This single-center retrospective study evaluated the diagnostic performance of a recently updated DLM (CorEx-2.0) for quantifying coronary stenosis on CCTA, using a binary CAD-RADS classification threshold of 70% stenosis (CAD-RADS 0-3 versus CAD-RADS 4-5) per patient and a 6-group CAD-RADS classification (CAD-RADS 0-5) per-vessel. CorEx-2.0 identified all patients with severe stenosis (CAD-RADS ≥ 4), yielding 100% sensitivity and 94% accuracy for the binary classification, while also demonstrating good performance using the 6-group classification per vessel (weighted kappa values > 0.70), both compared separately with two independent expert CCTA readers as references.

In a previous study, the pre-updated model (CorEx-1.0) achieved 96% accuracy for detecting ≥ 50% stenosis at patient level among 53 patients with predominantly non-obstructive CAD (72%) [20]. However, CorEx-1.0 demonstrated lower accuracy (84%) in 217 patients presenting with acute chest pain [25]. The updated model (CorEx-2.0) recently achieved 82% accuracy for detecting ≥ 50% stenosis in a high-risk population (62.7% obstructive CAD on CCTA) during transcatheter aortic valve replacement workup [22]. These previous studies were limited to the binary (CAD-RADS 0-2 versus CAD-RADS 3-5) or 3-group CAD-RADS classification (CAD-RADS 0, 1-2, or 3-4-5). In our validation study, the CAD-RADS distribution within the study population was more balanced, and a 6-group CAD-RADS classification per vessel was assigned by both CorEx-2.0 and the expert readers. This 6-group analysis is a more accurate method compared with the binary or 3-group analysis used in the previous studies. We further focused on the analysis of the binary classification (CAD-RADS 0-3 versus CAD-RADS 4-5) at patient level, which represents a threshold for medical therapy with potential additional interventional treatment (≥ 70% stenosis) [8]. Nearly two-thirds of the patients with ≥ 70% stenosis (CAD-RADS ≥ 4) undergo revascularization [26]. The performance characteristics of CorEx-2.0 in our study compare similarly or favorably to those reported in other studies [17,18,19, 23, 27, 28]. However, it should be mentioned that a direct comparison with the listed studies is not feasible, as different study populations with varying sizes and CAD-RADS distributions were used. One single-center retrospective study compared an in-house DLM with expert readers in 288 patients and demonstrated 71% accuracy for the binary classification (threshold 50% stenosis) at patient level [27]. Zreik et al achieved 85% accuracy for this binary analysis in 65 patients by comparing their research method with an expert reader [23]. These studies did not perform a binary classification (threshold 70% stenosis) or 6-group CAD-RADS analysis. Regarding this binary classification (threshold 50%) at patient level, CorEx-2.0 performed better than or similarly to their models, yielding an accuracy of 82% versus reader 1 and 90% versus reader 2. A multicenter study by Choi et al enrolled 232 patients and used a commercial AI model that identified all patients with severe stenosis (≥ 70%) versus consensus reading, similar to the performance of CorEx-2.0 [17]. On a per-vessel basis, the model showed weighted kappa values of 0.69, 0.57, and 0.67 versus individual expert readers for the 6-group CAD-RADS classification, which is lower than the performance of CorEx-2.0 versus expert readers (weighted kappa of 0.71 versus reader 1 and 0.73 versus reader 2). The weighted kappa of 0.72 at vessel level compared with consensus reading, as demonstrated by Choi et al, is comparable to our results. Notably, only 5% of the patients and only 1.5% of the vessels were classified as CAD-RADS ≥ 4 (≥ 70% stenosis), indicating a low disease prevalence in their study population. This may limit the ability to adequately assess the safety of this tool in detecting severe stenosis (≥ 70%). Another study evaluated the same AI model compared with invasive quantitative coronary angiography in a cohort of 303 patients with a higher disease prevalence (≥ 70% stenosis in 39% of patients and 19% of vessels), demonstrating a per-patient sensitivity and accuracy of 94% and 86%, respectively, for detecting ≥ 70% stenosis [28]. In our study, 28% of the patients and 15% of the vessels were classified as CAD-RADS ≥ 4 (≥ 70% stenosis) by expert reading, and CorEx-2.0 yielded 100% sensitivity and 94% accuracy. However, these results are difficult to compare, as we used expert CCTA readers as references. Moreover, their study did not conduct a 6-group CAD-RADS analysis per vessel. This 100% detection of ≥ 70% stenosis at patient level compared with expert readers, as achieved by CorEx-2.0, was also established by Lin and colleagues using a novel research DLM in 50 patients [18]. At vessel level, this DLM demonstrated a Cohen’s kappa of 0.78 compared to expert CCTA interpretation using a 5-group CAD-RADS classification (CAD-RADS 1-5). However, since a 6-group CAD-RADS analysis was not performed, it is difficult to compare their results with the weighted kappa values of 0.71 (versus reader 1) and 0.73 (versus reader 2) per vessel using the 6-group classification presented in our study. Recently, Van Herten et al designed a fully automated deep learning-based method for plaque segmentation and CAD-RADS grading [19]. This research model achieved a linearly weighted kappa of 0.71 for the 6-group CAD-RADS classification per patient in an external test set of 658 patients, which aligns with our results for the 6-group classification at patient level (weighted kappa of 0.67 versus reader 1 and 0.74 versus reader 2). However, their test set had a significantly lower disease prevalence (only 4.6% of patients with CAD-RADS ≥ 4), which may have influenced the model’s performance. Furthermore, despite good performance, their method tended to overestimate CAD severity [19]. These potential false-positive results may lead to unnecessary downstream testing [29], raising concerns about the significantly increased utilization of CCTA. With respect to CorEx-2.0, a marginal underestimation compared with human assessment was observed, yet the model proved to be safe for identifying patients with severe stenosis compared with a fully independent expert reader. These findings underscore the potential of CorEx-2.0 to flag patients at risk of severe obstructive CAD.

This study had several limitations. First, the retrospective single-center design and the exclusion of patients with non-diagnostic examinations or a history of coronary interventions could have resulted in selection bias. Second, the CAD-RADS distribution in the study population may not reflect local clinical practice; this might have influenced the performance of CorEx-2.0, and this holds implications for direct comparison with other studies for local performance and reproducibility of the presented results. Therefore, further studies with larger patient cohorts are needed to evaluate CorEx-2.0 performance in a more standard outpatient cardiology CCTA population. Third, one of the expert readers trained the DLM, which could induce bias for a one-to-one comparison with CorEx-2.0. However, the other expert reader (reader 1) showed similar performance and was not involved in the DLM training and development. Fourth, only the three main vessels (LAD, RCA, Cx) were included in this study. Finally, invasive coronary angiography was not performed routinely, and therefore, we could not adequately assess the performance of both expert readers in this subset of cases.