Diffusion deep learning for brain age prediction and longitudinal tracking in children through adulthood.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Developmental Neuroscience

Background:

• Deep learning (DL) models predict biological age from brain MRI (brain age) as a biomarker for brain health.
• Existing DL tools are limited by single-institution, cross-sectional data, hindering clinical translation.
• Advancing brain age prediction requires multi-site, longitudinal data for broader generalization.

Purpose of the Study:

• To develop and validate a novel DL framework for predicting brain age in developing humans using multi-site, longitudinal MRI data.
• To improve the accuracy and generalizability of brain age prediction for clinical applications.
• To provide an independently implementable code for wider research community use.

Main Methods:

• Leveraged 32,851 T1-weighted MRI scans from healthy individuals aged 3-30 across 16 multisite datasets.
• Developed and evaluated several DL frameworks, including a novel regression diffusion DL network (AgeDiffuse).
• Conducted multisite external validation on 5 and 3 independent datasets, assessing prediction accuracy (MAE) and correlation with brain structure changes.

Main Results:

• AgeDiffuse achieved a mean absolute error (MAE) of 2.78 years in the first external validation and 1.97 years in the second.
• AgeDiffuse brain age predictions better reflected age-related brain structure volume changes (R²=0.48) compared to chronological age (R²=0.37).
• Longitudinal brain age predictions closely tracked chronological age at the individual level.

Conclusions:

• The AgeDiffuse DL network demonstrates superior performance in predicting brain age across diverse populations and sites.
• This validated, publicly available tool serves as a robust, non-invasive biomarker for brain health and development in youth.
• AgeDiffuse facilitates clinical translation and further research into neurodevelopmental trajectories and brain health.