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

Area of Science:

• Neuroimaging analysis
• Medical data harmonization
• Machine learning in neuroscience

Background:

• Pooling MRI data from multiple sites requires harmonization to reduce inter-site variability.
• Traditional methods like ComBat use mixed-effects models, while deep learning approaches like cVAE are emerging.
• Current deep learning methods often neglect covariate distribution differences, potentially leading to suboptimal harmonization.

Purpose of the Study:

• To develop and evaluate novel deep learning-based MRI harmonization methods that explicitly account for covariates.
• To compare the performance of covariate-aware deep learning methods against existing approaches.
• To address the limitations of current deep learning harmonization techniques regarding covariate handling.

Main Methods:

• Proposed two covariate-aware deep learning harmonization methods: covariate VAE (coVAE) and DeepResBat.
• coVAE extends cVAE by incorporating covariates into latent representations.
• DeepResBat utilizes a residual framework, removing covariate effects first, then site effects, and finally reintroducing covariate effects.

Main Results:

• DeepResBat and coVAE outperformed ComBat, CovBat, and cVAE in reducing dataset differences and preserving biological effects.
• coVAE demonstrated a tendency to hallucinate spurious associations between MRI data and covariates.
• DeepResBat proved to be an effective deep learning alternative to ComBat for MRI harmonization.

Conclusions:

• Covariate-aware deep learning approaches can significantly improve MRI data harmonization.
• DeepResBat is a promising deep learning method for harmonizing multi-site MRI data, offering an effective alternative to ComBat.
• Researchers should be cautious of potential false positive findings with certain deep learning harmonization methods like coVAE.