Deep Residual Neural Network-based Standard CT Estimation from Ultra-Low Dose CT Imaging for COVID-19 Patients

Abstract

Chest computed tomography (CT) imaging was widely used for diagnosis and staging of severe acute respiratory syndrome coronavirus (SARS-CoV-2). CT can be utilized for initial diagnosis, severity scoring, serial monitoring, and patient status follow-up. For serial monitoring and follow-up, patients need to be scanned multiple times. The tendency in CT imaging is to minimize patient radiation dose. However, CT imaging is still considered as a high radiation dose modality. In this work, we proposed a deep residual neural network-based high quality (full dose) generation from ultra low-dose CT images to decrease the radiation dose for COVID-19 patients. In this multicenter study, we enrolled 1140 subjects with 313 PCR positive COVID-19 patients. The ultra low-dose CT images were analytically simulated, and then a deep residual neural network employed to estimate/generate full-dose images from the corresponding ultra-low-dose images. Various quantitative parameters, including the root mean square error (RMSE), structural similarity index (SSIM), and qualitative visual scoring were implemented to evaluate image quality of the generated CT images. The mean CTDIvol for full-dose images were 6.5 Gy (4.16-10.5 mGy), while, the simulated low-dose images were intended for a mean CTDIvol of 0.72 mGy (0.66-1.02 mGy). Regarding the external validation set (test set), the RMSE declined from 0.16±0.06 to 0.08±0.02 in low-dose and predicted standard-dose CT images, while the SSIM metric increased from 0.89±0.07 to 0.97±0.01, respectively. The highest visual scores (out of 5) were achieved by full-dose images (4.72±0.57) and predicted full-dose images (4.42±0.08). Conversely, ultra-low-dose images received the lowest score (2.78±0.9). In can be concluded that the proposed deep residual network improved image quality of ultra low-dose CT images, thus recovering their diagnostic value.