Empowering Data Sharing in Neuroscience: A Deep Learning Deidentification Method for Pediatric Brain MRIs.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Pediatric Neuroscience

Background:

• Privacy concerns with identifiable facial features in pediatric brain scans limit data sharing.
• Existing defacing tools are inadequate for pediatric neuroimaging, hindering research.
• National Institutes of Health mandates necessitate novel data sharing solutions.

Purpose of the Study:

• To develop an artificial intelligence (AI)-powered tool for automatic defacing of pediatric brain MRIs.
• To address limitations of current defacing tools in pediatric neuroimaging.
• To facilitate pediatric neuroimaging data sharing and research.

Main Methods:

• Utilized deep learning methodologies (nnU-Net) on a diverse multi-institutional dataset.
• Trained the model on 976 multiparametric MRI scans from 244 pediatric patients (7 days to 21 years).
• Included T1-weighted, T1-contrast-enhanced, T2-weighted, and T2-FLAIR MRI sequences.

Main Results:

• Achieved high accuracy (98%) in removing facial regions across various MRI types.
• Demonstrated lower accuracy (73%) for ear removal.
• Showed minimal impact on downstream research analyses, including global/regional brain measures and AI-generated volumes.

Conclusions:

• The AI defacing model is effective for pediatric brain MRIs across multiple sequences.
• Defacing minimally impacts downstream research utility, supporting data sharing.
• A freely available software package (pediatric-auto-defacer) is provided to advance research and data sharing practices.