Purpose or Case Report: We sought to evaluate the quality and interpretability of non-contrast chest CT images in pediatric patients for the diagnosis of airway disease at two low-dose levels with and without denoising using a deep convolutional neural network (CNN), with one of the dose levels approaching that in a standard chest radiograph.
Methods & Materials: 41 pediatric patients undergoing low-dose non-contrast chest CT for “airway disease” were included. CT scans were performed using a Siemens Force, 3.2 pitch, Sn100kV, and 250 quality reference mAs with contiguous inspiratory and expiratory images (0.75-mm slice thickness) acquired using filtered back projection (FBP, Qr54 kernel) and iterative reconstruction (IR, strength setting of 3). A validated noise insertion method was used to simulate 20% dose data for each case. A deep CNN model was trained and validated on 10 cases and then applied to the remaining 31 cases. Three CAQ-certified pediatric radiologists evaluated the 31 tests cases under 4 conditions: (1) routine low-dose (RD) + IR; (2) RD (FBP) + CNN; (3) 20% dose (TD) + IR; and (4) TD (FBP) +CNN, in a randomized, blinded fashion. Images were graded for spatial resolution (1-5, 5=excellent) and noise (1-4, 4=no noise). Objective assessment of spatial resolution was performed through visualization of the secondary pulmonary lobule (SPL) (1=seen, 2=not seen).
Results: Of the 31 patients evaluated [mean age 10.8 years, range, 0.17-17], 4 patients could not breath-hold and were scanned in only one phase of the breathing cycle. The volume CT dose index (CTDIvol) was 0.52±0.22 mGy per scan in the original RD exam (estimated effective dose per scan was 0.46±0.21 mSv). RD+CNN images showed the highest spatial resolution (3.6±1.1), followed by TD+CNN (3.4±1) and RD+IR (3.4±1), and TD+IR (2.6±0.7). RD+CNN showed the least noise (3.1±1), followed by TD+CNN (2.9±0.7), RD+IR (2.6±0.7), and TD+IR (2.1±0.4). There were statistically significant differences between reconstruction methods for both the subjective assessment of spatial resolution (p<0.001) and noise (p<0.001). The secondary pulmonary lobule was seen most consistently on the RD+CNN images (91% seen), followed by TD+CNN (90% seen), RD+IR (85%) and TD+IR (84% seen).
Conclusions: Deep CNN denoising improved spatial resolution and noise of routine low-dose images and ultra-low dose images corresponding to <0.10 mSv (dose level approaching that of a 2-view chest x-ray), which proved to be superior to RD images constructed with IR.