ABSTRACT
Chest X-ray (CXR) is a common imaging modality for examination of pneumonia. However, some pneumonia signs which are visible in CT may not be clearly identifiable in CXR. It is challenging to create a good ground truth for positive pneumonia cases based on CXR images especially for cases with small pneumonia lesions. In this paper, we propose a novel CT-based CXR synthesis framework, called ct2cxr, to perform data augmentation for pneumonia classification. Generative Adversarial Networks (GANs) were exploited and a customized loss function was proposed for model training to preserve the target pathology and maintain high image fidelity. Our results show that CXR images generated through style mixing can enhance the performance of general pneumonia classification models. Testing the models on a Covid-19 dataset shows similar improvements over the baseline models. © 2022, Springer Nature Switzerland AG.