Estimating cortical thickness trajectories in children across different scanners using transfer learning from...

Area of Science:

• Neuroimaging
• Developmental Neuroscience
• Biostatistics

Background:

• Longitudinal neuroimaging studies face challenges with multi-site data and varying scanner technologies.
• Establishing normative references is essential for interpreting individual brain development.
• Existing methods struggle to account for scanner and site-specific variations in developmental trajectories.

Purpose of the Study:

• To demonstrate the utility of normative models for longitudinal neuroimaging analysis across different scanners.
• To assess the effectiveness of transferring large-scale normative models to smaller developmental cohorts.
• To optimize sample size requirements for effective knowledge transfer in neuroimaging.

Main Methods:

• Estimation of a large-scale reference normative model using Hierarchical Bayesian Regression (N=42,993).
• Transfer of these models to a longitudinal developmental cohort (N=6,285) with multi-scanner data.
• Calculation of individual deviation scores independent of scanner and site effects.

Main Results:

• Transferred normative models provide scanner- and site-independent individual deviation scores.
• As few as 25 samples per site can ensure good performance for knowledge transfer.
• Deviation scores effectively detect morphological heterogeneity in preterm infants.

Conclusions:

• Normative models offer a powerful framework for harmonizing longitudinal neuroimaging data acquired across different scanners and sites.
• Efficient transfer learning strategies can significantly reduce the data requirements for establishing robust normative references.
• This approach enhances the clinical utility of neuroimaging by enabling precise characterization of individual neurodevelopmental trajectories and variations.