Application Value of Deep Learning in Spectral Computed Tomography of Upper Abdomen

Objective: To compare the image quality of virtual monochromatic (VIM) imaging of upper abdominal computed tomography (CT) energy spectrum based on deep learning image reconstruction algorithm (DLIR) and adaptive statistical iterative algorithm (ASIR-V) and explore the application value of DLIR algorithm in CT energy spectrum imaging. Methods: Thirty-three cases of upper abdominal CT spectral enhanced scans were prospectively collected. The arterial and venous phase images were reconstructed using ASIR-V and DLIR, respectively. ASIR-V uses the 50% weight reconstruction method, whereas DLIR uses the deep learning medium strength (DL-M) and high strength (DL-H) reconstruction methods. The virtual single energy images of arterial and venous phases were extracted using three different reconstruction algorithms: 40 keV, 50 keV, 60 keV and 70 keV. Two radiologists scored the images of 12 groups in arterial phase and 12 groups in venous phase, collecting the image data for each group. The statistical package for the social sciences (SPSS) was used to statistically process the obtained data and score the results. Results: There were significant differences in CT value, standard deviation (SD), contrast-to-noise ratio (CNR), and signal-to-noise ratio (SNR) among the 12 groups of images in arterial phase and venous phase. In the arterial and venous phases, the SD value of liver, kidney, spleen and pancreas in the ASIR-V 40 keV group was the highest, and the SNR and CNR values were the lowest. The SD value of liver, kidney, spleen and pancreas in the DL-H 70 keV group was the lowest, and the SNR and CNR values were the highest. The DL-H 60 keV group had the highest subjective scores for the liver, kidney, spleen and pancreas. Conclusion: Compared with ASIR-V, DLIR upper abdominal CT spectral VIM images can more effectively reduce image noise, reduce image artifacts, and obtain better image quality. Low keV VIM CT spectral images were reconstructed using deep learning.