Related Experiment Video
Updated: May 24, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
451
CGDM-GAN: An Adversarial Network Approach with Self-supervised Learning for Site Effect Removal.
Summary
We developed CGDM-GAN, a novel method to harmonize neuroimaging data across different sites. This approach effectively reduces site-specific bias, improving cross-site classification performance and enhancing model generalization.
Area of Science:
- Neuroimaging
- Medical Data Analysis
- Artificial Intelligence
Background:
- Multi-site imaging data presents challenges due to variations in acquisition protocols and scanners.
- Site-specific biases can hinder data pooling and model generalization in neuroimaging studies.
- Existing data harmonization methods often show unsatisfactory performance on specific tasks.
Purpose of the Study:
- To introduce a novel approach, CGDM-GAN, for harmonizing neuroimaging data across different sites.
- To mitigate site-specific bias while preserving intrinsic image properties.
- To improve cross-site classification performance in neuroimaging.
Main Methods:
- Proposed CGDM-GAN, combining generative models, maximum discrepancy theory, and gradient discrepancy minimization.
- Integrated self-supervised learning to enhance harmonization capabilities.
- Validated the method on synthetic, in-house, and ABCD datasets.
Main Results:
- CGDM-GAN successfully harmonized site effects in neuroimaging data.
- The method demonstrated superior performance compared to ComBat, CycleGAN, and MCD-GAN.
- Improved cross-site classification accuracy was observed using the proposed approach.
Conclusions:
- CGDM-GAN is a promising method for removing site effects in neuroimaging.
- The approach enhances the generalization of models by enabling effective data pooling.
- CGDM-GAN offers a potential solution for improving cross-site neuroimaging classification tasks.
