GAN based ROI conditioned Synthesis of Medical Image for Data Augmentation

Abstract

Synthetic data is considered to be a promising solution for data privacy and scarcity. Some studies have shown that synthetic data generated from a simple GAN-based model enables privacy-preserving data sharing and data augmentation also in the medical imaging field. However, there are some limitations in applying this approach to real world situations: 1) Since generative models needs large amount of data to be trained, it is hard to be applied for small data situation. 2) Even after successfully training generative models, it is hard to guarantee which class the synthesized data corresponds to, especially for non-conditional generative models, so it needs to be re-labeled. Here, we propose GAN based ROI conditioned synthesis of medical Image for data augmentation. We used StyleGAN2 to learn the distribution of CXR and Bayesian image reconstruction for ROI-conditioned synthesis from the distribution. In the 4-class classification of CXRs showing normal, pneumonia, pleural effusion, and pneumothorax, using synthetic data for data sharing showed comparable performance to centralized learning, slightly better in terms of AUROC. Also, using synthetic data for augmentation, the accuracy and AUROC showed up to 6.5% and 8.9% increases, respectively.