We introduce a method to calculate the Stenosis Regression Index (SRI) and lesion length across the full clinical spectrum, achieving an MAE of 12.4% without biased data filtering.

Our calcium scoring assistance tool was validated against five expert cardiologists, achieving an intraclass correlation coefficient (ICC) of 0.84, demonstrating “excellent” inter-rater agreement.

By prioritizing negative predictive value (NPV of 0.98), our model ensures reliable exclusion of significant CAD, making it an ideal screening tool in resource-constrained or high-throughput clinical environments.

We advanced a model capable of differentiating vulnerable plaques from normal segments, enhancing the diagnostic depth of cMPR-based screening.

To ensure consistency in reporting, we categorized our data into four hierarchical units: patients, vessels, cMPR scans, and image slices (Table 1). A total of 1,715 patients at Daejeon St. Mary’s Hospital, South Korea, spanning from April 2014 to February 2022 were included, with each patient providing up to three major coronary vessels (LAD, LCX, RCA). For each vessel, a single curved Multiplanar Reconstruction (cMPR) scan was generated using the CT Cardiac Package from TeraRecon Inc., which was further discretized into 13 sequential cross-sectional slices for deep learning analysis. The dataset was split into development (training and validation) and test sets using a time-independent approach based on the year of examination (2014–2020 for development; 2021-2022 for testing). Within the development set, we implemented a strict patient-level separation with a 4:1 ratio to ensure that slices from the same patient did not appear in both training and validation subsets, thereby preventing data leakage. The exact distribution across each unit and subset is summarized in Table 2.

Inclusion criteria were: (1) individuals aged 18–75 years, (2) those with a history of coronary artery disease or myocardial infarction, and (3) having no contraindication for cMPR. Exclusion criteria were: (1) images with compromised quality due to motion artifacts, calcification, or inadequate contrast enhancement, and (2) patients with coronary artery bypass grafts or stents. The process involved the following steps:

Figure 1 outlines patient enrollment and data preparation process for studying coronary artery disease, detailing the workflow from initial patient inclusion to the creation of datasets for training and testing models designed to detect stenosis and vulnerable plaque. Between April 1, 2014 and February 28, 2022, a total of 1,715 patients with heart pain were consecutively enrolled in this study. From these patients, 5,145 vessels, including the right coronary artery (RCA), left anterior descending (LAD), and left circumflex (LCX) arteries, were extracted for analysis. However, 33 cases were excluded due to various reasons, including 19 stented patients, 11 instances of blurred images, one missing slice, and two file errors. The remaining dataset comprised 5,112 vessels from 1,704 patients. This dataset was then divided into a stenosis dataset and a vulnerable plaque dataset. Each of these datasets was further split into subsets for training, validation, and testing. Specifically, the dataset was standardized to 1,704 patients (5,112 vessels). While preliminary tuning involved a pilot subset of 500 scans, all final results reported herein utilize the full reconciled dataset: 2,140 vessels for training, 535 for validation, and 2,437 for the independent test set (Table 2). The test set was segregated based on time to ensure model performance over different periods to evaluate temporal generalizability of predictive models.

To improve the identification of vessels, stenosis, calcification, and vulnerable plaque in cMPR images, we developed two deep learning-based vessel segmentation models. We used the 2D nnU-Net framework (Wang et al., 2023) to build vessel segmentation models. The nnU-Net framework was selected based on its demonstrated superiority in medical image segmentation tasks and its self-configuring properties that automatically optimize preprocessing, network architecture, and training procedures published in Nature Methods, 2021 (Isensee et al., 2021). This architecture has shown consistent performance across diverse medical imaging applications, making it particularly suitable for our multi-task approach involving stenosis and vulnerable plaque segmentation. The framework automatically configured a U-Net architecture with encoder-decoder structure utilizing residual connections and deep supervision for optimal segmentation performance. The first segmentation model was a stenotic segmentation model (SSM) for segmenting vessels and stenosis. SSM was achieved through a sophisticated deep learning algorithm that delineated vessel walls within cMPR images. Based on segmentation results of SSM, we employed a technique to identify sharp changes in vessel diameter, signaling potential stenosis. SSM was also used as a vessel segmentation model for calcification analysis. The second model was a vulnerable plaque segmentation model (VPSM) for segmenting vessels and vulnerable plaques. This model was designed to detect vulnerable plaques in cMPR images, which are difficult to detect with CCTA. We appraised the effectiveness of these segmentation models employing a suite of metrics, including accuracy, dice similarity coefficient (DSC), sensitivity, and specificity on a pixel-level scale.

For detailed analysis of CCTA images, we utilized SSM to delineate and measure coronary vessels and any present stenosis. SSM, a sophisticated fully convolutional neural network, is designed to process a two-dimensional input of CCTA images, from which it generates a three-dimensional segmentation map. For coronary artery assessment, we employed a 2D nnU-Net architecture optimized for cross-sectional analysis. The model receives individual 2D cMPR slices (512 × 512 pixels) as discrete inputs rather than volumetric data. To reconcile these 2D predictions with the longitudinal structure of the vessel, the resulting 2D segmentation masks from the 13 sequential slices per vessel are orthogonally stacked to reconstruct a pseudo-3D representation of the arterial lumen and plaque. This approach ensures high-resolution feature extraction within each slice while maintaining the interpretability of the entire vessel segment. The final quantitative metrics, including diameter stenosis and lesion length, are derived from this reconstructed sequence, effectively bridging the gap between 2D slice-based inference and 3D clinical assessment. This map distinctly identifies and marks vessel structures and stenosis regions. The network was rigorously trained on a substantial dataset comprising 2,140 vessels (cMPR scans; 13 slices per vessel), which were gathered and meticulously annotated by an expert cardiologist. A validation set comprising 535 vessels was used for model selection and tuning. This diverse and comprehensive training set ensured that the model could accurately interpret a wide range of image presentations, accounting for variations in patient anatomy and image quality.

To thoroughly assess the performance and reliability of SSM, we conducted evaluations with a distinct test set that included 2,437 vessels (time-independent test set). The effectiveness of SSM’s segmentation capabilities was quantified using a comprehensive suite of unbiased metrics to ensure clinical robustness: (1) Dice Similarity Coefficient (DSC) for measuring spatial overlap between the model’s output and ground truth annotations; (2) Sensitivity and Specificity to evaluate the detection of stenotic lesions; and (3) Positive Predictive Value (PPV) and Negative Predictive Value (NPV) to assess the clinical reliability of the model in a screening context. Notably, we prioritized NPV and PPV over overall accuracy, as accuracy can be misleadingly high in medical imaging due to class imbalance between narrow vessels and the vast background. Furthermore, for stenosis quantification, Mean Absolute Error (MAE) and Bland–Altman analysis were employed to evaluate the agreement between the predicted SRI and expert measurements across the full, unfiltered test set.

The following is a concrete representation of the stenosis detection and quantification process (Figure 3). The first process involved performing stenosis segmentation using the SSM, which delineated both normal vessel areas and stenotic regions. Post-processing was then applied to the segmentation masks of vessels and stenosis to refine the segmentation results. This post-processing step removed clusters of normal vessels and stenosis that were 100 pixels or smaller to address imperfections in the segmentation mask. Our statistical analysis indicated that less than 5% of clusters identified by the SSM were larger than 100 pixels. After post-processing, the second process defined various lengths for stenosis regression. This involved defining the proximal and distal segments based on the lengths of stenosis segmented by the SSM. In the third process, we employed the skeletonization method (Zhou and Toga, 1999) on the segmentation mask produced by the SSM. This enabled us to estimate the length of the stenosis along the skeleton. We calculated the change in vessel thickness as the full width at half maximum (FWHM) (Varma et al., 2004) by tracing the length of the normal along the skeletonized curved line. This method allowed us to determine the relative diameter of the cardiac vessel, as illustrated in Process 3 of Figure 3. For stenotic lesions, we extracted the minimum value, while for the proximal and distal regions, we used the median values as representative measures. The fourth process involved the quantification of coronary stenosis, including the assessment of stenosis severity and diagnosis. We calculated the stenosis severity as the percentage of vessel narrowing using the equation shown at the bottom of Figure 3.

We have standardized our stenosis detection and classification logic to align with the CAD-RADS™ 2.0 guidelines, ensuring clinical consistency as follows.

Category 0 (documented absence of disease): 0% stenosis (no visible plaque or luminal narrowing).

Category 1 (minimal): 1–24% stenosis (minimal stenosis with insignificant plaque).

Category 2 (mild): 25–49% stenosis (mild stenosis; no functional significance expected).

Category 3 (moderate): 50–69% stenosis (moderate stenosis; the threshold for potential ischemia, often requiring functional assessment).

Category 4 (severe):

4A: 70–99% stenosis in 1 or 2 vessels.

4B: left main stenosis >50% or 3-vessel obstructive disease (≥70%).

Category 5 (occluded): 100% stenosis (total vessel occlusion).

In this study, we specifically focused on the AI’s ability to differentiate between non-obstructive (Categories 0–2) and obstructive (Categories 3–5) disease, which is the most critical decision point in clinical triage. We diagnosed patients as having significant CAD if they had at least one stenosis with severity greater than or equal to 50%. We compared our diagnosis with the gold standard diagnosis by expert cardiologists who reviewed CCTA images using dedicated software. We evaluated our method for CSE using correlation coefficient, accuracy, sensitivity, specificity.

In our study, we focused on defining and quantifying calcification scoring (CS), a key indicator of coronary artery calcification crucial for assessing the risk of CAD. Utilizing a cMPR technique to obtain detailed CT images of the heart, we applied a previously developed SSM for isolating vessel regions within these images. The CS was determined by enumerating voxels exhibiting a CT attenuation value exceeding 2,100 HU, a threshold meticulously chosen based on its discriminatory power for distinguishing calcified regions in our training dataset. This quantification of CS was presented as a percentage relative to the total number of voxels within the vessel region, thereby providing a comprehensive measure of calcification extent (see Figure 4).

To enhance the robustness of our CS evaluation, a dual assessment approach was incorporated. Firstly, CS was visually appraised by expert cardiologists, each with over 15 years of experience, who assigned a score ranging from 0 to 10 based on calcification extent and severity. This scoring system, while subjective, provided a valuable qualitative assessment due to the absence of a definitive ground truth for CS. Secondly, a binary threshold method was employed, utilizing specific HU determined from histogram analysis of calcified patient images. This approach allowed for a more objective and replicable measure of calcification. Together, these methods established a comprehensive framework for evaluating coronary artery calcium, contributing significantly to our understanding and diagnosis of CAD.

In CAD diagnostics, the identification of vulnerable plaques holds paramount importance due to their potential to precipitate acute cardiac events. To address this critical need, our methodology encompassed a rigorous evaluation of the presence or absence of vulnerable plaques in each CA. To establish a meticulous reference standard for vulnerable plaque, we adopted a multi-parametric approach based on established CCTA-derived high-risk plaque (HRP) features. A plaque was labeled as “vulnerable” if it exhibited at least two of the following four criteria: (1) low-attenuation plaque (LAP < 30 HU), (2) positive remodeling (remodeling index > 1.1), (3) napkin-ring sign, and (4) spotty calcification (Motoyama et al., 2007; Motoyama et al., 2015; Ferencik et al., 2018; Cury et al., 2022). The labeling procedure was conducted by an expert cardiologist with over 10 years of experience. The expert manually annotated the regions of interest (ROIs) on cMPR images while cross-referencing the findings with official Invasive Coronary Angiography (ICA) reports to ensure clinical relevance. Although ICA primarily assesses luminal narrowing, the presence of complex lesions or filling defects documented in ICA was used to provide ancillary support for the vulnerable plaque labels assigned on CCTA. This rigorous labeling process ensures that our model targets morphologically unstable plaques that are clinically significant.

This process was facilitated by the development and training of a VPSM designed to differentiate between normal CA and those harboring vulnerable plaques. The efficacy of this VPSM was quantitatively assessed using key diagnostic metrics, specifically focusing on its sensitivity and specificity. These metrics were meticulously calculated for each CA, providing a detailed evaluation of the model’s ability to accurately identify vulnerable plaques crucial for risk stratification and subsequent clinical decision-making.

To validate the robustness of our model, we conducted extensive testing on a dataset comprising 2,437 angiograms, each meticulously annotated with ground truth data for vessel and stenosis. The performance of our model was evaluated based on four critical metrics: accuracy, DSC, sensitivity, and specificity. These metrics collectively offer a comprehensive overview of the model’s performance, encompassing its precision, reliability in plaque detection, and ability to minimize false positives and negatives—key factors in clinical diagnostics (see Figure 5).

A rule-based report generator for cMPR images was developed and evaluated. The report generator could extract information from cMPR images and provide useful information to clinicians about stenosis, vulnerable plaque, and calcification (Figure 6). We used 100 cMPR images from a test dataset as the stimuli for this study. These images were anonymized with various degrees of stenosis, vulnerable plaque, and calcification. We also developed a rule-based readout generator that could extract information from the cMPR images using image processing techniques. It generated readouts based on predefined templates. The report generator provided information about the location, severity, and length of stenosis as well as vulnerable plaque and calcification in each coronary artery segment.

The dataset was split into development and test sets based on the year of acquisition (Time-independent split). Within the development set, we maintained a strict 4:1 patient-level ratio for training and validation to prevent data leakage between slices of the same patient (Table 2). Training was performed using standard cross-validation with a 4:1 training-to-validation ratio. Data augmentation included random rotations, scaling, elastic deformations, and intensity variations to enhance model generalization. The models achieved convergence within 1,000 epochs with early stopping based on validation loss plateau. A time-independent test dataset was utilized to ensure temporal generalizability of the predictive models.