Key result
DLIR-H reduced in-stent lumen stenosis measurement error by ~15% versus ASIR-V on CCTA, without improving stent-level classification accuracy.
Why the study?
The comparative performance of high-strength deep learning image reconstruction (DLIR-H) versus adaptive statistical iterative reconstruction (ASIR-V) for evaluating image quality, stenosis measurement accuracy, and classification performance on CCTA of stented coronary vessels against invasive coronary angiography needed investigation.
Does high-strength deep learning image reconstruction (DLIR-H) improve image quality, stenosis measurement accuracy, and classification performance compared to adaptive statistical iterative reconstruction (ASIR-V) on coronary CT angiography in patients with stented coronary vessels?
Observational (n=20)
Blinded readers
No
Does high-strength deep learning image reconstruction (DLIR-H) improve image quality, stenosis measurement accuracy, and classification performance compared to adaptive statistical iterative reconstruction (ASIR-V) on coronary CT angiography in patients with stented coronary vessels?
In an exploratory study of stented coronary vessels on CCTA, high-strength deep learning image reconstruction modestly reduced in-stent lumen stenosis measurement error and improved subjective noise compared to iterative reconstruction, though it did not improve overall stenosis classification accuracy.
Answers are grounded in this paper and cite their sources.
Hypothesis-generating for DLIR-H in stented CT; should not change practice without larger prospective validation.
Background/Objectives: To compare high-strength deep learning image reconstruction (DLIR-H) with adaptive statistical iterative reconstruction (ASIR-V) for coronary computed tomography angiography (CCTA) of stented vessels, evaluating image quality, stenosis measurement accuracy, and classification performance against invasive coronary angiography (ICA) as reference standard. Methods: This retrospective single-centre study included 20 patients (30 coronary vessels) with prior stent implantation and ≥50% stenosis on CCTA. Datasets were reconstructed using ASIR-V and DLIR-H. Four blinded readers assessed subjective image quality and in-stent lumen and vessel stenosis degree; Bayesian hierarchical models accounted for reader and vessel variability. Results: DLIR-H improved subjective noise ratings (posterior probability >99%) and reduced in-stent lumen stenosis measurement error by approximately 15% (estimate: −0.16; 90% credible interval: −0.30 to −0.01; probability: 96%), with no clear evidence of an effect on native-vessel stenosis measurement error (probability: 72%). False-positive stent-level classifications (CAD-RADS ≥4) decreased from 10 to 4 reader assessments, whereas false-negative classifications increased from 2 to 4. Only one stent was reference-positive on ICA. Inter-reader agreement was similar for both reconstructions. Conclusions: In this exploratory study, DLIR-H was associated with a modest reduction in in-stent lumen stenosis measurement error, with no benefit for native vessel stenosis. At the stent level, DLIR-H shifted classifications towards CAD-RADS ≤3 but did not improve classification accuracy. Inter-reader agreement was limited with both reconstructions. Given the small sample and the single stent with ≥70% stenosis on ICA, these findings should be regarded as preliminary and require confirmation before informing routine clinical use.
No takes yet. Share an insight, caveat, or question.
Zdanowicz-Ratajczyk et al. (2026) conducted an observational in Prior coronary stent implantation with ≥50% stenosis on CCTA (n=20). High-strength deep learning image reconstruction (DLIR-H) vs. Adaptive statistical iterative reconstruction (ASIR-V) was evaluated on In-stent lumen stenosis measurement error (Estimate -0.16, 95% CI -0.30 to -0.01 (90% CrI), p=posterior probability 96%). DLIR-H reduced in-stent lumen stenosis measurement error by ~15% versus ASIR-V on CCTA, without improving stent-level classification accuracy.
Synapse has enriched 10 related papers. Consider them for comparative context: