Comparison of Super-Resolution Deep-Learning Reconstruction and Hybrid Iterative Reconstruction for Coronary Stent Assessment on CTA: A Prospective Multicenter Study.

Abstract

<b>Background:</b> Conventional deep-learning reconstruction (DLR) CT methods provide noise reduction but limited improvements in spatial resolution. <b>Objective:</b> To compare a novel super-resolution DLR (SR-DLR) algorithm and a hybrid iterative reconstruction (HIR) algorithm in terms of stent visualization and diagnostic performance for in-stent restenosis on coronary CTA using invasive coronary angiography (ICA) as the reference. <b>Methods:</b> This prospective study enrolled patients undergoing coronary CTA at 11 centers in China from September 2023 to November 2024. Participants underwent coronary CTA using a normal-resolution 320-row MDCT scanner with reconstruction of HIR and SR-DLR images. The final study sample included participants with a coronary stent who underwent ICA within 2 months after CTA. CNRstent, stent-lumen attenuation increase ratio (SAIR), and stent edge sharpness were calculated. Two radiologists (R1, R2) independently assessed stents for subjective image quality (1-5 scale; 5=highest quality), diagnostic confidence (1-5 scale; 5=greatest confidence), and in-stent restenosis (defined as ≥50% stenosis). HIR and SR-DLR were compared using ICA as the reference for in-stent restenosis. <b>Results:</b> The analysis included 73 participants (62 men, 11 women; mean age, 65.5±10.3 years) with 110 stents. SR-DLR, compared with HIR, showed greater CNRstent (44.0±20.1 vs 33.6±15.7), lower SAIR (0.44±0.36 vs 0.71±0.55), and greater edge sharpness (469±261 vs 221±130 HU/mm) (all p<.05). SR-DLR, compared with HIR, showed greater subjective image quality (both readers: median, 4 vs 3) and greater diagnostic confidence (both readers: median, 4 vs 3) (all p<.001). SR-DLR, compared with HR, showed greater accuracy for in-stent restenosis among all stents in per-stent analysis (R1: 89.1% vs 79.1%; R2: 88.2% vs 77.3%) and per-patient analysis (R1: 89.0% vs 79.5%; R2: 87.7% vs 72.6%) and among stents with diameter ≤3.0 mm in per-stent analysis (R1: 91.5% vs 81.4%; R2: 93.2% vs 81.4%) and per-patient analysis (R1: 90.7% vs 76.7%; R2: 93.0% vs 76.7%). <b>Conclusions:</b> SR-DLR, compared with HIR, yielded objective and subjective improvements in stent visualization, including differences attributable to reduced blooming artifacts, and greater diagnostic performance for in-stent restenosis. <b>Clinical Impact:</b> The novel SR-DLR algorithm improves the performance of coronary CTA obtained on conventional 320-row MDCT scanners and could support expanded use of CTA for stent evaluation.