Abstract
Background and Objectives: We attempted to determine the optimal radiation dose to maintain image quality using a deep learning application in a physical human phantom. Materials and Methods: Three 5 × 5 × 5 mm(3) uric acid stones were placed in a physical human phantom in various locations. Three tube voltages (120, 100, and 80 kV) and four current-time products (100, 70, 30, and 15 mAs) were implemented in 12 scans. Each scan was reconstructed with filtered back projection (FBP), statistical iterative reconstruction (IR, iDose), and knowledge-based iterative model reconstruction (IMR). By applying deep learning to each image, we took 12 more scans. Objective image assessments were calculated using the standard deviation of the Hounsfield unit (HU). Subjective image assessments were performed by one radiologist and one urologist. Two radiologists assessed the subjective assessment and found the stone under the absence of information. We used this data to calculate the diagnostic accuracy. Results: Objective image noise was decreased after applying a deep learning tool in all images of FBP, iDose, and IMR. There was no statistical difference between iDose and deep learning-applied FBP images (10.1 ± 11.9, 9.5 ± 18.5 HU, p = 0.583, respectively). At a 100 kV-30 mAs setting, deep learning-applied FBP obtained a similar objective noise in approximately one third of the radiation doses compared to FBP. In radiation doses with settings lower than 100 kV-30 mAs, the subject image assessment (image quality, confidence level, and noise) showed deteriorated scores. Diagnostic accuracy was increased when the deep learning setting was lower than 100 kV-30 mAs, except for at 80 kV-15 mAs. Conclusions: At the setting of 100 kV-30 mAs or higher, deep learning-applied FBP did not differ in image quality compared to IR. At the setting of 100 kV-30 mAs, the radiation dose can decrease by about one third while maintaining objective noise.