TS-AI: A deep learning pipeline for multimodal subject-specific parcellation with task contrasts synthesis

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Accurate individual brain mapping is vital for understanding function.
• Task-based functional MRI (tfMRI) provides subject-specific activation but is resource-intensive.
• Existing methods lack efficiency and individual specificity.

Purpose of the Study:

• To develop an efficient, individualized brain atlas creation method.
• To predict tfMRI data and localize functional subregions without actual scans.
• To enhance the accuracy and generalizability of individual brain parcellation.

Main Methods:

• A two-stage network model (TS-AI) was proposed.
• TS-AI synthesizes task contrast maps using anatomical and resting-state connectivity.
• A deep neural network with a novel feature consistency loss individualizes the atlas.

Main Results:

• TS-AI successfully generated individualized brain atlases.
• The method demonstrated superior performance, reliability, and generalizability.
• Validated parcellations showed high homogeneity and predictive power for cognitive behaviors.
• Identified accelerated shrinkage in Alzheimer's disease progression.

Conclusions:

• TS-AI offers an efficient alternative to traditional tfMRI for individual brain atlasing.
• The model enhances individual specificity and mitigates overfitting.
• TS-AI shows significant potential for clinical research, particularly in neurodegenerative diseases.