Deep-learning reconstruction for the evaluation of lumbar spinal stenosis in computed tomography

Abstract

To compare the quality and interobserver agreement in the evaluation of lumbar spinal stenosis (LSS) on computed tomography (CT) images between deep-learning reconstruction (DLR) and hybrid iterative reconstruction (hybrid IR). This retrospective study included 30 patients (age, 71.5 ± 12.5 years; 20 men) who underwent unenhanced lumbar CT. Axial and sagittal CT images were reconstructed using hybrid IR and DLR. In the quantitative analysis, a radiologist placed regions of interest within the aorta and recorded the standard deviation of the CT attenuation (i.e., quantitative image noise). In the qualitative analysis, 2 other blinded radiologists evaluated the subjective image noise, depictions of structures, overall image quality, and degree of LSS. The quantitative image noise in DLR (14.8 ± 1.9/14.2 ± 1.8 in axial/sagittal images) was significantly lower than that in hybrid IR (21.4 ± 4.4/20.6 ± 4.0) (P < .0001 for both, paired t test). Subjective image noise, depictions of structures, and overall image quality were significantly better with DLR than with hybrid IR (P < .006, Wilcoxon signed-rank test). Interobserver agreements in the evaluation of LSS (with 95% confidence interval) were 0.732 (0.712–0.751) and 0.794 (0.781–0.807) for hybrid IR and DLR, respectively. DLR provided images with improved quality and higher interobserver agreement in the evaluation of LSS in lumbar CT than hybrid IR.