Low-dose Covid-19 CT Imaging: Noise-to-noise Versus Supervised Deep Learning-based Denoising

Abstract

Low-dose CT imaging has been proposed to reduce radiation dose to patients, and deep learning-based algorithms have been used to restore the original quality and underlying/crucial structures in low-dose CT pictures. The extra noise from low-dose CT imaging may affect the accuracy of diagnosis due to the modest expression of Covid-19 infections in the lung volume. Due to the significant noise signals, deep learning-based techniques proposed to reduce increased noise levels in low-dose CT images may generate artifacts and/or pseudo structures in the lung volume. The performance of a supervised deep learning network to estimate full-dose CT images from low-dose counterparts is investigated in this paper. Moreover, the performance of the supervised deep learning model is compared with an unsupervised method based on the noise-to-noise technique. These models were evaluated for two low-dose levels of 10% and 5% of standard imaging for Covid-19 patients. The quality of predicted full-dose CT images was quantified in terms of structural similarity index (SSIM), root mean square error (RMSE), and peak signal-noise ratio (PSNR). For qualitative assessment of Covid-19 pneumonia within the lung in the predicted CT images, a 5-grade (1: uninterruptable and 5: excellent) scoring scheme was adopted. Regarding the quantitative analysis, both supervised and noise-to-noise models exhibited similar performance at 10% low-dose level (RMSE=0.07±0.02). However, at 5% low-dose level, the supervised model outperformed the noise-to-noise model with RMSE=0.09±0.02 vs. RMSE=0.11±0.02 (p-value<0.02). Regarding lesion detectability, both supervised and noise-to-noise models obtained a mean score of 4.2±0.3 (good) at 10% low-dose level. At 5% dose level, despite the superior performance of the supervised model, the qualitative evaluation demonstrated similar performance of these two models (3.9±0.3). This observation is due to the generation of pseudo-structures by the supervised model at high noise levels (5% low-dose level). At high noise levels, despite the superior quantitative performance, the supervised model exhibited more outliers (pseudo lesions).