Generalization of diffusion magnetic resonance imaging-based brain age prediction model through transfer learning

Area of Science:

• Neuroimaging
• Machine Learning
• Biostatistics

Background:

• Diffusion magnetic resonance imaging (dMRI) and machine learning enable brain age prediction.
• High inter-site variability in dMRI data limits model generalizability.
• Brain age prediction models struggle to perform across different imaging conditions.

Purpose of the Study:

• To generalize dMRI-based brain age prediction models to diverse datasets using transfer learning.
• To investigate optimal tuning data size and feature types for domain adaptation.
• To compare the transfer learning approach with statistical covariate methods.

Main Methods:

• Employed a transfer learning approach for domain adaptation of brain age prediction models.
• Utilized a large dMRI dataset as the source domain and three distinct datasets as target domains.
• Evaluated model performance based on prediction accuracy, tuning data size, feature types, and test-retest reliability.

Main Results:

• Transferred models achieved significantly improved prediction performance across all target cohorts (p < 0.001).
• Mean absolute error in age prediction reduced substantially, e.g., from 13.89 to 4.78 years in Cohort 1.
• The transferred model demonstrated high test-retest reliability (ICC = 0.950) and clinical sensitivity in schizophrenia patients.

Conclusions:

• Transfer learning is an efficient method for generalizing dMRI-based brain age prediction models.
• Optimal transfer learning strategies and tuning data sizes are crucial for different dMRI acquisition scenarios.
• The generalized model shows promise for reliable and accurate brain age assessment in diverse clinical settings.