Low-dose chest X-ray image super-resolution using generative adversarial nets with spectral normalization

Abstract

Chest X-ray (CXR) imaging is one of the most widely-used and cost-effective technology for chest screening and diagnosis of Pulmonary diseases. An always concerned improvement about CXR is to reduce X-ray radiation while achieving ultra-high quality imaging with fine structural details since CXR involves ionizing radiation and tolerance of different populations. In this paper, we present a supervised generative adversarial nets approach to accurately recover high-resolution (HR) CXR images from low-resolution (LR) counterparts while keep pathological invariance. Specifically, the auxiliary label information is introduced to constrain the feature generation to attack the potential risk of pathological variance. Then, spectral normalization is designed to control the performance of discriminative network with the guarantee of theoretical demonstration in controlling Lipschitz bound of discriminator. Results from quantitative and qualitative evaluations demonstrate that our method delivers more authentic improvement for CXR super-resolution (SR) compared to recent state-of-the-art methods. The proposed method has outperformed average 13.0%, 12.2% in FSIM and 13.7%, 12.5% in MSIM on two datasets, respectively. Besides, the index of generative performance GAN-train and GAN-test have achieved average increment 9.3% and 10.5% on CXR2 dataset. Subjective evaluation on SR CXR has outperformed average score 0.425 and 0.525 in terms of pathological invariance and acceptability, respectively.