Artificial intelligence in coronary plaque characterization and risk assessment: from images to impact

INTRODUCTION

IMAGING TECHNIQUES AND THEIR LIMITATIONS

Coronary computed tomography angiography (CCTA) is widely used for plaque detection, while intravascular ultrasound (IVUS) and optical coherence tomography (OCT) offer high-resolution views of the vessel wall^[3]. Although evidence suggests that specific plaque characteristics, such as a thin fibrous cap, large lipid core, and signs of inflammation, effectively predict future cardiovascular events, intra-coronary imaging often performs suboptimally in cath labs, limited by interobserver variability, manual analysis, and suboptimal identification of vulnerable plaque features^[4,5]. There is perhaps a critical need for methods able to integrate anatomical, morphological, and potentially functional data in a consistent and scalable manner, increasing exam reproducibility and improving workflow in cath labs^[6]. Coronary OCT resolves structures at roughly 10-20 µm axially and 20-40 µm laterally, with tissue penetration of about 1-2 mm, which helps distinguish thin fibrous caps and lipid pools with high contrast^[7]. By comparison, conventional 40-MHz IVUS has lower spatial resolution (about 100-150 µm axially and 200-250 µm laterally) but penetrates deeper into the vessel wall. Higher-frequency (~60-MHz) IVUS improves axial resolution relative to 40-MHz systems, although OCT remains finer overall^[7,8]. Reproducibility depends on what is being measured: OCT lumen measurements are generally highly reproducible, whereas fibrous-cap thickness shows only moderate agreement in routine practice [Intraclass Correlation Coefficient (ICC) ~0.52] that improves to ~0.88 with standardized methods; plaque-type classification reaches substantial agreement (κ ~0.71)^[8]. For IVUS, volumetric analyses-including Virtual Histology Intravascular Ultrasound (VH-IVUS)-show good-to-excellent reproducibility across centers for vessel and plaque volumes (ICC ~0.83-0.99)^[9].

AI IN CORONARY PLAQUE CHARACTERIZATION

AI algorithms have demonstrated high accuracy in automating plaque detection in CCTA, quantifying plaque burden, and distinguishing high-risk morphologies, such as low-attenuation plaque, positive remodelling, or the napkin-ring sign^[10,11]. Companies such as Cleerly, HeartFlow, and others have developed tools that provide rapid, reproducible, and quantitative assessments of plaque components on CCTA^[12,13]. Similarly, deep learning models have shown reliability in analyzing OCT and IVUS images for core-lipid characterization and fibrous cap thickness^[14,15]. The integration of AI into coronary imaging involves multiple stages, from image acquisition to automated plaque analysis and clinical decision support. Figure 1 illustrates a typical workflow for AI-assisted coronary plaque characterization across imaging modalities.

Figure 1. Schematic overview showing how AI integrates data from CCTA, OCT, and IVUS to identify high-risk plaque features, such as thin fibrous caps, lipid cores, and inflammation, for improved coronary risk assessment. (A) Image of coronary angiography from my clinical activity in Santa Maria Goretti Hospital, Latina; (B) Image of IVUS adapted from reference 26^[26]. Licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Changes: cropped and relabeled in GIMP; (C) Images of CCTA by MBq, via Wikimedia Commons, licensed under CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/). Changes: cropped/relabeled/recolored^[27]; (D) Image of OCT adapted from reference 28^[28]. Licensed under CC BY 4.0. Changes: cropped/relabeled/recolored. AI: Artificial intelligence; CCTA: coronary computed tomography angiography; OCT: optical coherence tomography; IVUS: intravascular ultrasound; GIMP: GNU image manipulation program.

AI makes it possible to combine clinical information such as age, cholesterol levels, and symptoms with imaging findings, offering a more comprehensive assessment than traditional tools such as Coronary Artery Disease - Reporting and Data System (CAD-RADS) or Agatston scores^[16]. Although the coronary artery calcium (CAC) score, most commonly calculated using the Agatston method, is a robust marker of total atherosclerotic burden and widely used for cardiovascular risk stratification, it has important limitations^[17]. By design, the CAC score quantifies only macroscopic calcified plaque and fails to capture non-calcified components, which are often more vulnerable and prone to rupture^[17]. Moreover, it does not differentiate between stable, densely calcified lesions and unstable plaques containing microcalcifications, low-attenuation components, or necrotic cores, features that are more directly linked to acute coronary events^[17,18]. Figure 2 provides an example of high-risk lipid-rich plaques in the context of IVUS/OCT/CCTA. As highlighted in recent literature, this lack of granularity reduces its utility in assessing plaque vulnerability and limits its predictive value for cardiovascular events in certain populations^[17,18]. This fusion of image-derived and clinical features represents a step forward in personalized cardiovascular risk stratification. AI tools can synthesize features from multiple imaging techniques, such as CCTA, OCT, and IVUS, to identify high-risk plaque components. Figure 1 provides an overview of this multimodal integration.

Figure 2. Representative intracoronary imaging of high-burden lipid-rich plaques. (A) Grayscale IVUS of the left circumflex artery showing an eccentric, predominantly hypoechoic plaque with marked deep ultrasound attenuation (arrow) and no bright calcific leading edge, consistent with attenuated plaque; (B) OCT from routine clinical imaging demonstrating a large, signal-poor, diffusely bordered lipid-rich plaque with long contiguous involvement on the longitudinal view. Scale bars =1 mm. IVUS: Intravascular ultrasound; OCT: optical coherence tomography. Images were acquired during routine care at our institution and are fully de-identified.

PROGNOSTIC APPLICATIONS AND RISK PREDICTION

AI models trained on large datasets have shown reliability in predicting major adverse cardiovascular events (MACEs), sometimes outperforming decisions made by clinicians^[19]. Recent findings from the CONFIRM2 Registry highlight how advances in quantitative imaging, particularly AI-derived plaque characteristics, provide a detailed and powerful tool for predicting Major Adverse Cardiovascular Events^[20]. These AI-based assessments offer a more comprehensive view of coronary artery disease burden than traditional human evaluations of stenosis severity alone^[20]. The Scottish Computed Tomography of the Heart (SCOT-HEART) study’s semi-automatic-based sub-analyses demonstrated that plaque burden was more strongly associated with MACE than human-read stenosis assessments^[21]. In a post hoc analysis involving 1,769 patients, those with low-attenuation plaque (LAP) burden exceeding 4% experienced a nearly fivefold increase in the risk of myocardial infarction (MI) compared to those with lower LAP burden {hazard ratio [HR] 4.65; 95%confidence interval [CI], 2.06-10.5; P < 0.001)^[21]. Additionally, the presence of adverse plaque features, such as positive remodeling or low-attenuation plaques, was associated with a threefold higher incidence of coronary heart disease death or nonfatal MI (HR 3.01; 95%CI, 1.61-5.63; P = 0.001)^[21]. Notably, patients exhibiting both obstructive disease and adverse plaque characteristics faced the highest event rates, with a tenfold increase in risk compared to individuals with normal coronary arteries (HR 11.50; 95%CI, 3.39-39.04; P < 0.001)^[21]. Important insights also come from the Decisions for Treating Coronary Disease are Changed in Patients Evaluated With Quantified Plaque Analysis (DECODE) study, which evaluated 100 patients who underwent CCTA^[12,22]. Initially, three expert cardiologists assessed each case and formulated a management plan based on standard CCTA interpretation and clinical information. Subsequently, AI-based quantitative plaque characterization was introduced^[20]. The inclusion of AI data led to a change in the initial management strategy in 66% of cases. Reclassification rates varied by risk profile: 47% in patients with a CAC score of 0, 96% in those with a CAC score > 400, and 89.5% in patients with stenoses ≥ 50%^[12,22]. Although limited, emerging evidence suggests that AI-guided imaging may offer comparable performance to conventional intravascular imaging in guiding percutaneous coronary intervention (PCI)^[22]. The Fully Automated Quantitative Coronary Angiography Versus Optical Coherence Tomography Guidance for Coronary Stent Implantation (FLASH) trial, a prospective, multicenter, randomized controlled study, evaluated the effectiveness of AI-based quantitative coronary angiography (AI-QCA) compared to OCT guidance during PCI^[22]. Involving 400 patients, the trial demonstrated noninferiority of AI-QCA-guided PCI in achieving optimal stent expansion, with a mean post-PCI minimal stent area (MSA) of 6.3 ± 2.2 mm² in the AI-QCA group versus 6.2 ± 2.2 mm² in the OCT group (P for noninferiority < 0.001)^[22]. At six-month follow-up, clinical outcomes, including target lesion failure and stent thrombosis, were comparable, supporting AI-QCA as a viable and practical alternative to intravascular imaging^[22]. Moreover, explainable AI (XAI) methods are being developed to visualize the contribution of each variable to the prediction, which may increase clinical trust and usability^[23]. A comparison between conventional imaging interpretation and AI-enhanced analysis is presented in Table 1, highlighting differences in reproducibility, integration, and prognostic value.

CHALLENGES IN IMPLEMENTATION

FUTURE DIRECTIONS

The future of artificial intelligence in plaque characterization will depend on several factors: the launch of large, multicenter prospective studies that assess real clinical outcomes; the use of federated learning to train AI models across institutions without sharing raw patient data, improving both privacy and generalizability. Current literature points toward the integration of AI models in daily practice to support clinicians, rather than replace human expertise. Progress will also be linked to the development of hybrid platforms that bring together imaging data, such as perfusion and plaque morphology, with molecular, genetic, or biomarker information to provide a more comprehensive view of disease. Another promising direction lies in the integration of imaging-derived data with clinical, laboratory, and non-invasive instrumental parameters^[13]. Combining high-sensitivity biomarkers [e.g., troponin, C-Reactive Protein (CRP)], clinical features (e.g., symptom presentation, risk factor burden), and echocardiographic findings (e.g., regional wall motion abnormalities) with imaging data could enable the development of complex and multimodal AI systems^[3,13]. To illustrate how hybrid platforms integrating imaging and clinical data may enhance coronary risk stratification in practice, we propose the following hypothetical clinical scenario: A 63-year-old man with hypertension and a history of elevated low-density lipoprotein (LDL) cholesterol presents with atypical chest pain. He undergoes coronary Computed Tomography (CT) angiography, which is analyzed using an advanced platform that integrates imaging findings with clinical and laboratory data. The automated report identifies high-risk plaque features, such as low-attenuation plaque and positive remodeling, and combines these with his risk factors, including blood pressure and lipid profile, to estimate his individualized cardiovascular risk. Although no significant stenosis is present, the overall assessment supports intensification of medical therapy with high-dose statins and lifestyle changes. The platform allows for follow-up imaging and biomarker tracking, enabling more precise monitoring over time and personalized adjustments to therapy.

These platforms may substantially enhance risk stratification and support personalized decision making in coronary artery disease. An overview of commercially available AI-based systems, including their imaging modality, availability, and core features, is provided in Table 3.

CONCLUSION

AI-based analysis of coronary plaque constitutes a great technological advancement with an increasing role in personalized cardiovascular care. Beyond its technical capabilities, the true value of AI in coronary imaging lies in its progressive integration into daily cardiology practice. AI tools can significantly reduce interpretation time by automating plaque quantification and risk assessment, freeing clinicians to focus on decision making and patient care. Embedding AI outputs directly into PACS or electronic health records can substantially improve communication and collaboration across the care team, supporting a shared understanding of risk profiles. Future improvements in accuracy, reproducibility, and scalability of imaging-based risk assessment by AI models will be determinant for their integration into routine clinical practice. Nonetheless, robust validation, ethical safeguards, and clinician engagement are essential to ensure that AI complements, rather than compromises, evidence-based cardiology.