1. Low-contrast assessment on deep learning-based reconstruction of CT images.
- Author
-
Setiawan, Adiwasono M. B., Anam, Choirul, Sutanto, Heri, and Naufal, Ariij
- Subjects
MICROCOMPUTER workstations (Computers) ,IMAGE reconstruction ,COMPUTED tomography ,SIGNAL-to-noise ratio ,COMPUTER systems ,IMAGE reconstruction algorithms ,DEEP learning - Abstract
Deep learning image reconstruction (DLIR) is a most recent computed tomography (CT) image reconstruction method for reducing image noise while maintaining the spatial resolution of the image. It is well-understood that image noise affects the low-contrast detectability (LCD) of the images. We assessed the LCD of DLIR images using parameters of CT number accuracy, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR) and visual assessment on the low-contrast module in a Catphan phantom and compared it to images reconstructed with filtered-back projection (FBP). We scanned the Catphan 604 phantom (The Phantom Laboratory, USA) with a 512 multi-slice computed tomography (MSCT) scanner. The images were reconstructed with different strengths of the DLIR algorithm (i.e., low, medium, and high). SNR was measured by placing region of interest (ROI) in the center of CTP729 module (i.e. module for measuring CT number uniformity) of the Catphan phantom. CNR was measured by placing two ROIs in the center and a circular object with diameter of 15 mm on CTP730 module (i.e., module for measuring low-contrast detectability) of the Catphan phantom. Low contrast objects were also assessed visually on a 2D viewer software provided by Advantage Workstation Computer System (GE Healthcare, USA). The results show that the CT number for DLIR and FBP algorithms had no significant changes, but the noise levels on DLIR images were reduced compared to those from FBP. The noise levels were reduced to 2.9 HU (or 38.16%) for low-DLIR, 3.74 HU (or 49.21%) for medium-DLIR and 4.7 HU (or 61.84%) for high-DLIRh. Therefore, the SNR improved by 37.5, 49.35 and 61.94% for low-, medium-, and high-DLIR, respectively. The CNR were 0.68, 0.98, 1.2, 1.46 for FBP, low-DLIR, medium-DLIR and high-DLIR, respectively. Compared to CNR from FBP, the CNR improved by 30.61% for low-DLIR, 43.33% for medium-DLIR and 53.42% for high-DLIR. These results corresponded to our visual assessment that LCD is improved as the strength of the DLIR algorithm is increased. In conclusion, it is found that if the strength of the DLIR algorithm increases, then the noise level is reduced and the SNR and CNR are increased. Hence, the DLIR algorithm has a better conspicuity to detect low-contrast objects compared to those from FBP. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF