Automatic diagnosis of coronary artery stenosis by deep learning based on X-ray coronary angiography

Introduction

Coronary artery disease (CAD) has emerged as the principal cause of global mortality (1). This disease emerges with the accumulation of atheromatous plaques in the coronary arteries, resulting in their narrowing or obstruction (2). Such narrowing reduces the flow of oxygen-rich blood to the heart muscle, elevating the risk of heart attacks and, in acute cases, resulting in death. X-ray coronary angiography is the current gold standard for diagnosing clinically significant CAD (3). This modality provides invaluable insights into early CAD markers by evaluating the morphological features of coronary arteries, including their diameter, bifurcation angles, and tortuosity (4). The application of coronary angiography has broadened and is now integral to several advanced cardiac interventions (5).

The primary objective in coronary angiography interpretation is discerning and quantifying stenosis severity within the coronary vessels (6). However, angiograms frequently pose challenges due to complex vessel morphologies, inconsistent contrast, and varied degrees of illumination (4). These complications can adversely affect the precise determination of coronary artery stenosis severity. Moreover, significant inconsistencies often arise in stenosis assessments when a variety of measurement techniques are applied to the same angiogram (7). Although traditional manual assessments remain crucial, they are time-consuming and subject to the variability of individual training and judgment (8).

Consequently, there is a growing preference for the automated detection and grading of coronary artery stenoses via X-ray angiograms by computer-aided diagnosis systems (9). Such automated methods reduce variability among observers, granting cardiologists a supplementary diagnostic perspective (10). This could potentially enhance diagnostic precision and operational efficacy. Numerous frameworks have been advanced for the automated quantitative assessment of stenosis in coronary angiography (11). Broadly, these can be categorized into two primary methodologies based on workflow characteristics: the modular pipeline approach and the end-to-end deep learning approach. The modular pipeline method involves a sequence of distinct steps, such as vessel detection (or enhancement), vessel segmentation, diameter calculation of the vessel, and subsequent stenosis analysis (12). Each module functions independently in this approach, cumulatively contributing to the final determination of stenosis.

The modular pipeline approach offers certain advantages, particularly in situations in which detailed control and interpretability are essential. However, the end-to-end deep learning method is often more streamlined, adaptable, and efficient, especially when there are large amounts of annotated data available for training. Recent advancements in end-to-end deep learning for stenosis assessment have emphasized two key objectives: (I) discerning the presence of stenosis in images; and (II) identifying its location (Table S1). Ovalle-Magallanes et al. introduced several models using convolutional neural networks (CNN) paired with attention modules to determine the presence of stenosis, achieving binary classification accuracies ranging from 87.87% to 95.49% (13-17). Similarly, Moon et al. and Stralen et al. deployed varying CNN architectures to identify stenosis (18,19). Visual interpretation tools, such as gradient-weighted class activation mapping, have been used to explain the relevant features within these models (20). Consequently, the stenosis location can be visually represented by highlighting the pixels. In addition to these aforementioned methods, Han et al. adopted deep learning techniques to delineate the stenosis by drawing bounding boxes around the affected areas (21). Such an approach offers clear advantages: it provides a more explicit spatial context of stenosis, aiding cardiologists in precisely identifying its location. Moreover, bounding boxes can facilitate streamlined integration into medical imaging software, allowing for efficient overlay and comparison against other diagnostic tools.

However, only a handful of methodologies have been proposed within the scope of end-to-end coronary stenosis quantification. Cong et al. proposed categorization methods that differentiate stenosis into three distinct grades: below 25%, between 25% and 99%, and total occlusion (22,23). Similarly, Danilov et al. advanced in this domain by grading stenosis into small, medium, or large classifications based on the bounding box’s area (24). However, these pioneering methods involve certain limitations, as they primarily group stenosis severity into broad categories, neglecting the clinically significant finer grades within the 71–100% range. Transitioning from simple binary detection to a more granular grading system holds considerable promise. Such a grading approach offers a richer insight into stenosis severity, which is essential for developing tailored treatment plans (25). Furthermore, it is necessary for tracking disease evolution and can comprehensively characterize a patient’s cardiovascular status, thus facilitating better clinical decision-making.

In this study, we aimed to develop a novel deep learning-based approach to segment and grade vessel stenosis based on a large, single-center retrospective X-ray coronary angiography cohort and to evaluate its performance. We present this article in accordance with the CLEAR reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-24-2415/rc).

Methods

Study population

Data for coronary angiography in this study were retrospectively collected from the Department of Cardiology at the Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine, between January 2020 and December 2022. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Ethical approval was obtained from the Ethics Committee of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (approval No. SH9H-2019-T160-5). Informed consent was obtained from all participants for this retrospective analysis, and all procedures conformed to the established guidelines and protocols. All cases were anonymized to maintain confidentiality. Exclusion criteria included patients with in-stent restenosis, significant eccentric lesions, coronary artery anomalies, including arteriovenous or coronaries-to-cardiac chamber fistulae, notable orifice anomalies, congenital absences, presence of bridging vessels, and moderate-to-severe myocardial bridges.

X-ray coronary angiography collection

Image data from coronary angiography procedures were collected with the Innova digital subtraction X-ray vascular imaging system (GE HealthCare, Chicago, IL, USA) and subsequently stored in Digital Imaging and Communications in Medicine (DICOM) format on workstations postexportation. These DICOM images were accessed and analyzed with the eFilm workstation (version 4.10; Merge Healthcare, Chicago, IL, USA), with a focus on the right coronary artery (left anterior oblique 30°) and left coronary artery (cranial 30° and caudal 30°) views. Lesion-containing DICOM images were exported as single-frame images at a rate of 25 frames per second. For all patient data, a leading cardiologist in cardiac catheterization and angiography handpicked the illustrative frame from the image sequences, which showed the detailed vascular structure. Most of the chosen frames from the X-ray angiography were the initial image captured upon the contrast agent reaching the distal end of the targeted vessel during injection.

Ground truth annotation

Stenosis annotations were facilitated with specialized software (Dataloop AI, Herzliya, Israel). Different degrees of stenosis were delineated in five distinct colors (Figure 1), corresponding to ranges of 1–49, 50–70, 71–95, 96–99, and 100. Each annotated image underwent a verification process by two cardiologists specialized in coronary artery intervention. In particular, well-trained experts completed a thorough examination and labelling to ensure the accuracy of these reference annotations. Stenosis points shown in the images were recognized only if both readers agreed on their assessments, and the color-coded scheme described above was used to represent various degrees of stenosis. Only those candidate images that received consistent stenosis gradings from both cardiologists were included in this analysis. In cases in which the two cardiologists disagreed, a senior interventional cardiologist was consulted to provide an authoritative judgment.

Figure 1 X-ray images showing manually annotated stenosis. Each set consists of the original X-ray image (top) and the corresponding annotated X-ray (bottom) highlighting different severity levels. The annotations include five distinct colors, each representing a specific degree of stenosis. Particular emphasis is placed on identifying and annotating the most severely stenotic segment within a continuous lesion, as this severity typically carries greater clinical weight in determining the optimal treatment strategy.

Model architecture

The overall workflow of the proposed framework is illustrated in Figure 2. To address the challenge of identifying stenotic regions in vascular structures, we developed a segmentation model that integrates deep learning with guidance in vessel structure. In this model, after preprocessing and augmentation (described in section “Model training and performance evaluation index”), each X-ray angiography image is processed through a shared encoder-decoder backbone. A Frangi filter is applied to enhance the tubular features of coronary arteries, producing a vessel map that serves as an auxiliary learning target. The network then generates two outputs from the shared feature representation: (I) the primary stenosis segmentation, color-coded by clinically defined severity grades; and (II) the auxiliary vessel map, which reproduces the precomputed Frangi-based vessel structure. This dual-output design, extending the conventional U-Net framework, incorporates an additional output head for vessel reconstruction alongside the stenosis segmentation. The auxiliary supervision guides the model to preserve vessel topology, acting as an anatomical constraint during training and enabling anatomically consistent localization of stenoses while improving segmentation accuracy and clinical applicability.

Figure 2 Workflow of the proposed stenosis segmentation framework. The input X-ray angiography image is first subjected to data augmentation and then processed through an encoder-decoder backbone to extract shared feature maps. The model produces two outputs: a primary output for stenosis segmentation, where colored segments denote different levels of stenosis severity, and an auxiliary output representing vessel-like structures. The auxiliary output is supervised with a precomputed vessel map derived via Frangi filtering and thresholding, providing anatomical guidance to enhance segmentation of tubular vascular structures.

The encoder–decoder structure is illustrated in Figure 3. The encoder comprises a series of contraction blocks, each including a max pooling layer followed by two 2D convolutional layers with activation functions and regularization through batch normalization. Channel depth increases progressively at each stage: from an initial 32 channels to 64, 128, 256, and finally 480. Convolutional layers in the first block use a stride of [1, 1], which is modified to [2, 2] in subsequent blocks to downsample the input. The decoder mirrors this structure through a series of expanding blocks, each containing a transpose convolution followed by two convolutional layers. These blocks receive dual inputs: one from the corresponding encoder block and another from the previous decoder layer. Channels are gradually reduced from 960 to 32 in powers of two (960, 480, 256, 128, 64, and 32), with all convolutional layers using a kernel size of [3, 3], a stride of [1, 1], and a padding of [1, 1]. Instance normalization and leaky rectified linear unit activations follow each layer to improve convergence and model stability.

Figure 3 Topology of the convolutional neural networks employed in this work for automatic feature extraction.

We developed our model in Python 3.10.9 using PyTorch 1.13.1 with CUDA 11.6 and NumPy 1.23.5. Training and evaluation were conducted on an RTX 3090 Ti GPU (Nvidia, Santa Clara, CA, USA).

Model training and performance evaluation index

Model weights were initialized via He initialization in PyTorch (26). We trained the model to identify narrowing regions for 600 epochs, implementing an early stopping mechanism if there was no improvement in the loss function for 10 consecutive epochs. If this criterion was met, the training halted, reverting the model to its optimal state. When the loss function last showed improvement. Across the 5-fold cross-validation, the best validation performance generally emerged between epochs 450 and 510, after which early stopping was typically triggered. The choice of 600 as the maximum epoch count provided sufficient headroom to accommodate convergence across different folds and augmentation variations without prematurely cutting off training. Before inputting the training set images into the model for every epoch, they were cropped to maintain the central field of view (512×512 pixels) and augmented through random scaling, rotation, elastic deformation, gamma correction, and mirroring.

To mitigate the imbalance in the levels of stenosis severity, particularly for 96–99% and 100% cases, we employed patch-based foreground oversampling during training. Patches were preferentially selected from regions containing stenotic lesions, increasing the model’s exposure to clinically significant but underrepresented cases. The dataset was split into a training set and a test set via a stratified fivefold cross-validation; that is, 80% of the images were used to train the model and the remaining 20% to monitor their generalization capability as a test set. The total loss was computed as the sum of the stenosis segmentation loss (Dice loss) and the vessel structure loss (cross-entropy loss). Six performance metrics were evaluated during the training: intersection over union [IoU; ], Dice score [], precision [], sensitivity [], where TP = true positives, FP = false positives, FN = false negatives. Dice loss [], and cross-entropy loss [] (27).

Results

Patient demographics

A total of 398 coronary angiographic images from 179 patients were initially collected (Figure 4). The degree of coronary artery stenosis in each image was independently assessed by two cardiologists according to clinical standards. When interpretations were inconsistent, a senior cardiologist adjudicated the disputed cases. If consensus among all three cardiologists could not be reached, the image was excluded from the analysis. Following this adjudication process and the application of exclusion criteria, 15 images from 11 patients were excluded. Ultimately, 383 images from 168 patients were included in the final dataset. The mean age of the included patients was 68.640±0.705 years, with a median age of 69 years (range, 36–83 years). The gender distribution was 125 males and 43 females. Out of 383 angiography images (Table 1), we identified vessel segments with 288 nonobstructive lesions (1–49%), 223 intermediate lesions (50–70%), 236 severe lesions (71–95%), 134 sub-total occlusions (96–99%), and 63 total occlusions (100%). All patients presented two or more instances of stenosis.

Figure 4 Flowchart illustrating the inclusion and exclusion process for images from coronary angiography. Out of 398 images assessed, 15 were excluded due to unresolved interpretation discrepancies, resulting in 383 images from 168 patients included in the final analysis.

Qualitative evaluation of the stenosis segmentation model

The primary approach we adopted was qualitative analysis. We used our model to segment images, highlighting narrowed arteries, especially for degrees of stenosis categorized as 1–49%, 50–70%, and 71–95%. The segmentation results are provided in Figure 5. From this figure, it is evident that the segmented area of arteries, particularly in the ranges of 1–49%, 50–70%, and 71–95%, is identified with significant accuracy by our model. However, stenosis degrees of 96–99% and 100% produced a higher incidence of false negatives. To further illustrate the limitations of the proposed model, Figure 6A presents representative false-negative cases. These examples highlight that most missed stenoses occurred in segments with low angiographic contrast or overlapping vessel structures. Such conditions reduced the visual distinction between normal and narrowed segments, making accurate localization challenging for both automated methods. In the first example, the segment was only partially detected due to its short length and distal location within an overlapping vessel path. In the second example, branch overlap obscured the 96–99% component of the lesion, resulting in underestimation of the total stenotic burden. The confusion matrix in Figure 6B indicates that false negatives were relatively common, whereas false positives were comparatively rare across all categories.

Figure 5 Prediction of stenosis. Five distinct colors were used to represent the specific degree of stenosis.

Figure 6 False-negative examples and confusion matrix of the stenosis segmentation model. (A) Representative false negative cases from the proposed model. In the first case (top row), the false negative arises from reduced contrast in the segment and vessel overlap, leading to incomplete detection of a short segment. In the second case (bottom row), the missed segment occurs in a region with overlapping branches, causing reduced visibility and difficulty in delineating the full extent of the stenosis. (B) Confusion matrix for the stenosis segmentation model. Each cell represents the number of pixels. Rows denote the ground truth categories, and columns indicate the predicted categories.

Furthermore, to underscore the significance of the Frangi filter in our methodology, we conducted an ablation study by excluding it from our model. The Frangi filter, in essence, acts as a constraint on the feature map, enhancing the vascular structures and emphasizing the variations in the arterial diameter. Our ablation study showed that without this specialized filter, the model had reduced capacity to effectively detect all arterial narrowings.

Quantitative evaluation of the stenosis segmentation model

In our study, we used fivefold cross-validation, data augmentation, and early stopping to prevent overfitting. This approach achieved strong results, with an IoU of 0.92, a Dice score of 0.95, a precision of 0.93, and a sensitivity of 0.96. Table 2 shows the segmentation outcomes for different stenosis degrees, with the best performance for stenosis categories of 71–95% and 96–99%. The IoU for 71–95% was 93.73%, slightly higher than that for 96–99% at 93.03%. The 50–70% category had the lowest IoU at 88.72%, but all metrics remained above 88%. Precision peaked at 95.16%, and sensitivity was highest for 96–99% stenosis at 97.97%, indicating strong overall performance.

The confusion matrix in Figure 6B illustrates the algorithm’s effectiveness in pixel-wise segmentation across stenosis grades. It indicates some misclassification occurred in the 0 category, suggesting challenges in distinguishing nonstenosis areas from regions with minimal or significant stenosis. However, there was remarkable precision and minimal misclassifications for stenosis grades of 71–95%, 96–99%, and 100%, highlighting the model’s strength. In contrast, nonobstructive to intermediate lesions (1–49% and 50–70%) produced more confusion, particularly with the 0 category.

We compared our model with other top-performing models such as U-Net, U-Net++, residual neural network (ResNet), MobileNetV2, and Vision Transformer. U-Net is widely used in biomedical image segmentation, and its advanced version, U-Net++, enhances performance by capturing finer details (28). ResNet introduces skip connections for deeper networks (29), MobileNetV2 focuses on efficiency for mobile devices (30), and the Vision Transformer applies the Transformer architecture to visual data (31). Detailed configurations of these models are provided in Appendix 1.

Tables 3,4 show a significant difference in performance between models for different degrees of stenosis. Our model consistently outperformed others, particularly for stenosis of 1–49%, with IoU decreasing steeply from 92.87% to 19.22% without Frangi Filtering. Across all stenosis degrees, our model had the highest IoU, with the closest competitor being our model without Frangi Filtering in the 96–99% range at 46.20%. Other models, including U-Net and U-Net++, performed reasonably well but still fell short of our model.

Similarly, our model with Frangi filtering had the highest Dice coefficient across all stenosis levels. In the 96–99% category, although U-Net++ and the version without Frangi filtering performed relatively better as compared to other categories, our model remained superior, with a Dice coefficient of 95.45%.

Discussion

In this study, we demonstrated the feasibility of using a dual-output deep learning model to accurately diagnose coronary artery stenosis from X-ray angiography. The proposed approach achieved automatic segmentation of stenotic regions and incorporated a vessel-like auxiliary output, resulting in a high Dice coefficient (>0.90). These findings indicate that deep learning-based analysis of X-ray angiography can provide accurate, efficient, and reproducible stenosis assessment, offering a viable alternative to time-consuming and operator-dependent expert analysis.

Accurate segmentation of coronary artery stenosis is essential for the reliable postprocessing of angiographic data. Manual annotating, however, is labor-intensive and subject to interobserver variability. Although previous studies (Appendix 1, Table S1) have shown that deep learning can identify stenosis in images (13-19), many have been limited to classification or localization tasks, such as using bounding boxes to mark suspected regions (21,32). In contrast, our study used 383 images, exceeding the sample size of many previous works, to train a segmentation model capable of segmenting stenosis at the pixel level. By integrating an auxiliary task derived from Frangi filter-based vessel enhancement into a standard U-Net architecture, we achieved high segmentation accuracy compared with previous studies, as reflected in the Dice coefficient exceeding 0.90. Although U-Net’s proven performance in coronary artery tree segmentation [34] suggested potential for stenosis segmentation, our ablation studies suggested that applying U-Net alone to this task is challenging. This is largely due to the need to simultaneously identify vessel boundaries and determine stenosis severity grades. Our dual-output design addressed this challenge by incorporating vessel structure information into the segmentation process. The auxiliary branch improved performance by providing a detailed vessel map that guided the detection and grading of stenoses.

Despite the strong overall performance, errors were most frequently observed in intermediate stenosis cases (50–70%) and in severe lesions exhibiting complex morphology. Intermediate lesions often exhibit heterogeneous features, including eccentric narrowing, calcification, vascular remodeling, or bifurcation involvement, which can visually overlap with milder or more severe categories, leading to boundary misclassification. In some high-grade lesions, particularly those with diffuse narrowing or overlapping vessel segments, the model underestimated stenosis severity due to reduced contrast or vessel foreshortening in angiographic projections. To address these challenges, future work will expand the training dataset to include a greater number of such problem cases and integrate intravascular imaging as a reference standard to improve accuracy in complex 2D angiographic interpretations.

Some prior studies have explored automated evaluation of coronary artery stenosis using machine learning or deep learning. For example, Cong et al. employed a hybrid long short-term memory (LSTM) and InceptionV3 network to classify stenosis severity into three categories (<25%, 25–99%, or total occlusion) (22), while Danilov et al. automatically estimated stenosis size (small, medium, or large) using deep learning (24). However, clinically significant stenosis is typically defined as >70% narrowing in a major coronary branch, and few studies have specifically addressed this critical range, underscoring the need for improved patient care in such cases (33).

More recent advances in stenosis segmentation and assessment include attention-based modules. However, their performance tends to decline in cases of severe stenosis (34). Furthermore, an integrated detection–segmentation pipeline combining YOLOv8 for region detection with DeepLabV3+ for stenosis delineation and thickness measurement has been reported, yet this work focused solely on detection and did not report performance for severity grading (35). Compared with these approaches, our method provides a finer-grained and clinically relevant severity stratification across five categories—nonobstructive (1–49%), intermediate (50–70%), severe (71–95%), sub-total occlusion (96–99%), and total occlusion (100%), while maintaining high segmentation precision, particularly within the critical 71–100% range.

This study involved several limitations that should be addressed. First, due to its retrospective design, we included only patients who had previously undergone X-ray angiography, preventing prospective comparison of the deep learning method with traditional image analysis. Nevertheless, by training the stenosis segmentation model on an isolated cohort and comparing the results with traditional reports, we obtained conclusive and clinically consistent findings. Patients with eccentric lesions and in-stent restenosis were excluded due to their significant impact on image interpretation. Given that computed tomography coronary angiography (CTCA) provides 3D visualization capable of resolving interpretative difficulties in eccentric lesions, incorporating CTCA alongside X-ray angiography may enable broader inclusion of such cases in future studies (36). Future research should aim to develop automated approaches that incorporate additional anatomical, morphological, and functional information on coronary lesions.

Second, our evaluation relied on stratified fivefold cross-validation without a separately held-out test set, and all data were drawn from a single institution. Although this strategy ensured representation across stenosis severity levels, particularly for underrepresented high-grade cases, the absence of external validation may limit generalizability to other imaging systems, acquisition protocols, and patient populations. Incorporating an independent test cohort and evaluating the model on external datasets or public benchmarks would provide a more rigorous assessment of performance. To address this, we are actively recruiting more patients with high-grade stenosis (96–99% and 100%) and plan to validate the model across a diverse clinical setting, similar to the dataset-diversification strategies successfully applied in other medical imaging domains, such as force map-enhanced segmentation frameworks for early cancer detection (37).

Third, our study did not include advanced agreement analyses, such as Dice-area under the curve curves or Bland-Altman plots, nor did it implement visual interpretability mechanisms such as class activation maps, attention maps, or saliency maps. As our proposed framework produces segmentation outputs via an argmax operation, these threshold-dependent analyses were not performed. Nonetheless, such methods could provide deeper insight into model performance relative to expert annotations, assess stability across probability thresholds, and improve transparency by confirming that predictions are based on anatomically relevant features. Incorporating these analyses and interpretability tools in future work could enhance clinical confidence and support the integration of our framework into real-world workflows.

Conclusions

Automatically detecting and grading vessel stenosis poses challenges, especially with complex vascular structures and backgrounds of varied intensities. This work introduced an innovative, fully automated technique that uses deep learning for precise quantification of stenosis severity in X-ray coronary angiography.

The deep learning component streamlined the stenosis detection process, ensuring reliable and precise tracking. Meanwhile, the vessel-like auxiliary output enhanced the model’s focus, as stenosis within coronary arteries represents only a minute portion of X-ray images. Our evaluations, both qualitative and quantitative, highlight the method’s efficacy. Notably, it surpassed other popular models with an average IoU of 0.92, a Dice score of 0.95, a precision of 0.93, and a sensitivity of 0.96. The method precisely localized and classified stenoses across different grades, achieving notable IoUs for each grade.

These findings suggest the potential of our approach in fostering a computer-aided diagnostic system to automate vessel stenosis quantification, enhancing clinical assessments. Nonetheless, there is room for improvement, particularly in differentiating nonstenosis from milder stenosis grades. Given the model’s proficiency in identifying stenosis of greater severity, refining its performance in boundary areas prone to misclassifications should be prioritized. Future studies could also explore the paths for other cohorts, such as patients with eccentric lesions or in-stent restenosis.

Acknowledgments

We extend our gratitude to Villanelle Life (Shanghai) Co. Ltd. for their valuable support in data processing and consultation.

Footnote

Reporting Checklist: The authors have completed the CLEAR reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-24-2415/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-24-2415/dss

Funding: This work was supported by grants from the Clinical Research Program of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. JYLJ202014 to J.Z.), Clinical Research Project of Multi-Disciplinary Team, Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. 201911 to J.Z.), Biobank for Coronary Heart Disease of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. YBKA202206 to J.Z.) and Shanghai Hospital Development Center Three-Year Action Plan for Promoting Clinical Skills and Innovation Ability of Municipal Hospitals (No. SHDC2022CRD045 to J.Z.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-2415/coif). J.Z. reports this work was supported by grants from the Clinical Research Program of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. JYLJ202014 to J.Z.), Clinical Research Project of Multi-Disciplinary Team, Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. 201911 to J.Z.), Biobank for Coronary Heart Disease of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. YBKA202206 to J.Z.) and Shanghai Hospital Development Center Three-Year Action Plan for Promoting Clinical Skills and Innovation Ability of Municipal Hospitals (No. SHDC2022CRD045 to J.Z.). The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (approval No. SH9H-2019-T160-5), and informed consent was obtained from all participants for this retrospective analysis.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ralapanawa U, Sivakanesan R. Epidemiology and the Magnitude of Coronary Artery Disease and Acute Coronary Syndrome: A Narrative Review. J Epidemiol Glob Health 2021;11:169-77. [Crossref] [PubMed]
2. Collet JP, Zeitouni M, Procopi N, Hulot JS, Silvain J, Kerneis M, Thomas D, Lattuca B, Barthelemy O, Lavie-Badie Y, Esteve JB, Payot L, Brugier D, Lopes I, Diallo A, Vicaut E, Montalescot GACTION Study Group. Long-Term Evolution of Premature Coronary Artery Disease. J Am Coll Cardiol 2019;74:1868-78. [Crossref] [PubMed]
3. Athanasiou LS, Fotiadis DI, Michalis LK. Atherosclerotic plaque characterization methods based on coronary imaging. Amsterdam, The Netherlands: Academic Press; 2017.
4. Çimen S, Gooya A, Grass M, Frangi AF. Reconstruction of coronary arteries from X-ray angiography: A review. Med Image Anal 2016;32:46-68. [Crossref] [PubMed]
5. Rear R, Bell RM, Hausenloy DJ. Contrast-induced nephropathy following angiography and cardiac interventions. Heart 2016;102:638-48. [Crossref] [PubMed]
6. Nallamothu BK, Spertus JA, Lansky AJ, Cohen DJ, Jones PG, Kureshi F, et al. Comparison of clinical interpretation with visual assessment and quantitative coronary angiography in patients undergoing percutaneous coronary intervention in contemporary practice: the Assessing Angiography (A2) project. Circulation 2013;127:1793-800. [Crossref] [PubMed]
7. Gould KL, Kirkeeide RL. Assessment of stenosis severity. In: Reiber JHC, Serruys PW, editors. State of the Art in Quantitative Coronary Arteriography. Dordrecht: Springer Netherlands; 1986:209-28.
8. Zhang H, Mu L, Hu S, Nallamothu BK, Lansky AJ, Xu B, et al. Comparison of Physician Visual Assessment With Quantitative Coronary Angiography in Assessment of Stenosis Severity in China. JAMA Intern Med 2018;178:239-47. [Crossref] [PubMed]
9. Alizadehsani R, Abdar M, Roshanzamir M, Khosravi A, Kebria PM, Khozeimeh F, Nahavandi S, Sarrafzadegan N, Acharya UR. Machine learning-based coronary artery disease diagnosis: A comprehensive review. Comput Biol Med 2019;111:103346. [Crossref] [PubMed]
10. Sanchez-Martinez S, Camara O, Piella G, Cikes M, González-Ballester MÁ, Miron M, Vellido A, Gómez E, Fraser AG, Bijnens B. Machine Learning for Clinical Decision-Making: Challenges and Opportunities in Cardiovascular Imaging. Front Cardiovasc Med 2021;8:765693. [Crossref] [PubMed]
11. Zhou Y, Guo H, Song J, Chen Y, Wang J. Review of vessel segmentation and stenosis classification in X-ray coronary angiography. 2021 13th International Conference on Wireless Communications and Signal Processing (WCSP), Changsha, China, 2021. doi: 10.1109/WCSP52459.2021.9613197.
12. Wan T, Feng H, Tong C, Li D, Qin Z. Automated identification and grading of coronary artery stenoses with X-ray angiography. Comput Methods Programs Biomed 2018;167:13-22. [Crossref] [PubMed]
13. Ovalle-Magallanes E, Alvarado-Carrillo DE, Avina-Cervantes JG, Cruz-Aceves I, Ruiz-Pinales J, Contreras-Hernandez JL. Attention Mechanisms Evaluated on Stenosis Detection using X-ray Angiography Images. Journal of Advances in Applied & Computational Mathematics 2022;9:62-75.
14. Ovalle-Magallanes E, Avina-Cervantes JG, Cruz-Aceves I, Ruiz-Pinales J. Hybrid classical–quantum Convolutional Neural Network for stenosis detection in X-ray coronary angiography. Expert Systems with Applications 2022;189:116112.
15. Ovalle-Magallanes E, Avina-Cervantes JG, Cruz-Aceves I, Ruiz-Pinales J. Transfer learning for stenosis detection in X-ray coronary angiography. Mathematics 2020;8:1510.
16. Ovalle-Magallanes E, Avina-Cervantes JG, Cruz-Aceves I, Ruiz-Pinales J. LRSE-Net: Lightweight Residual Squeeze-and-Excitation Network for Stenosis Detection in X-ray Coronary Angiography. Electronics 2022;11:3570.
17. Ovalle-Magallanes E, Avina-Cervantes JG, Cruz-Aceves I, Ruiz-Pinales J. Improving convolutional neural network learning based on a hierarchical bezier generative model for stenosis detection in X-ray images. Comput Methods Programs Biomed 2022;219:106767. [Crossref] [PubMed]
18. Moon JH, Lee DY, Cha WC, Chung MJ, Lee KS, Cho BH, Choi JH. Automatic stenosis recognition from coronary angiography using convolutional neural networks. Comput Methods Programs Biomed 2021;198:105819. [Crossref] [PubMed]
19. Stralen PV, Rodrigues DL, Oliveira AL, Menezes MN, Pinto FJ. Stenosis detection in X-ray coronary angiography with deep neural networks leveraged by attention mechanisms. Proceedings of the 9th International Conference on Bioinformatics Research and Applications; 2022.
20. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-cam: Visual explanations from deep networks via gradient-based localization. 2017 IEEE International Conference on Computer Vision (ICCV), Venice, Italy, 2017. doi: 10.1109/ICCV.2017.74.
21. Han T, Ai D, Li X, Fan J, Song H, Wang Y, Yang J. Coronary artery stenosis detection via proposal-shifted spatial-temporal transformer in X-ray angiography. Comput Biol Med 2023;153:106546. [Crossref] [PubMed]
22. Cong C, Kato Y, Vasconcellos HD, Lima J, Venkatesh B. Automated stenosis detection and classification in x-ray angiography using deep neural network. 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), San Diego, CA, USA, 2019. doi: 10.1109/BIBM47256.2019.8983033.
23. Cong C, Kato Y, Vasconcellos HD, Ostovaneh MR, Lima JAC, Ambale-Venkatesh B. Deep learning-based end-to-end automated stenosis classification and localization on catheter coronary angiography. Front Cardiovasc Med 2023;10:944135. [Crossref] [PubMed]
24. Danilov V, Gerget O, Klyshnikov K, Ovcharenko E, Frangi A. Comparative study of deep learning models for automatic coronary stenosis detection in x-ray angiography. Proceedings of the 30th International Conference on Computer Graphics and Machine Vision; 2020: CEUR Workshop Proceedings.
25. Ovalle-Magallanes E, Alvarado-Carrillo DE, Avina-Cervantes JG, Cruz-Aceves I, Ruiz-Pinales J, Correa R. Deep Learning-based Coronary Stenosis Detection in X-ray Angiography Images: Overview and Future Trends. Artificial Intelligence and Machine Learning for Healthcare: Vol 2: Emerging Methodologies and Trends 2022:197-223.
26. He K, Zhang X, Ren S, Sun J, editors. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. 2015 IEEE International Conference on Computer Vision (ICCV), Santiago, Chile, 2015. doi: 10.1109/ICCV.2015.123.
27. Taha AA, Hanbury A. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC Med Imaging 2015;15:29. [Crossref] [PubMed]
28. Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, Cham, 234-241. doi: 10.1007/978-3-319-24574-4_28.
29. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016. doi: 10.1109/CVPR.2016.90.
30. Sandler M, Howard A, Zhu M, Zhmoginov A, Chen LC. Mobilenetv2: Inverted residuals and linear bottlenecks. Proceedings of the IEEE conference on computer vision and pattern recognition; 2018. doi: 10.1109/CVPR.2018.00474.
31. Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T, Dehghani M, Minderer M, Heigold G, Gelly S, Uszkoreit J, Houlsby N, editors. An image is worth 16×16 words: Transformers for image recognition at scale. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. International Conference on Learning Representations; 2021 May 3-7; Online. OpenReview; 2021.
32. Pang K, Ai D, Fang H, Fan J, Song H, Yang J. Stenosis-DetNet: Sequence consistency-based stenosis detection for X-ray coronary angiography. Comput Med Imaging Graph 2021;89:101900. [Crossref] [PubMed]
33. Neglia D, Rovai D, Caselli C, Pietila M, Teresinska A, Aguadé-Bruix S, et al. Detection of significant coronary artery disease by noninvasive anatomical and functional imaging. Circ Cardiovasc Imaging 2015;8:e002179. [Crossref] [PubMed]
34. Stralen PV, Rodrigues DL, Oliveira AL, Menezes MN, Pinto FJ. Stenosis detection in X-ray coronary angiography with deep neural networks leveraged by attention mechanisms. Proceedings of the 9th International Conference on Bioinformatics Research and Applications (ICBRA); 2022 Sep 18-20; Berlin, Germany. New York: ACM; 2022. p. 123-8.
35. Evtyukhov D, Kopanitsa G, Metsker O, Mogilevskii A, Yakovlev A, Kovalchuk S. Automatic Detection and Segmentation of Coronary Artery Stenosis in Coronary Angiography Images. International Conference on Computational Science; 2025: Springer.
36. Meng Q, Yu P, Yin S, Li X, Chang Y, Xu W, Wu C, Xu N, Zhang H, Wang Y, Shen H, Zhang R, Zhang Q. Coronary computed tomography angiography analysis using artificial intelligence for stenosis quantification and stent segmentation: a multicenter study. Quant Imaging Med Surg 2023;13:6876-86. [Crossref] [PubMed]
37. Umirzakova S, Muksimova S, Baltayev J, Cho YI. Force Map-Enhanced Segmentation of a Lightweight Model for the Early Detection of Cervical Cancer. Diagnostics (Basel) 2025;15:513. [Crossref] [PubMed]