DeepResBat: deep residual batch harmonization accounting for covariate distribution differences.

Area of Science:

• Neuroimaging
• Medical Data Analysis
• Machine Learning

Background:

• Pooling MRI data from multiple sources requires harmonization to minimize site-specific variations.
• Existing methods like ComBat use mixed-effects models, while deep learning approaches like conditional variational autoencoders (cVAE) are emerging.
• Current deep learning methods often overlook covariate distribution differences, potentially impacting harmonization quality.

Purpose of the Study:

• To evaluate the impact of covariate distribution differences on MRI harmonization.
• To propose novel deep neural network (DNN)-based harmonization methods that explicitly account for covariates.
• To compare the performance of new methods against existing techniques.

Main Methods:

• Developed two covariate-aware DNN harmonization approaches: covariate VAE (coVAE) and DeepResBat.
• coVAE extends cVAE by incorporating covariates into latent representations.
• DeepResBat uses a residual framework, removing covariate effects, harmonizing site differences with cVAE, and reintroducing covariate effects.

Main Results:

• DeepResBat and coVAE demonstrated superior performance in reducing dataset differences and enhancing biological effects compared to ComBat, CovBat, and cVAE.
• The study utilized three large datasets with 2787 participants and 10085 T1 scans.
• coVAE exhibited a tendency to generate false positive associations between MRI data and covariates.

Conclusions:

• Deep learning-based harmonization methods that account for covariates can outperform traditional and non-covariate-aware DNN approaches.
• DeepResBat is presented as a robust deep learning alternative to ComBat for MRI data harmonization.
• Researchers should be cautious of potential false positive findings with certain DNN harmonization methods like coVAE.