Deep learning reconstruction dual-energy computed tomography for gastrointestinal system tumors: low-kiloelectron...

To qualitatively and quantitatively compare dual-energy computed tomography (DECT)-derived 55 keV virtual monochromatic images (VMIs) using deep learning image reconstruction (DLIR) with 70 keV VMIs using adaptive statistical iterative reconstruction (ASiR-V) in the Gastrointestinal (GI) system tumors and to investigate whether DLIR can influence the measured iodine parameters.

This prospective, observational study included 47 patients with GI system tumors who underwent DECT. Qualitative and quantitative assessment was performed on 70 keV VMIs using ASiR-V40%, and 55 keV VMIs using DLIR at low (DLIR-L), medium (DLIR-M), and high strength (DLIR-H). Quantitative evaluation included calculation of the contrast-to-noise ratio (CNR) and the signal-to-noise ratio (SNR). Overall image quality, lesion conspicuity, and diagnostic confidence were evaluated qualitatively by 2 radiologists. Iodine parameters of tumor lesions were compared among four groups of algorithms.

There were no significant differences in the image noise between DLIR-M and ASiR-V40%, whereas image noise was significantly lower in DLIR-H than ASiR-V40% and DLIR-M (both phases, P < 0.001). DLIR-M and DLIR-H had significantly higher CNR and SNR compared with ASiR-V40% on both phases (all P ≤ 0.001). When compared to ASiR-V40%, DLIR-M and DLIR-H had similar or improved qualitative assessment metrics. DLIR-L yielded the highest image noise and the lowest overall image quality score among these algorithms. Lastly, the measured iodine parameters were equivalent among different algorithms (all P > 0.1).

Abdominal DECT at 55 keV with DLIR (-M and -H) provided qualitative and quantitative performance for GI tumor evaluation comparable to or exceeding that of ASiR-V 40% at 70 keV.