DeepComBat: A statistically motivated, hyperparameter-robust, deep learning approach to harmonization of neuroimaging...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Multi-scanner/protocol neuroimaging data present batch artifacts, confounding analyses and reducing reproducibility.
• Existing harmonization methods struggle to fully eliminate batch effects, especially with complex downstream models.
• Batch effects can be implicitly incorporated by models, leading to unreliable results.

Purpose of the Study:

• To introduce DeepComBat, a novel deep learning harmonization method.
• To address limitations of current methods in removing batch effects from neuroimaging data.
• To improve the reliability and reproducibility of multi-batch neuroimaging data analysis.

Main Methods:

• Developed DeepComBat, integrating a conditional variational autoencoder with the ComBat statistical method.
• Leveraged deep learning to model complex feature relationships while retaining statistical rigor.
• Applied the method to cortical thickness measurements from a cognitive-aging cohort.

Main Results:

• DeepComBat demonstrated superior performance in removing batch effects compared to existing methods.
• The method successfully preserved biological heterogeneity within the data.
• Qualitative and quantitative analyses confirmed the effectiveness of DeepComBat.

Conclusions:

• DeepComBat offers a powerful approach for harmonizing multi-batch neuroimaging data.
• The method enhances data quality, reproducibility, and analytical reliability.
• DeepComBat represents a significant advancement in statistically motivated deep learning harmonization.