Deep Learning Imaging Reconstruction Algorithm for Carotid Dual Energy CT Angiography: Opportunistic Evaluation of Cervical Intervertebral Discs-A Preliminary Study

Abstract

Thus, the aim of this study is to evaluate the performance of deep learning imaging reconstruction (DLIR) algorithm in different image sets derived from carotid dual-energy computed tomography angiography (DECTA) for evaluating cervical intervertebral discs (IVDs) and compare them with those reconstructed using adaptive statistical iterative reconstruction-Veo (ASiR-V). Forty-two patients who underwent carotid DECTA were included in this retrospective analysis. Three types of image sets (70 keV, water-iodine, and water-calcium) were reconstructed using 50% ASiR-V and DLIR at medium and high levels (DLIR-M and DLIR-H). The diagnostic acceptability and conspicuity of IVDs were assessed using a 5-point scale. Hounsfield Units (HU) and water concentration (WC) values of the IVDs; standard deviation (SD); and coefficient of variation (CV) were calculated. Measurement parameters of the 50% ASIR-V, DLIR-M, and DLIR-H groups were compared. The DLIR-H group showed higher scores for diagnostic acceptability and conspicuity, as well as lower SD values for HU and WC than the ASiR-V and DLIR-M groups for the 70 keV and water-iodine image sets (all p < .001). However, there was no significant difference in scores and SD among the three groups for the water-calcium image set (all p > .005). The water-calcium image set showed better diagnostic accuracy for evaluating IVDs compared to the other image sets. The inter-rater agreement using ASiR-V, DLIR-M, and DLIR-H was good for the 70 keV image set, excellent for the water-iodine and water-calcium image sets. DLIR improved the visualization of IVDs in the 70 keV and water-iodine image sets. However, its improvement on color-coded water-calcium image set was limited.

Keywords: Deep learning; Dual-energy computed tomography; Image reconstruction; Intervertebral disc.

Fig. 1

Axial CT images of a 66-year-old man with central disc herniation at the C2/3 level. The 70 keV, water (iodine), and color-coded virtual noncalcium (VNCa) images were reconstructed using iterative reconstruction-Veo (ASiR-V), deep learning image reconstruction at medium level (DLIR-M), and high level (DLIR-H). The deep learning image reconstruction algorithms effectively reduced image noise of grayscale CT images compared to ASiR-V

Fig. 2

Box and dot plots compared image quality (diagnostic acceptability and conspicuity) of 70 keV (a, g), water (iodine) (b, h) and color-coded virtual noncalcium (VNCa) (c, i) image sets among different reconstruction algorithm, and iterative reconstruction-Veo (ASiR-V) (d, j), deep learning image reconstruction algorithm at medium level (DLIR-M) (e, k), at high level (DLIR-H) (f, i) among different image sets by both readers (a–f for reader 1; g–i for reader 2). **** indicates p < 0.001, *** indicates p < 0.01, and ns indicates means result was not statistical significance