Deep learning reconstruction improves diagnostic confidence for coronary stent assessment at coronary CT angiography: a technical note

Abstract

Coronary stent evaluation remains one of the most demanding applications of coronary CT angiography (CCTA) because metallic struts, blooming artifact, image noise, and cardiac motion may reduce lumen interpretability. Reconstruction strategy therefore plays a critical role in determining diagnostic confidence. In this retrospective single-center technical note, we compared deep learning reconstruction (DLR), hybrid iterative reconstruction (HIR), and model-based iterative reconstruction (MBIR) for coronary stent assessment on a 320-row whole-heart CT platform. Twenty patients with 35 evaluable coronary stents were included. For each examination, three reconstruction sets were generated using vendor-available HIR, MBIR, and DLR algorithms. Stents were classified as proximal or distal according to coronary location. An experienced cardiothoracic radiologist assessed diagnostic confidence on a 5-point Likert scale, considering in-stent lumen visibility, evaluation of relevant residual or recurrent stenosis, and delineation of stent edges. Ordinal data were analyzed using a cumulative link mixed model with reconstruction type and stent location as fixed effects and patient/stent clustering as random effects. DLR achieved significantly higher diagnostic confidence than both HIR and MBIR (both p < 0.001), whereas no significant difference was observed between HIR and MBIR (p = 0.957). The superiority of DLR was maintained in both proximal and distal stents. These preliminary findings suggest that DLR may offer a practical advantage for routine coronary stent evaluation at CCTA, although larger multicenter studies with invasive reference standards are needed to confirm its impact on diagnostic accuracy and clinical decision-making.