joint spatial and temporal encoding of fMRI data.
• Employed spatially constrained windowed independent component analysis (ICA) components as weak supervision for training.
• Evaluated the model on large-scale resting-state fMRI datasets using various statistical metrics.

Main Results:

• The model successfully generated dynamic 4D brain maps capturing inter-subject and temporal variations, effectively denoising noisy data.
• Statistically significant differences in brain maps were observed between schizophrenia patients and healthy controls, particularly within the Default Mode Network (DMN).
• Identified higher activity in the thalamus for healthy controls compared to schizophrenia patients within the DMN.

Conclusions:

• The proposed model provides an effective method for dynamic brain mapping, capturing significant spatiotemporal variations.
• Weakly supervised learning with ICA components allows for robust dynamic pattern learning without direct ground-truth data.
• The model demonstrates potential for differentiating clinical populations and advancing the study of brain dynamics.