Extensive T1-weighted MRI preprocessing improves generalizability of deep brain age prediction models

Area of Science:

• Neuroimaging
• Radiology
• Artificial Intelligence

Background:

• Brain age estimation from T1w MRI is a key biomarker for brain aging and diseases.
• Current brain age prediction accuracy is within 2-3 years, but cross-study comparisons are difficult due to varied preprocessing.
• Deep learning models are increasingly used for brain age prediction.

Purpose of the Study:

• To investigate the impact of T1w image preprocessing on the performance of deep learning brain age models.
• To compare the effects of different registration transforms, grayscale correction, and software implementations.
• To determine optimal preprocessing strategies for robust brain age prediction.

Main Methods:

• Evaluated four deep learning brain age models using four distinct T1w preprocessing pipelines.
• Pipelines varied in registration transform (rigid vs. affine), grayscale correction, and software.
• Assessed prediction error (Mean Absolute Error - MAE) across different preprocessing conditions and model types (2D vs. 3D).

Main Results:

• Preprocessing choices significantly affected prediction error, with MAE increasing by up to 0.75 years.
• Affine registration to a brain atlas statistically improved MAE compared to rigid registration.
• 3D models with 1mm³ resolution were less sensitive to preprocessing variations than 2D or downsampled 3D models.
• Extensive preprocessing improved MAE for new datasets, contradicting previous findings.

Conclusions:

• T1w image preprocessing critically influences brain age prediction accuracy.
• Extensive preprocessing, especially affine registration, enhances model performance on unseen data.
• Offset correction is essential for generalizing brain age models to diverse datasets, irrespective of their preprocessing.