Image denoising with conditional generative adversarial networks (CGAN) in low dose chest images

Abstract

Recently, low-dose medical imaging attracts a significant interest owing to the harmfulness of ionized radiations including X-rays. However, when the radiation dose is reduced during the medical image acquisition process, a significant quantum noise is commonly generated. The purpose of this study is to develop a deep-learning-based image-denoising method for low-dose chest imaging, which is a commonly performed medical imaging for diagnosis. Conditional generative adversarial networks (CGANs) were used in the development of the denoising algorithm. In order to train the deep-learning model, we used the SPIE American-Association-of-Physicists-in-Medicine lung-CT-challenge, and Lung-Image-Database-Consortium and Image-Database-Resource-Initiative databases. The obtained image demonstrated that the proposed method achieved an excellent image quality by removing the noise component. Compared with conventional denoising algorithms such as the total-variation (TV) minimization and non-local means (NLM), the proposed method exhibited a superior quality of the obtained images. Losses of image information, detrimental in medical diagnoses, occurred in the medical images obtained using conventional denoising algorithms. Unlike the conventional denoising algorithms, the proposed algorithm restored the corrupted image resolution owing to image noise. The quantitative evaluation through structure similarity index measure (SSIM) demonstrated the superiority of the proposed method over the conventional methods. The SSIM of the proposed method was improved by 1.5 and 2.5 times, compared to those of the NLM and TV methods, respectively. Therefore, we developed a denoising algorithm for medical imaging with CGAN, which is one of the latest deep-learning structures, for low-dose chest images. The proposed denoising method is expected to contribute to the improvement of image quality and reduction of the patient dose.