CGDM-GAN: An Adversarial Network Approach with Self-supervised Learning for Site Effect Removal

CGDM-GAN: An Adversarial Network Approach with Self-supervised Learning for Site Effect Removal

연구 분야: Artificial Intelligence

학회: 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)

Imaging data collected from different sites is difficult to pool together due to unwarranted variations introduced by different acquisition protocols or scanners. Data harmonization is an effective way to mitigate site-specific bias while preserving the intrinsic image properties, thereby increasing the sample size and enhancing the generalization of models. Although various harmonization methods exist, their performance on specific tasks is often unsatisfactory. Here, we proposed a novel approach, CGDM-GAN, by combining the advantages of generative models, maximum discrepancy theory, and gradient discrepancy minimization with self-supervised learning to harmonize site effects and improve cross-site classification performance. The proposed CGDM-GAN was successfully conducted on synthetic dataset, and further validated on in-house and ABCD datasets, outperforming three data harmonization methods, including ComBat, CycleGAN, and MCD-GAN, suggesting its potential for removing site effects and improving cross-site neuroimaging classification.