Towards ultra-low-dose CT for detecting pulmonary nodules using DenseNet.

Abstract

Low-radiation techniques should be used to detect and follow lung nodules on CT images, but reducing radiation dose to ultra-low-dose CT with submilliSievert dose level would drastically impede image quality and sensitivity for nodule detection. This study investigated the feasibility of using DenseNet to suppress image noise in ultra-low-dose CT for lung cancer screening. DenseNet was trained using input-label pairs from 1, 2, 4, and 6 patients. After training, the model was tested with chest CT from 14 patients that were not used in training process. Seven patients have solid nodules and 7 patients have subsolid nodules. Root mean square error (RMSE) and peak signal-to-noise ratio (PSNR) were calculated to quantify the difference between reference and test images. The contrast-to-noise ratio (CNR) between lung nodule and lung parenchyma was calculated to evaluate the target detectability of chest CT. Subjective image quality assessment was performed using 4-point ranking scale to evaluate the visual quality of CT images perceived by end user. Substantial improvements in RMSE and PSNR were observed after denoising. The lung nodules in denoised images could be distinguished more easily in comparison with those in the original ultra-low-dose CT, which is supported by the CNRs and subjective image quality scores. The comparison of intensity profiles for lung nodules demonstrated that the image noise in ultra-low-dose CT could be suppressed effectively after denoising without causing edge blurring or variation in Hounsfield unit (HU) values. A two-sample t-test revealed no statistically significant differences between full-dose CT and denoised ultra-low-dose CT in the evaluation of lung nodules, lung parenchyma, paraspinal muscle, or vertebral body. Since the linear no-threshold model suggests that no amount of ionizing radiation is entirely risk-free, the quest for further dose reduction remains a consistently important focus in radiology. Overall, our findings suggest that DenseNet could be a viable approach for reducing image noise in ultra-low-dose CT scans used for lung cancer screening.