Comparing two deep learning spectral reconstruction levels for abdominal evaluation using a rapid-kVp-switching dual-energy CT scanner.

Abstract

Purpose: Deep Learning Spectral Reconstruction (DLSR) potentially improves dual-energy CT (DECT) image quality, but there is a paucity of research involving human abdominal DECT scans. The purpose of this study was to comprehensively evaluate image quality by quantitatively and qualitatively comparing strong and standard levels of a DLSR algorithm. Optimal virtual monochromatic image (VMI) energy levels were also evaluated.

Methods: DECT scans of the abdomen/pelvis from 51 patients were retrospectively evaluated. VMIs were reconstructed at energy levels ranging from 35 to 200 keV using both standard and strong DLSR levels. For quantitative analysis, various abdominal structures were assessed using regions of interest, and mean signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) values were calculated. This was supplemented with a qualitative evaluation of VMIs reconstructed at 35, 45, 55, and 65 keV.

Results: The strong-level DLSR demonstrated significantly better SNR and CNR values (p < 0.0001) compared to standard-level DLSR across all structures. The optimal SNR was observed at 70 keV (p < 0.0001), while the optimal CNR was found at 65 keV (p < 0.0001). The average qualitative scores between standard and strong DLSR were significantly different at 45, 55, and 65 keV (p < 0.0001). There was a moderate level of agreement between observers (ICC = 0.427, p < 0.0001).

Conclusion: A DLSR set to a strong level significantly improves image quality compared to standard-level DLSR, potentially enhancing the diagnostic evaluation of abdominal DECT scans. In addition to achieving a very high SNR, 65 keV VMIs had the highest CNR, which differs from what is typically observed with traditional DECT using non-deep learning reconstruction approaches.