'Harmless' adversarial network harmonization approach for removing site effects and improving reproducibility in...

Area of Science:

• Neuroimaging
• Data Science
• Biomarker Discovery

Background:

• Multi-site collaborations are crucial for overcoming small sample sizes in neuroimaging.
• Site-specific factors introduce confounds that reduce cross-site reproducibility.
• Robust biomarkers require harmonized datasets from diverse sources.

Purpose of the Study:

• To develop a method for removing confounds and enhancing cross-site task performance in neuroimaging.
• To improve the reproducibility of findings in large-cohort neuroimaging studies.
• To introduce a novel Maximum Classifier Discrepancy Generative Adversarial Network (MCD-GAN) for data harmonization.

Main Methods:

• Proposed a Maximum Classifier Discrepancy Generative Adversarial Network (MCD-GAN).
• Combined generative models with maximum discrepancy theory for data harmonization.
• Visualized MCD-GAN mechanisms using simulated data.
• Compared MCD-GAN with existing methods (ComBat, cycle-GAN) on the ABCD dataset.

Main Results:

• MCD-GAN effectively improved cross-site gender classification performance.
• Demonstrated successful harmonization of site effects in neuroimaging data.
• Outperformed state-of-the-art methods in improving cross-site task performance.

Conclusions:

• MCD-GAN offers an efficient solution for removing site effects in large-cohort neuroimaging studies.
• The proposed framework facilitates cross-site reproducibility for various classification and prediction tasks.
• This method holds promise for enhancing the reliability of neuroimaging biomarkers.