Brain Age Prediction: Deep Models Need a Hand to Generalize.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Brain age prediction from T1-weighted MRI is a key marker for brain aging.
• Deep learning models show potential but struggle with generalization to new data due to limited training data and model complexity.
• A generalization gap often exists between training and unseen data performance.

Purpose of the Study:

• To assess the SFCN-reg deep learning model for brain age prediction.
• To address the generalization gap in brain age prediction models.
• To improve the clinical applicability of deep learning in neuroimaging.

Main Methods:

• Utilized a VGG-16 based deep model (SFCN-reg).
• Employed comprehensive preprocessing, extensive data augmentation, and model regularization.
• Trained the model using UK Biobank data.

Main Results:

• Reduced generalization Mean Absolute Error (MAE) by 47% on the ADNI dataset (2.79 years).
• Reduced generalization MAE by 12% on the AIBL dataset (3.75 years).
• Achieved up to 13% reduction in scan-rescan error (0.70 years) and improved robustness.

Conclusions:

• High-quality preprocessing and robust training are crucial for accurate brain age prediction.
• The study demonstrates a pathway to narrow the generalization gap for clinical use.
• Improved models enhance neuroimaging research and clinical applications of brain age prediction.