Related Experiment Video
Updated: Jul 9, 2025
03:37
Author Spotlight: Impact of Intergenic Interactions on Disease-Identifying Dark Biomarkers
Published on: March 1, 2024
767
Maximum Classifier Discrepancy Generative Adversarial Network for Jointly Harmonizing Scanner Effects and Improving
IEEE Transactions on Bio-Medical Engineering
|December 7, 2023
Summary
We developed a new framework, MCD-GAN, to remove MRI scanner effects and improve data harmonization across sites. This method enhances cross-scanner classification performance while preserving biological information for reproducible biomarker discovery.
Area of Science:
- Neuroimaging
- Machine Learning
- Biomarker Discovery
Background:
- Multi-site MRI studies face challenges with scanner-specific variations impacting reproducibility.
- Existing harmonization methods often fail to improve downstream task performance.
Purpose of the Study:
- To introduce a novel multi-scanner harmonization framework, MCD-GAN, for removing scanner effects.
- To preserve biological information crucial for downstream analytical tasks.
Main Methods:
- Proposed the Maximum Classifier Discrepancy Generative Adversarial Network (MCD-GAN).
- Utilized adversarial generative networks to maintain sample structural integrity.
- Incorporated a maximum classifier discrepancy module to regulate GANs using downstream tasks.
Main Results:
- MCD-GAN demonstrated superior performance compared to ComBat and CycleGAN on simulated and real-world data (ABCD dataset).
- The framework effectively improved cross-scanner classification accuracy.
- MCD-GAN preserved the anatomical layout of original MRI images.
Conclusions:
- MCD-GAN is the first generative model to integrate downstream tasks into the harmonization process.
- This approach offers a promising solution for enhancing cross-site reproducibility in MRI studies.
- Facilitates reproducible biomarker discovery for tasks like classification and regression.
