The Feasibility of Deep Learning–Based Reconstruction for Low-Tube-Voltage CT Angiography for Transcatheter Aortic Valve Implantation

Abstract

Objective The purpose of this study is to evaluate the efficacy of deep learning reconstruction (DLR) on low-tube-voltage computed tomographic angiography (CTA) for transcatheter aortic valve implantation (TAVI). Methods We enrolled 30 patients who underwent TAVI-CT on a 320-row CT scanner. Electrocardiogram-gated coronary CTA (CCTA) was performed at 100 kV, followed by nongated aortoiliac CTA at 80 kV using a single bolus of contrast material. We used hybrid-iterative reconstruction (HIR), model-based IR (MBIR), and DLR to reconstruct these images. The contrast-to-noise ratios (CNRs) were calculated. Five-point scales were used for the overall image quality analysis. The diameter of the aortic annulus was measured in each reconstructed image, and we compared the interobserver and intraobserver agreements. Results In the CCTA, the CNR and image quality score for DLR were significantly higher than those for HIR and MBIR (P < 0.01). In the aortoiliac CTA, the CNR for DLR was significantly higher than that for HIR (P < 0.01) and significantly lower than that for MBIR (P ≤ 0.02). The image quality score for DLR was significantly higher than that for HIR (P < 0.01). No significant differences were observed between the image quality scores for DLR and MBIR. The measured aortic annulus diameter had high interobserver and intraobserver agreement regardless of the reconstruction method (all intraclass correlation coefficients, >0.89). Conclusions In low tube voltage TAVI-CT, DLR provides higher image quality than HIR, and DLR provides higher image quality than MBIR in CCTA and is visually comparable to MBIR in aortoiliac CTA.