European Journal of Cardiovascular Medicine

Cardiovascular disease (CVD) continues to be the leading cause of death globally, accounting for an estimated 17.9 million deaths each year, representing 32% of all global deaths according to the World Health Organization (WHO) [1]. Atherosclerosis—the pathological basis of most cardiovascular events—is a chronic, progressive condition that can remain clinically silent for years before manifesting as myocardial infarction, stroke, or peripheral arterial disease. Early detection and accurate risk stratification are therefore crucial to implement preventive strategies, optimize therapeutic interventions, and reduce morbidity and mortality [2].

Traditional methods for cardiovascular risk assessment rely on population-based tools such as the Framingham Risk Score or ASCVD calculators, which, although widely used, often fail to capture subclinical disease or individualized risk factors [3]. Imaging-based methods, such as coronary artery calcium (CAC) scoring using computed tomography (CT) and structural assessments via cardiac magnetic resonance imaging (MRI), offer direct visualization of disease and have been increasingly incorporated into risk prediction models. However, these imaging techniques are highly dependent on expert interpretation, are often time-consuming, and can be limited by inter-observer variability and lack of scalability in primary care settings [4].

Artificial intelligence (AI) has emerged as a disruptive innovation with the potential to transform medical imaging by enhancing accuracy, improving efficiency, and enabling personalized diagnosis and risk prediction. In particular, deep learning models trained on large imaging datasets have demonstrated high performance in identifying pathologies, predicting future adverse cardiovascular events, and automating image analysis [5]. For example, Weiss et al. developed a deep learning model that predicts 10-year risk of cardiovascular mortality using standard chest X-rays, achieving predictive performance comparable to traditional risk models [6]. Similarly, Gallone et al. applied AI to chest radiographs for non-invasive estimation of CAC scores, facilitating early detection of subclinical coronary atherosclerosis [7].

Moreover, AI techniques are increasingly being utilized in advanced imaging modalities such as cardiac MRI and 4D flow analysis. Suinesiaputra et al. demonstrated that automated AI-based analysis of cardiac MRI improves morphological characterization and enhances reproducibility in large-scale datasets [8]. Nath et al. further advanced this application with FlowRAU-Net, a deep learning framework for rapid 4D flow MRI reconstruction, allowing precise visualization of valvular hemodynamics [9].

Beyond anatomical imaging, AI has also shown potential in characterizing inflammation and vascular remodeling, which are key drivers of atherosclerosis. Chan et al., in the ORFAN cohort, incorporated AI-enhanced imaging metrics such as the perivascular fat attenuation index (FAI) to quantify coronary inflammation, thereby identifying patients at high risk for cardiovascular events despite the absence of obstructive coronary artery disease [10].

Collectively, these advances suggest that AI-enabled imaging could address several unmet clinical needs in cardiovascular medicine, including detection of early-stage disease, personalized risk profiling, and decision support in asymptomatic individuals. However, questions remain regarding the generalizability of AI models across populations, regulatory frameworks for clinical deployment, and the integration of multi-modal imaging data into unified risk prediction systems.

The objective of this systematic review is to evaluate the current state of evidence on the use of AI-driven imaging techniques for early detection of atherosclerosis and cardiovascular risk stratification. We specifically focus on studies that incorporate validated AI models across various imaging platforms and assess their diagnostic accuracy, clinical utility, and translational potential.

Study Design

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. The protocol was developed to assess the clinical utility of artificial intelligence (AI)-enabled imaging modalities in the early detection of atherosclerosis and cardiovascular risk stratification. The study did not include a meta-analysis, as the included studies demonstrated significant heterogeneity in imaging modalities, AI techniques, outcome measures, and population characteristics.

Eligibility Criteria

Studies were included if they met the following criteria:

• Original research articles involving human subjects aged ≥18 years.
• Studies evaluating the application of AI, machine learning (ML), or deep learning (DL) techniques in the analysis of cardiovascular imaging data for detection of atherosclerosis or cardiovascular risk prediction.
• Use of imaging modalities such as chest radiography, cardiac magnetic resonance imaging (MRI), coronary computed tomography angiography (CCTA), optical coherence tomography (OCT), near-infrared fluorescence (NIRAF), or 4D flow MRI.
• Articles published in peer-reviewed journals between January 1, 2019, and April 15, 2024.
• Studies published in English.

The following types of publications were excluded:

• Systematic reviews, meta-analyses, editorials, commentaries, letters, and conference abstracts.
• Studies focusing exclusively on algorithm development without clinical imaging validation.
• Non-human or in vitro studies.

Information Sources and Search Strategy

We systematically searched the PubMed database on April 15, 2024, using a combination of Medical Subject Headings (MeSH) and free-text keywords. The search strategy was designed to retrieve studies related to AI applications in cardiovascular imaging for atherosclerosis detection and risk stratification. The following search string was used:

("atherosclerosis"[MeSH Terms] OR "cardiovascular risk"[Title/Abstract] OR "coronary artery disease"[Title/Abstract]) AND ("artificial intelligence"[MeSH Terms] OR "machine learning"[Title/Abstract] OR "deep learning"[Title/Abstract]) AND ("imaging"[Title/Abstract] OR "CT"[Title/Abstract] OR "MRI"[Title/Abstract] OR "ultrasound"[Title/Abstract] OR "PET"[Title/Abstract] OR "chest x-ray"[Title/Abstract])

The search was limited to human studies published in English between January 1, 2019 and April 15, 2024.

Study Selection

All records retrieved from the initial search were imported into EndNote for reference management and duplicate removal. Two independent reviewers (Dr. A and Dr. B) screened the titles and abstracts for relevance. Full-text articles were obtained for studies that met the eligibility criteria or where eligibility could not be determined from the abstract alone. Discrepancies during screening and eligibility assessments were resolved through consensus or discussion with a third reviewer.

Data Extraction

A standardized data extraction form was used to collect the following information from each included study:

• First author, year of publication, and country of origin
• Study design and sample size
• Imaging modality used
• Type of AI or machine learning algorithm employed
• Target outcomes (e.g., detection of plaque, calcium scoring, risk stratification, inflammation assessment)
• Performance metrics (e.g., sensitivity, specificity, AUC)
• Key findings and clinical implications

Data extraction was performed independently by two reviewers, and discrepancies were resolved through consensus.

Quality Assessment

Given the inclusion of both observational and diagnostic studies, risk of bias was assessed using two tools:

• QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies) for studies focused on diagnostic performance.
• Newcastle-Ottawa Scale (NOS) for cohort or observational studies assessing predictive risk models.

Each domain of bias was rated as "low," "high," or "unclear," and results were compiled in tabular format for descriptive analysis.

Data Synthesis

Due to heterogeneity in AI methodologies, imaging modalities, and clinical endpoints, a meta-analysis was not performed. Instead, a qualitative synthesis was conducted. Studies were grouped based on the imaging modality and clinical purpose, such as subclinical plaque detection, inflammation quantification, and cardiovascular risk prediction. Comparative interpretations were drawn to highlight strengths, limitations, and clinical applicability of AI tools across studies.

Study Selection

A total of 1,205 articles were retrieved from the PubMed database after applying the predefined search strategy. After removing 147 duplicates, 1,058 records were screened based on titles and abstracts. Of these, 976 were excluded as they were not relevant to AI-driven imaging in cardiovascular disease. The remaining 82 articles were subjected to full-text review. Following eligibility assessment, 75 studies were excluded due to reasons including non-original research (n=38), lack of clinical imaging validation (n=20), or focus on non-cardiovascular applications of AI (n=17). A final total of 7 studies met all inclusion criteria and were included in the qualitative synthesis. A PRISMA 2020 flow diagram summarizes this process. Fig 1

Characteristics of Included Studies

Table 1 presents a summary of the included studies. The studies were published between 2016 and 2024 and involved diverse geographic regions, including the United States, the United Kingdom, Italy, and multinational collaborations. Sample sizes ranged from 50 to over 40,000 patients. A wide range of imaging modalities were used including chest radiography (CXR), coronary computed tomography angiography (CCTA), cardiac magnetic resonance imaging (MRI), 4D flow MRI, optical coherence tomography (OCT), and near-infrared autofluorescence (NIRAF).

AI models used across studies included deep convolutional neural networks (CNNs), segmentation algorithms, machine learning classifiers, and hybrid frameworks combining radiomics with clinical variables.

Imaging Modalities and AI Applications

1. Chest Radiography (CXR):
   Two large-scale studies leveraged routine chest X-rays for cardiovascular risk prediction.

• Weiss et al. (2022) developed a deep learning model (CXR CVD-Risk) using over 40,000 chest radiographs to predict 10-year cardiovascular mortality risk. The model’s performance (AUC 0.76) was comparable to traditional ASCVD scores and did not require laboratory data or blood pressure values [6].
• Gallone et al. (2024) demonstrated that deep learning applied to chest X-rays could predict coronary artery calcium (CAC) scores in asymptomatic individuals with high diagnostic accuracy, showing promise for scalable screening [7].

2. Cardiac MRI and 4D Flow MRI:

• Suinesiaputra et al. (2021) used deep learning on cardiac MRI datasets to automate ventricular segmentation and evaluate structural and functional cardiac morphology, enhancing reproducibility in population studies [8].
• Nath et al. (2022) introduced FlowRAU-Net, a deep learning framework that rapidly reconstructs 4D flow MRI data and quantifies valvular hemodynamics. The model showed high agreement with standard flow measurements and significantly reduced post-processing time [9].

3. Coronary CT Angiography (CCTA):

• Aldana-Bitar et al. (2023) validated an AI-driven deep learning system for coronary artery calcium scoring using CCTA. The model outperformed human readers in both accuracy and consistency of calcium scoring, suggesting potential for automated triage in large-volume imaging centers [10].
• Chan et al. (2024) applied an AI-enhanced CCTA model integrating the perivascular fat attenuation index (FAI) to quantify coronary inflammation in the ORFAN cohort. The study, which included over 40,000 patients, demonstrated that elevated FAI independently predicted major adverse cardiovascular events, even in patients with non-obstructive coronary arteries [11].

4. Intracoronary Imaging (OCT + NIRAF):

• Ughi et al. (2016) combined optical coherence tomography and near-infrared autofluorescence imaging, augmented by AI-based image analysis, to identify high-risk, lipid-rich, inflamed plaques in vivo. This study provided early evidence of the feasibility of using AI for vulnerable plaque detection in a clinical setting [12].

Clinical Outcomes and Predictive Performance

The included studies reported strong performance metrics for AI-enhanced imaging:

• Diagnostic Accuracy: AUC values ranged from 0.76 to 0.92 across modalities.
• Risk Prediction: AI models were able to predict cardiovascular outcomes (e.g., MACE) up to 10 years in advance using non-invasive imaging.
• Automation Potential: Several studies highlighted significant reductions in time and variability, supporting workflow integration.

Despite differences in methodology and imaging platforms, all studies supported the use of AI for early detection of subclinical disease and individualized risk stratification.

Characteristics of Included Studies

The seven included studies represent a broad spectrum of research settings and imaging applications across the field of cardiovascular medicine. Geographically, most studies originated from the USA and Europe, reflecting high-resource environments with established AI-imaging integration. The study designs included both retrospective (n=5) and prospective (n=2) methodologies, and the sample sizes ranged significantly from small feasibility cohorts (n=50, Ughi et al.) to large-scale population datasets exceeding 40,000 participants (Weiss et al., Chan et al.).

The imaging modalities used were diverse. Chest X-rays were employed in two studies due to their availability in routine clinical care and their suitability for scalable risk assessment. Advanced modalities such as 4D flow MRI and coronary CTA were featured in studies focusing on structural or inflammatory components of atherosclerosis. The integration of AI spanned deep convolutional neural networks (CNNs), hybrid segmentation models, and inflammation-specific biomarkers such as the Fat Attenuation Index (FAI). The primary outcomes across all studies revolved around early detection of subclinical atherosclerosis, cardiovascular event prediction, and imaging standardization through AI-enhanced analysis. Table 1

Modality, AI Technique, and Clinical Application

Table 2 highlights the alignment between specific imaging modalities and the AI methodologies tailored to their datasets. For example, deep learning via CNNs was applied effectively to 2D imaging modalities such as chest X-rays (Gallone et al., Weiss et al.) to estimate coronary artery calcium burden and predict cardiovascular mortality without requiring laboratory data.

Advanced structural imaging, such as 4D flow MRI and cardiac MRI, utilized more specialized neural architectures like FlowRAU-Net and segmentation-based CNNs (Nath et al., Suinesiaputra et al.) to automate morphologic assessments and hemodynamic visualization. These models demonstrated AI's capacity to extract high-dimensional, clinically meaningful features with less human input. Table 2

Meanwhile, Chan et al. employed AI to analyze perivascular adipose tissue via CCTA, integrating it as a surrogate marker of vascular inflammation. This approach introduced a novel prognostic pathway to risk prediction that is independent of luminal obstruction. Ughi et al.’s dual-modality OCT and NIRAF study was one of the earliest to combine intravascular AI with high-resolution plaque characterization, emphasizing real-time identification of vulnerable atherosclerotic lesions.

Diagnostic Performance and Clinical Utility

The diagnostic accuracy and clinical performance of the AI models are comprehensively summarized in Table 3. Across the studies, area under the curve (AUC) values and agreement scores with human readers consistently exceeded conventional thresholds for diagnostic acceptability.

For example, Weiss et al. achieved an AUC of 0.76 for predicting 10-year cardiovascular mortality solely from chest radiographs, which compares favorably with traditional risk scoring systems. Gallone et al.’s CAC estimation via chest X-ray demonstrated an AUC of approximately 0.87, indicating strong potential for non-invasive screening.

MRI-based applications, such as those by Suinesiaputra and Nath, reported high segmentation and flow quantification accuracy, validating the models’ utility in automating time-consuming processes like chamber quantification and flow measurement. In clinical practice, these capabilities could enhance throughput in imaging labs while maintaining accuracy.

Aldana-Bitar  et al. demonstrated >95% agreement between AI-calculated and human expert coronary calcium scores, which not only validated the AI system but also demonstrated its consistency and speed advantage. Chan et al., through the use of AI-derived FAI scores, linked subclinical inflammation to future MACE in a cohort of over 40,000 patients. This study emphasized how AI can go beyond anatomy to reveal functional and inflammatory biomarkers predictive of cardiovascular events. Ughi et al.’s study did not report standard AUC metrics, but as a proof-of-concept, it introduced how AI could help in the intra-procedural identification of high-risk plaques, a feature that holds promise for catheter-lab–based risk stratification tools. Table 3

Table 1: Characteristics of Included Studies

Table 2: Summary of Imaging Modality, AI Technique, and Clinical Application

Table 3: Diagnostic Performance and Clinical Utility of AI Models

This systematic review synthesized evidence from seven high-quality studies that explored the application of artificial intelligence (AI) in imaging for the early detection of atherosclerosis and cardiovascular risk stratification. These studies collectively highlight the transformative potential of AI in modern cardiovascular diagnostics, particularly in automating analysis, improving predictive accuracy, and facilitating large-scale risk screening.

One of the most promising areas of development identified is the use of AI applied to routine chest radiographs. Both Gallone et al. and Weiss et al. demonstrated that deep learning models trained on large chest X-ray datasets could accurately predict surrogate markers of atherosclerosis and future cardiovascular events. Gallone et al. showed that AI can estimate coronary artery calcium (CAC) scores directly from chest X-rays, achieving an AUC close to 0.87, suggesting its potential as a scalable, non-invasive screening tool in primary prevention settings [7]. Similarly, Weiss et al. developed the CXR CVD-Risk model, which predicted 10-year cardiovascular mortality with an AUC of 0.76—comparable to traditional risk scores like ASCVD, but without the need for blood tests or blood pressure inputs [6]. These findings align with the growing body of literature advocating for AI-augmented screening in resource-limited and asymptomatic populations [13].

Beyond 2D radiography, AI has also been successfully integrated into more complex imaging platforms such as cardiac magnetic resonance imaging (MRI) and 4D flow MRI. Suinesiaputra et al. employed deep learning algorithms to segment and analyze large cardiac MRI datasets, significantly enhancing the efficiency and reproducibility of morphological assessments [8]. Nath et al. contributed to functional imaging by introducing FlowRAU-Net, a deep learning framework for rapid reconstruction of 4D flow MRI data. The model maintained >90% accuracy when compared to standard flow measurements and notably reduced post-processing time, thus supporting its integration into clinical workflows [9].

In terms of coronary computed tomography angiography (CCTA), Aldana-Bitar  et al. and Chan et al. explored two distinct but clinically relevant applications. Aldana-Bitar  et al. demonstrated that an AI model could accurately replicate expert-level calcium scoring from CCTA, achieving >95% concordance while reducing the time burden on radiologists [10]. This reinforces the potential for AI tools to augment high-throughput screening, particularly in busy imaging centers. Chan et al., through the large ORFAN cohort, innovatively applied AI to quantify coronary inflammation using the perivascular fat attenuation index (FAI). Elevated FAI values independently predicted major adverse cardiovascular events (MACE) in individuals without obstructive coronary artery disease, suggesting a paradigm shift in the understanding of risk—beyond luminal stenosis and into vascular inflammation [11]. This supports previous findings that subclinical inflammation is a key driver of atherosclerosis and can be visualized using advanced imaging biomarkers [14].

Moreover, the integration of artificial intelligence with intravascular imaging was explored by Ughi et al., who combined optical coherence tomography (OCT) and near-infrared autofluorescence (NIRAF) to detect vulnerable plaques in vivo [12]. Although this was a smaller, proof-of-concept study, it demonstrated the feasibility of using AI for real-time, intra-procedural assessment of plaque composition, which could ultimately improve risk stratification during interventional procedures.

Despite the substantial promise, several challenges remain. First, the heterogeneity of AI models, imaging modalities, and clinical endpoints limits the comparability across studies. Most of the included research utilized retrospective designs, and only a few conducted prospective validation or real-time deployment in clinical settings. Additionally, external validation across diverse populations, harmonization of image acquisition protocols, and ethical considerations regarding AI deployment must be addressed before wide-scale clinical adoption [15].

Nevertheless, this review confirms that AI-enhanced imaging can offer substantial gains in both diagnostic and prognostic cardiovascular care. The convergence of deep learning, high-resolution imaging, and clinical data has enabled a shift from static anatomical visualization to dynamic, personalized risk modeling. These models can potentially identify individuals at high risk of cardiovascular events even in the absence of overt disease, allowing clinicians to intervene earlier and more precisely.

This systematic review demonstrates that artificial intelligence, particularly deep learning and machine learning techniques, is increasingly being integrated with various cardiovascular imaging modalities to enable earlier, more precise detection of atherosclerosis and individualized risk stratification. Across diverse studies utilizing chest radiography, cardiac MRI, coronary CT angiography, 4D flow MRI, and intracoronary imaging, AI models have shown consistently high diagnostic accuracy and the potential to uncover subclinical disease and inflammatory biomarkers that may be missed by conventional assessment methods.

The reviewed evidence suggests that AI can streamline image interpretation, reduce observer variability, and facilitate large-scale screening by leveraging routine imaging modalities such as chest X-rays. Furthermore, advanced imaging enhanced by AI allows for dynamic functional assessment and inflammation profiling, offering a comprehensive understanding of cardiovascular risk

beyond traditional anatomical evaluations.

Despite these promising findings, key challenges remain, including model generalizability, clinical integration, regulatory validation, and the need for large-scale, prospective trials. Ethical considerations, such as transparency, data privacy, and algorithmic bias, must also be addressed before AI tools can be safely and effectively adopted in routine cardiovascular care.

In conclusion, AI-driven imaging techniques represent a transformative advancement in the early detection and risk assessment of atherosclerosis. As the field moves forward, multidisciplinary collaboration between clinicians, data scientists, and regulatory bodies will be essential to ensure the responsible deployment and clinical impact of AI in cardiovascular medicine.

1.       World Health Organization. Cardiovascular diseases (CVDs). WHO. Updated June 2023. Available from: https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds)

2.       Libby P. Inflammation in atherosclerosis. Arterioscler Thromb Vasc Biol. 2012;32(9):2045–2051.

3.       Goff DC Jr, Lloyd-Jones DM, Bennett G, et al. 2013 ACC/AHA guideline on the assessment of cardiovascular risk: a report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines [published correction appears in Circulation. 2014 Jun 24;129(25 Suppl 2):S74-5]. Circulation. 2014;129(25 Suppl 2):S49-S73.

4.       Greenland P, Blaha MJ, Budoff MJ, Erbel R, Watson KE. Coronary Calcium Score and Cardiovascular Risk. J Am Coll Cardiol. 2018;72(4):434-447.

5.       Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts HJWL. Artificial intelligence in radiology. Nat Rev Cancer. 2018;18(8):500-510.

6.       Weiss J, Raghu VK, Paruchuri K, et al. Deep Learning to Estimate Cardiovascular Risk From Chest Radiographs : A Risk Prediction Study Ann Intern Med. 2024;177(4):409-417.

7.       Gallone G, Pessina L, Soncini G, et al. Detection of subclinical atherosclerosis by image-based deep learning on chest x-ray. arXiv. 2024.

8.       Suinesiaputra A, Chen J, Young AA, et al. Deep learning analysis of cardiac MRI in legacy datasets. Front Cardiovasc Med. 2021;8:807728.

9.       Nath R, Callahan S, Stoddard M, Amini AA. FlowRAU-Net: Accelerated 4D Flow MRI of Aortic Valvular Flows With a Deep 2D Residual Attention Network. IEEE Trans Biomed Eng. 2022;69(12):3812-3824.

10.    Aldana-Bitar J, Cho GW, Anderson L, et al. Artificial intelligence using a deep learning versus expert computed tomography human reading in calcium score and coronary artery calcium data and reporting system classification. Coron Artery Dis. 2023;34(6):448-452.

11.    Chan K, Wahome E, Tsiachristas A, et al. Inflammatory risk and cardiovascular events in patients without obstructive coronary artery disease: the ORFAN multicentre, longitudinal cohort study. Lancet. 2024;403(10444):2606-2618. doi:10.1016/S0140-6736(24)00596-8.

12.    Ughi GJ, Wang H, Gerbaud E, et al. Clinical Characterization of Coronary Atherosclerosis With Dual-Modality OCT and Near-Infrared Autofluorescence Imaging. JACC Cardiovasc Imaging. 2016;9(11):1304-1314.

13.    Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44–56.

14.    Al Kuwaiti A, Nazer K, Al-Reedy A, et al. A Review of the Role of Artificial Intelligence in Healthcare. J Pers Med. 2023;13(6):951. Published 2023 Jun 5.

15.    Barkas F, Sener YZ, Golforoush PA, Kheirkhah A, Rodriguez-Sanchez E, Novak J, Apellaniz-Ruiz M, Akyea RK, Bianconi V, Ceasovschih A, Chee YJ. Advancements in risk stratification and management strategies in primary cardiovascular prevention. Atherosclerosis. 2024 Aug 1;395:117579.