Fully synthetic neuroimaging data for replication and exploration.

Area of Science:

• Neuroimaging
• Data Science
• Open Science

Background:

• Data sharing is crucial for scientific transparency, exploration, and education.
• Privacy concerns and regulations hinder sharing of human subject/patient neuroimaging data.
• Fully synthetic data offers an alternative for data sharing when real data cannot be shared.

Purpose of the Study:

• To develop and validate a method for generating fully synthetic neuroimaging datasets.
• To ensure synthetic data accurately represents the covariance structure of observed data.
• To facilitate data sharing while mitigating privacy risks.

Main Methods:

• Utilized principles from multiple imputation to replace observed values with synthetic values.
• Created synthetic predictor tables (demographics, behavior, ICV) from 264 pediatric cases.
• Synthesized gray matter images from T1-weighted data using synthetic predictors.

Main Results:

• Synthetic predictor tables closely approximated observed data's pooled variance and statistical estimates.
• Synthetic gray matter data accurately represented variance and voxel-level associations with predictors (age, sex, IQ, ICV).
• Replicated the magnitude and spatial distribution of gray matter effects observed in the original data.

Conclusions:

• The described approach generates fully synthetic neuroimaging data that preserves statistical properties of the original data.
• This method enables widespread data sharing for replication, discovery, and education.
• Fully synthetic neuroimaging datasets reduce privacy disclosure risks inherent in sharing real neuroimaging data.