Automated quantification of pericoronary fat attenuation index using deep learning: a validation study against expert manual measurement

Introduction

Background

Stable coronary artery disease (CAD) primarily stems from atherosclerosis and stands as a prominent cause of morbidity and mortality worldwide, posing significant challenges to the healthcare system (1-3). While lifestyle modifications and pharmacological interventions are effective in preventing CAD (4-6), identifying individuals who would derive the greatest benefit from these interventions remains a persistent challenge (7). Therefore, continuous efforts are required to enhance the prediction and stratification of cardiovascular risk at an individual level.

Rationale and knowledge gap

Recently, pericoronary fat attenuation index (FAI) (8) on coronary computed tomography angiography (CCTA) imaging has been proposed as a novel marker of coronary vascular inflammation with prognostic value for major cardiovascular events (9-11). However, the conventional FAI measurement method comes with limitations, such as the need for radiological expertise and manual delineation of coronary arteries. This manual process is time-consuming and prone to inter-reader variability, rendering it less feasible for routine clinical practice. While AI has shown promise in medical imaging (12-14), robust validation of automated FAI measurement, especially in the presence of artifacts like calcification, remains limited.

Objective

In this study, a commercial deep learning-based FAI (DL-FAI) measurement algorithm was proposed as a time-saving and bias-mitigating method that automatically delineates coronary arteries and measures FAI based on CCTA images. The primary objective of this study was to investigate the correlation and agreement between DL-FAI and manual FAI measurements in patients with CAD. We present this article in accordance with the STARD reporting checklist (available at https://rx.amegroups.com/article/view/10.21037/rx-26-0027/rc).

Methods

Data and study design

This retrospective study included 8,025 consecutive individuals who underwent CCTA at Shanghai General Hospital between January 2010 and December 2018. This cohort was used for development and optimization of the automated analysis workflow, including cardiac chamber segmentation, coronary artery segmentation, continuity refinement, centerline extraction, and branch labeling. The dataset was randomly divided at the patient level into a training set (n=6,063) and an internal validation set (n=1,962).

For external validation, 209 patients who underwent CCTA at Zhongshan Hospital, Fudan University, between January 2022 and December 2022 and also had invasive coronary angiography data were retrospectively enrolled. This cohort was used to evaluate the agreement, stability, and processing efficiency of the automated DL-FAI workflow against manual reference measurement in real-world cases with relatively complex disease phenotypes.

Patients with severely impaired CCTA image quality that did not meet basic diagnostic requirements or with tube voltage other than 100 or 120 kV were excluded. Image quality was pre-evaluated by experienced cardiovascular radiologists. Because of the retrospective design and image-quality-based screening, potential selection bias should be acknowledged.

Protocol of CCTA

All imaging data were anonymized before model development and validation to protect patient privacy. CCTA examinations were performed using multiple computed tomography (CT) platforms, including a 320-detector-row CT scanner (Aquilion One Vision, Canon Medical Systems), a dual-source CT scanner (SOMATOM Force, Siemens Healthineers, Germany), a 256-slice wide-detector CT scanner (Brilliance iCT, Philips Healthcare, the Netherlands), a 320-slice wide-detector CT scanner (uCT 960+, United Imaging, China), and a 128-slice multidetector CT scanner (uCT 780, United Imaging, China), according to routine cardiovascular CT protocols.

Sublingual nitroglycerin was routinely administered before scanning, and metoprolol was given when appropriate to achieve a relatively stable heart rate. Major acquisition parameters included tube voltage of 100 or 120 kV, detector collimation of 320 mm × 0.5 mm, tube current of 500 mA, and gantry rotation time of 275 ms. Iodinated contrast medium (iopamidol, 370 mg iodine/mL) was injected intravenously at 4.2 mL/s for a total volume of 70 mL. All image data were stored in DICOM format.

Expert reader manual measurements of pericoronary FAI

Manual FAI measurement served as the reference method for methodological evaluation of DL-FAI. Measurements were independently performed by two radiologists with substantial experience in cardiovascular imaging, both blinded to each other’s results and to the automated FAI results. The average of the two measurements was used as the manual reference value. A semiautomated software package (syngo.via, Siemens Healthineers) was used for manual measurement. Interobserver agreement was assessed in a randomly selected subset of cases, yielding an intraclass correlation coefficient (ICC) of 0.871.

It should be noted that manual measurement was used as a reference method to assess agreement with current clinical practice, rather than as an absolute biological ground truth.

The FAI algorithm

The FAI algorithm employs advanced DL techniques to determine the coronary perivascular adipose tissue (PVAT) attenuation pertaining to three primary coronary branches: the left anterior descending artery (LAD), the left circumflex artery (LCX), and the right coronary artery (RCA). This algorithm is aligned with the uAI coronary CTA reconstruction framework (United Imaging Intelligence, Shanghai, China) and comprises four integrated modules that operate cohesively.

The initial module employs a cascaded V-net-shaped fully convolutional neural network, following a coarse-to-fine approach, to delineate the heart chambers. Such segmentation is crucial to the accurate localization of coronary arteries. The subsequent module emphasizes fine-grained segmentation of the coronary arteries, combining both global and local anatomical contextual information and a decoding feature-guided encoding channel attention strategy (referred to as DEA, as shown in Figure 1). Moreover, a vascular tracing network is incorporated to mend discontinuities, enhancing the comprehensiveness of coronary segmentation. In the third module, we utilize a novel spatial anatomical dependency-based point-cloud network (called SAD-PointNet) that integrates spatial anatomical features and pre-trained segmentation network features. This specialized model is employed to precisely interpret and assign labels to the trajectory of coronary artery vessels. This stage plays a critical role in improving segmentation precision by applying coronary topological restrictions and eliminating erroneous segmentations like neighboring enhanced veins. Ultimately, perivascular regions to derive the PVAT attenuation are automatically isolated from the coronary segmentation, thereof various quantitative measurements, including FAI, are computed.

Figure 1 Schematic illustration of the DEA module used for fine-grained coronary artery segmentation. DEA, Decoding feature-guided Encoding channel Attention; FAI, fat attenuation index.

DL networks

The proposed framework consisted of three interconnected DL modules: cardiac chamber segmentation, coronary artery segmentation, and coronary artery trajectory labeling (Figure 2).

Figure 2 Overview of the deep learning framework. The architecture consists of three modules: (I) Cardiac chamber segmentation using a cascaded V-Net; (II) Coronary artery segmentation utilizing V-Net with DEA and TRACT-NET for discontinuity repair; and (III) Vessel trajectory labeling using SAD-PointNet. DEA, Decoding feature-guided Encoding channel Attention; SAD-PointNet, Self-Attention Distillation PointNet; TRACT-NET, TRACT Network; V-Net, Volumetric Convolutional Network.

The first module used a cascaded coarse-to-fine V-Net architecture for cardiac chamber segmentation (15). In the coarse stage, input CCTA images were resampled to 2.0 mm³ × 2.0 mm³ × 2.0 mm³ to rapidly localize the heart. The localized heart region was then cropped and resampled to 0.8 mm³ × 0.8 mm³ × 0.8 mm³ for the fine segmentation stage, generating refined chamber masks for subsequent coronary artery analysis.

The second module performed coronary artery segmentation by integrating anatomical constraints and local vessel continuity refinement. Based on the spatial relationship between the cardiac chambers and coronary arteries, a search band covering the expected coronary distribution was generated, and segmentation was restricted to this region to improve computational efficiency. Distance maps from the outer walls of the four cardiac chambers were used as additional input channels to provide anatomical context. A V-Net with a decoding feature-guided encoding channel attention (DEA) module was used for initial coronary artery segmentation. The DEA module computed channel attention weights from encoding- and decoding-stage feature maps, which were then used to enhance informative encoding features before concatenation with decoding features. To reduce discontinuities in segmented vessels, TRACT-NET was applied around the endpoints of the initial coronary segmentation. This patch-based network focused on local image details and helped restore vessel continuity in small, stenotic, or anatomically variable coronary segments.

The third module used SAD-PointNet for coronary artery trajectory labeling (16,17). SAD-PointNet was based on PointNet++ and assigned anatomical labels to coronary artery centerline points by combining image features from the preceding segmentation network with shape and spatial position features. The shape feature was defined as the local vessel axial direction vector. The position features were defined as three-dimensional coordinate vectors from each coronary artery point to the centroids of the left atrium, left ventricle, right atrium, and right ventricle. To improve robustness to different patient positions and imaging orientations, these features were calculated in a normalized cardiac coordinate system (Figure 3A,3B). In this coordinate system, the z-axis was defined as the intersection line between the atrioventricular groove plane and the interventricular groove plane, the y-axis was defined as parallel to the atrioventricular groove plane and orthogonal to the z-axis, and the x-axis was defined as the cross-product of the y- and z-axes. The origin was defined as the point on the z-axis closest to the centroid of the heart. High-level coronary image features extracted from the penultimate convolutional layer of the coronary segmentation network were integrated with the point-cloud feature vector to improve vessel labeling.

Figure 3 Schematic of the coordinate system used in SAD-PointNet for spatial labeling. (A) The x–y plane projection illustrates the horizontal spatial relationship of the coronary trajectory relative to the normalized cardiac coordinate system. (B) The x–z plane projection illustrates the longitudinal spatial relationship of the coronary trajectory relative to the normalized cardiac coordinate system. SAD-PointNet, Self-Attention Distillation PointNet.

Training of the neural network

All networks and experiments were conducted using PyTorch 2.0 and executed on two NVIDIA Tesla V100S (40 GB) GPUs. The network was trained with mini-batches consisting of eight cropped CTA images, resulting in a total of 600 training epochs. The Adam optimizer was utilized with an initial learning rate set at 10⁻⁴. In the case of the chamber segmentation and coronary segmentation models, the average Dice similarity coefficient was employed as the loss function, while for the parsing model, focal loss was used. To determine the optimal training epoch for each model, we selected the epoch that yielded the lowest validation data loss.

Internal data

Supervised by a senior radiologist, a team consisting of two proficient technicians meticulously performed manual segmentation on an internal dataset comprising 8,025 cases using ITK-SNAP 3.8 software. To ensure accuracy, the resulting segmentations were further assessed by two radiologists with three or more years of experience in order to minimize any potential errors in segmentation. These meticulously obtained segmentations were then utilized for both training and validating the chamber segmentation, coronary artery segmentation and coronary artery parsing models.

Automatic FAI calculation

FAI was measured in the proximal segments of the three major epicardial coronary arteries: RCA, LAD, and LCX. To improve interindividual consistency, predefined fixed-length proximal segments were analyzed along the vessel centerline. Because the ostial RCA may be influenced by adjacent aortic structures, the most proximal 10 mm was excluded, and the 10–50 mm segment of the RCA was analyzed. For the LAD and LCX, the proximal 40 mm segment of each vessel was analyzed. The left main coronary artery was not included because of marked interindividual variability in its length.

PVAT was defined as adipose voxels located within a radial distance from the outer vessel wall equal to the local vessel diameter and with CT attenuation values between −190 and −30 Hounsfield units (HU). To account for attenuation differences across tube voltage settings, FAI values obtained at 100 kV were divided by a conversion factor of 1.11485 to ensure comparability with those obtained at 120 kV.

Measurement time assessment

Processing time was recorded to compare workflow efficiency between automated and manual FAI measurement. For the automated workflow, processing time was defined as the interval from loading anonymized DICOM images into the software to generation of vessel-specific FAI results, including preprocessing, coronary segmentation, centerline extraction, branch labeling, PVAT extraction, and FAI calculation. For manual measurement, processing time was defined as the interval from opening the CCTA dataset to completion of FAI measurement for the three major coronary arteries. Manual measurement time was recorded during routine analysis by the experienced reader. Because computational time depends on hardware configuration and image size, the reported 30-second processing time should be interpreted as workflow-specific rather than a hardware-independent metric.

Statistical analysis

Continuous data are presented as mean ± standard deviation. Normality was tested with the Shapiro-Wilk test. The correlation and agreement between the standard reference and the automatic software were assessed with Pearson’s correlation coefficient (r) and ICC. Bland-Altman plots displayed bias and limits of agreement (LoA) within the 95% confidence interval (CI). Generalized linear regression models were used to evaluate the influence of manual measurements on DL-FAI. Model performance was assessed using the Akaike information criterion (AIC), root mean square error (RMSE), and adjusted R-squared values. A two-sided P<0.05 was considered statistically significant. All analyses were performed using IBM SPSS v.24 and R version 4.3.0. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the institutional Ethics Committees of Shanghai General Hospital (approval No. 2023049) and Zhongshan Hospital, Fudan University (approval No. B2020088R). The requirement for informed consent was waived because of the retrospective design and anonymized data processing.

Results

Patients and vessel characteristics

The training and internal validation sets included a total of 8,025 patients. The external test set comprised 209 CAD patients [age: 60.1±8.0 years, male: 138 (66.0%)]. A total of 627 vessels were analyzed in the external set (209 each for LAD, LCX, and RCA).

Comparison of DL-FAI and manual FAI

The automated DL-FAI workflow required approximately 30 seconds per case on the specified workstation, whereas manual measurement required approximately 15 minutes per case in the routine reader workflow, corresponding to an estimated 96% reduction in post-processing time.

Bland-Altman analysis demonstrated high agreement between the two methods. The mean bias (DL-FAI minus manual FAI) was 2.69 HU (LoA: −3.60 to 8.98) for LAD, 3.65 HU (LoA: −3.99 to 11.29) for LCX, and 2.03 HU (LoA: −3.71 to 7.77) for RCA. Across all vessels, 95.22% of the measurements fell within the 95% LoA (Table 1).

Correlation analysis revealed excellent consistency. Correlation analysis revealed strong consistency. The overall Pearson correlation coefficient for all vessels was 0.913 (P<0.001). Subgroup analysis confirmed strong correlations for each vessel type: LAD (r=0.910), LCX (r=0.874), and RCA (r=0.917). Furthermore, the ICC for the overall agreement was 0.857 (95% CI: 0.834–0.876), demonstrating excellent reliability (Table 2).

Linear regression analysis further substantiated the robust relationship. For the total cohort, manual FAI was a significant predictor of DL-FAI (R²=0.833), with a regression coefficient of 0.972 (95% CI: 0.938 to 1.006) and a non-significant intercept (P=0.65), indicating a near 1:1 linear relationship. Detailed regression parameters are presented in Table 3.

Subgroup analysis: impact of calcification

To preliminarily assess the performance of the automated method under challenging imaging conditions, a subgroup analysis was performed in 58 vessels with evident calcification. In this subgroup, DL-FAI remained highly correlated with manual reference measurement (Pearson r=0.952, P<0.001), with an R² of 0.906. Bland-Altman analysis showed a mean bias of −2.87 HU, and 93.1% of measurements fell within the LoA.

These findings suggest that the method retained a certain degree of robustness in vessels with evident calcification. However, given the limited sample size and the absence of an independent ground truth, these results should not be interpreted as evidence that the automated method is superior to manual measurement.

Discussion

In this study, we developed and preliminarily validated a DL-based workflow for automated coronary FAI quantification. This workflow integrates cardiac chamber segmentation, coronary artery segmentation with continuity refinement, and coronary artery parsing with anatomical label assignment, enabling automated extraction of the coronary tree, centerline localization, standardized segment identification, PVAT extraction, and final FAI calculation with minimal manual intervention. External validation showed good agreement between DL-FAI and manual reference measurement, together with a substantial reduction in post-processing time, supporting its potential value for standardized and large-scale analysis.

FAI has attracted attention because it may reflect the response of PVAT to changes in the local vascular micro-environment (18-22), thereby providing information beyond luminal stenosis alone. However, its broader application has been limited by a complex workflow, strong operator dependence, and variability related to reader experience (23,24). In this context, the main contribution of our study is not to re-establish the clinical significance of FAI itself, but rather to address a more fundamental methodological question: whether FAI can be measured in a relatively stable, rapid, and standardized manner using an automated pipeline.

From the perspective of CCTA automation research, previous DL studies have mainly focused on coronary stenosis, plaque burden, or vessel reconstruction. In contrast, the present work extends automated analysis to FAI, an imaging phenotype related to the perivascular microenvironment. Its task orientation also differs from conventional coronary segmentation studies: rather than optimizing a single segmentation step alone, our approach was designed as a complete workflow serving the final purpose of standardized FAI quantification.

Methodologically, we evaluated the automated workflow mainly in terms of agreement and efficiency relative to manual reference measurement. The results showed good agreement at the group level, with most measurements falling within the Bland-Altman LoA, and with supportive ICC and linear regression results. At the same time, the average processing time was reduced from approximately 15 minutes to approximately 30 seconds, highlighting its practical value for batch processing and standardized analysis.

Nevertheless, some directional bias remained between DL-FAI and manual reference measurement, and the magnitude of bias varied across vessel types. This suggests that although the automated method can reproduce manual measurement trends reasonably well at the group level, it should not be assumed to be fully interchangeable with manual analysis in every individual case. This is particularly relevant for the LCX, where the ICC CI was relatively wide, indicating potential instability related to anatomical complexity, local boundary definition, minor centerline deviations, or differences in PVAT extraction range. Therefore, the current findings support DL-FAI as an important alternative to manual analysis, rather than an unconditional replacement in all scenarios.

In the subgroup with evident calcification, DL-FAI still showed high correlation and acceptable agreement with manual measurement, suggesting a certain degree of robustness under complex imaging conditions. This is clinically relevant because calcification may interfere with coronary boundary delineation and attenuation assessment. However, the subgroup sample size was limited, and no independent ground truth was available to determine whether either method was closer to the true biological state. Accordingly, these findings should be regarded as supplementary evidence of feasibility in challenging conditions rather than proof of superiority over manual measurement.

The findings should also be interpreted in light of the biological nature of FAI itself. Although FAI is widely used as an imaging phenotype related to local coronary inflammation, it remains an indirect imaging marker and may be influenced not only by local inflammatory activity but also by scan parameters, local anatomy, and systemic metabolic background. Thus, what this study validates is primarily whether an automated method can reproduce manual reference measurement with reasonable stability, not whether it directly quantifies the true inflammatory state.

Several limitations should be noted. First, manual measurement used as the reference standard, but it is not an absolute biological ground truth; therefore, the absolute accuracy of DL-FAI cannot be established from this study alone. Second, the external validation analysis was conducted mainly at the vessel level, whereas the three vessels originated from the same patient; thus, within-patient correlation should be considered when interpreting the statistical inference. Third, the external validation sample size was still limited and involved a certain degree of clinical selection, so further validation is needed in broader CCTA populations. Fourth, although a calcification subgroup analysis was performed, systematic evaluation remains lacking in heavily calcified vessels, cases with substantial motion artifact, segments adjacent to stents, and more complex anatomical variants. Fifth, body mass index, body composition, and other systemic metabolic parameters were not included, limiting further interpretation of the relationship between automated FAI and whole-body metabolic status. Future studies should further determine whether automated FAI quantification provides incremental value for CAD diagnosis, risk prediction, treatment monitoring, or clinical decision-making beyond conventional CCTA findings such as stenosis severity and plaque characteristics.

Overall, our findings suggest that automated coronary FAI quantification based on DL is feasible. Compared with conventional manual or semiautomated analysis, the proposed method demonstrated good agreement and substantial efficiency advantages in external validation, while also showing preliminary stability in calcified vessels. Its principal contribution lies in providing a methodological foundation for standardized, scalable, and reproducible quantification of this inflammation-related imaging biomarker.

Conclusions

We developed a DL-based workflow for automated coronary FAI quantification. In the external validation cohort, this method showed good agreement with manual reference measurement and substantially reduced post-processing time; it also demonstrated preliminary stability in vessels with evident calcification. These findings support DL-FAI as a feasible approach for automated coronary FAI analysis and provide a methodological basis for future clinical translation studies.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://rx.amegroups.com/article/view/10.21037/rx-26-0027/rc

Data Sharing Statement: Available at https://rx.amegroups.com/article/view/10.21037/rx-26-0027/dss

Peer Review File: Available at https://rx.amegroups.com/article/view/10.21037/rx-26-0027/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://rx.amegroups.com/article/view/10.21037/rx-26-0027/coif). The 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. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the institutional Ethics Committees of Shanghai General Hospital (approval No. 2023049) and Zhongshan Hospital, Fudan University (approval No. B2020088R). The requirement for informed consent was waived because of the retrospective design and anonymized data processing.

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. Senecal C, Widmer RJ, Lerman LO, et al. Association of Search Engine Queries for Chest Pain With Coronary Heart Disease Epidemiology. JAMA Cardiol 2018;3:1218-21. [Crossref] [PubMed]
2. Sanchis-Gomar F, Perez-Quilis C, Leischik R, et al. Epidemiology of coronary heart disease and acute coronary syndrome. Ann Transl Med 2016;4:256. [Crossref] [PubMed]
3. Hajra A, Li Y, Siu S, et al. Risk of coronary disease in the South Asian American population. J Am Coll Cardiol 2013;62:644-5. [Crossref] [PubMed]
4. Anderson L, Brown JP, Clark AM, et al. Patient education in the management of coronary heart disease. Cochrane Database Syst Rev 2017;6:CD008895. [Crossref] [PubMed]
5. Blumenthal JA, Sherwood A, Babyak MA, et al. Exercise and pharmacological treatment of depressive symptoms in patients with coronary heart disease: results from the UPBEAT (Understanding the Prognostic Benefits of Exercise and Antidepressant Therapy) study. J Am Coll Cardiol 2012;60:1053-63. [Crossref] [PubMed]
6. Mega JL, Stitziel NO, Smith JG, et al. Genetic risk, coronary heart disease events, and the clinical benefit of statin therapy: an analysis of primary and secondary prevention trials. Lancet 2015;385:2264-71. [Crossref] [PubMed]
7. Williams MC, Hunter A, Shah ASV, et al. Use of Coronary Computed Tomographic Angiography to Guide Management of Patients With Coronary Disease. J Am Coll Cardiol 2016;67:1759-68. [Crossref] [PubMed]
8. Oikonomou EK, Marwan M, Desai MY, et al. Non-invasive detection of coronary inflammation using computed tomography and prediction of residual cardiovascular risk (the CRISP CT study): a post-hoc analysis of prospective outcome data. Lancet 2018;392:929-39. [Crossref] [PubMed]
9. Sagris M, Antonopoulos AS, Simantiris S, et al. Pericoronary fat attenuation index-a new imaging biomarker and its diagnostic and prognostic utility: a systematic review and meta-analysis. Eur Heart J Cardiovasc Imaging 2022;23:e526-36. [Crossref] [PubMed]
10. Theofilis P, Sagris M, Oikonomou E, et al. Inflammatory Mechanisms Contributing to Endothelial Dysfunction. Biomedicines 2021;9:781. [Crossref] [PubMed]
11. Antonopoulos AS, Sanna F, Sabharwal N, et al. Detecting human coronary inflammation by imaging perivascular fat. Sci Transl Med 2017;9:eaal2658. [Crossref] [PubMed]
12. Fu F, Shan Y, Yang G, et al. Deep Learning for Head and Neck CT Angiography: Stenosis and Plaque Classification. Radiology 2023;307:e220996. [Crossref] [PubMed]
13. Huang B, Sollee J, Luo YH, et al. Prediction of lung malignancy progression and survival with machine learning based on pre-treatment FDG-PET/CT. EBioMedicine 2022;82:104127. [Crossref] [PubMed]
14. Zeleznik R, Foldyna B, Eslami P, et al. Deep convolutional neural networks to predict cardiovascular risk from computed tomography. Nat Commun 2021;12:715. [Crossref] [PubMed]
15. Milletari F, Navab N, Ahmadi SA. V-Net: fully convolutional neural networks for volumetric medical image segmentation. arXiv:1606.04797v1 [Preprint]. 2016. Available online: https://doi.org/10.48550/arXiv.1606.04797.
16. Qi CR, Yi L, Su H, et al. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. arXiv:1706.02413v1 [Preprint]. 2017. Available online: https://doi.org/10.48550/arXiv.1706.02413.
17. Qi CR, Su H, Mo K, et al. PointNet: deep learning on point sets for 3D classification and segmentation. arXiv:1612.00593v2 [Preprint]. 2017. Available online: https://doi.org/10.48550/arXiv.1612.00593.
18. Kaptoge S, Seshasai SR, Gao P, et al. Inflammatory cytokines and risk of coronary heart disease: new prospective study and updated meta-analysis. Eur Heart J 2014;35:578-89. [Crossref] [PubMed]
19. Zhang X, Bishawi M, Zhang G, et al. Modeling early stage atherosclerosis in a primary human vascular microphysiological system. Nat Commun 2020;11:5426. [Crossref] [PubMed]
20. Meng B, Li Y, Ding Y, et al. Myeloid-derived growth factor inhibits inflammation and alleviates endothelial injury and atherosclerosis in mice. Sci Adv 2021;7:eabe6903. [Crossref] [PubMed]
21. Chang L, Villacorta L, Li R, et al. Loss of perivascular adipose tissue on peroxisome proliferator-activated receptor-γ deletion in smooth muscle cells impairs intravascular thermoregulation and enhances atherosclerosis. Circulation 2012;126:1067-78. [Crossref] [PubMed]
22. Antoniades C, Kotanidis CP, Berman DS. State-of-the-art review article. Atherosclerosis affecting fat: What can we learn by imaging perivascular adipose tissue?. J Cardiovasc Comput Tomogr 2019;13:288-96. [Crossref] [PubMed]
23. van Velzen SGM, Lessmann N, Velthuis BK, et al. Deep Learning for Automatic Calcium Scoring in CT: Validation Using Multiple Cardiac CT and Chest CT Protocols. Radiology 2020;295:66-79. [Crossref] [PubMed]
24. Winkel DJ, Suryanarayana VR, Ali AM, et al. Deep learning for vessel-specific coronary artery calcium scoring: validation on a multi-centre dataset. Eur Heart J Cardiovasc Imaging 2022;23:846-54. [Crossref] [PubMed]