Deep Generative Medical Image Harmonization for Improving Cross-Site Generalization in Deep Learning Predictors.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning in medical imaging faces challenges in clinical adoption due to poor generalization across different devices and protocols.
• Brain age prediction models can serve as biomarkers for brain health, but require improved cross-site generalizability.

Purpose of the Study:

• To develop and evaluate a deep learning-based image harmonization method.
• The goal is to enhance the cross-site generalizability of deep learning models for brain age prediction.

Main Methods:

• Retrospective analysis of 8,876 subjects from six sites using brain imaging data (1.5T and 3T MRI).
• StarGAN v2 was employed for image harmonization, mapping diverse datasets to a reference domain.
• Age prediction model performance was evaluated using harmonized, histogram-matched, and unharmonized data.

Main Results:

• Generative adversarial network (GAN)-based harmonization reduced the mean absolute error (MAE) of brain age prediction from 15.81 years to 7.21 years overall.
• In multisite testing, GAN-based harmonization improved MAE from 9.78 years to 5.32 years.
• This demonstrates substantial improvements in out-of-sample prediction accuracy.

Conclusions:

• GAN-based medical image harmonization shows promise for improving cross-site deep learning generalization.
• This technique is a potential tool to overcome limitations in current AI-based medical imaging applications.
• Further research is warranted to fully explore its clinical utility.