Roadmap on the use of artificial intelligence for imaging of vulnerable atherosclerotic plaque in coronary arteries

Bernhard Föllmer, Michelle C. Williams, Damini Dey, Armin Arbab‐Zadeh, Pál Maurovich‐Horvat, Rick Volleberg, Daniel Rueckert, Julia A. Schnabel, David E. Newby, Marc R. Dweck, Giulio Guagliumi, Volkmar Falk, Aldo J. Vázquez Mézquita, Federico Biavati, Ivana Išgum, Marc DeweyNature Reviews Cardiology 2023 · DOI

Introduction

Published in Nature Reviews Cardiology, this roadmap provides a strategic framework for integrating artificial intelligence (AI) into the detection and analysis of "vulnerable" atherosclerotic plaques. These fatty deposits in coronary arteries are prone to rupture, often leading to sudden myocardial infarction. By synthesizing expert consensus, the authors outline a path for AI to transition from a research tool to a clinical standard, enabling faster and more objective cardiac imaging.

The Challenge of Identifying Vulnerable Plaques

Coronary artery disease remains a leading global cause of mortality. A primary goal in preventative cardiology is the identification of vulnerable plaques—specifically thin-cap fibroatheromas (TCFA)—before they rupture.

Currently, identifying these features requires manual review of images from Coronary Computed Tomography Angiography (CCTA) or invasive modalities like Intravascular Ultrasound (IVUS) and Optical Coherence Tomography (OCT). This manual process is hindered by several factors:

  • Efficiency: Detailed analysis can exceed one hour per patient.
  • Subjectivity: High inter-reader variability often leads to inconsistent risk assessments.
  • Complexity: Quantifying exact volumes and textures of plaque components (e.g., lipid vs. calcium) is difficult for the human eye.
Imaging features of vulnerable atherosclerotic plaques across CCTA, IVUS, and OCT modalities.

AI Methodologies in Cardiac Imaging

The roadmap, developed through a multi-round Delphi consensus process involving 14 interdisciplinary experts, categorizes the primary AI approaches currently utilized:

  • Deep Learning (CNNs): Primarily used for automated segmentation, drawing vessel wall and plaque boundaries.
  • Radiomics: Extracts quantitative features (texture, shape, heterogeneity) invisible to the human eye, providing a "super-human" perspective on plaque vulnerability.
  • Multimodal Learning: Integrates imaging data with clinical variables (age, cholesterol, genetics) to generate holistic risk scores.
Hierarchy and applications of AI, machine learning, and deep learning in coronary imaging.

Datasets and Validation

The roadmap reviews evidence from landmark trials such as SCOT-HEART, PROSPECT, PROMISE, and ISCHEMIA. For instance, AI-quantified "low-attenuation plaque volume" (≥238 mm³) was associated with a fivefold higher risk of heart attack in the SCOT-HEART cohort.

However, a significant bottleneck remains: the lack of large, diverse, and open-access annotated datasets. Most current tools are trained on curated data that may not reflect the "messy" reality of clinical practice, such as images with motion artifacts or heavy calcification.

Key Insights: Speed, Precision, and Trust

  • Efficiency Gains: AI can reduce analysis time from approximately 25 minutes to under 6 seconds, facilitating real-time decision-making during invasive procedures.
  • The Trust Requirement: Experts emphasized that AI must be "explainable." Clinicians require visual overlays or heatmaps to understand why an algorithm flagged a specific area, rather than relying on a "black box" output.
  • Domain Shift: A major technical challenge is ensuring robustness across different hardware (e.g., different scanner manufacturers) and diverse patient populations.

Clinical Maturity and Future Directions

Several FDA-cleared and CE-marked tools (e.g., Cleerly, HeartFlow, Elucid) are already entering the market, though most currently require human over-reading. Despite this progress, hurdles remain:

  1. Artifacts: "Blooming" from calcium and motion blur continue to challenge AI accuracy.
  2. Evidence Gap: There is a need for large-scale Randomized Controlled Trials (RCTs) to prove that AI-guided treatment improves patient outcomes compared to standard care.
  3. Bias Mitigation: Models must be trained on diverse datasets to ensure equitable performance across different ethnic and age groups.

The authors advocate for a "human-in-the-loop" approach, where AI handles tedious quantification tasks, allowing physicians to focus on high-level clinical interpretation and patient communication.

Summary updated

BibTeX
@article{follmer2023roadmap,
  author = {Bernhard Föllmer and Michelle C. Williams and Damini Dey and Armin Arbab‐Zadeh and Pál Maurovich‐Horvat and Rick Volleberg and Daniel Rueckert and Julia A. Schnabel and David E. Newby and Marc R. Dweck and Giulio Guagliumi and Volkmar Falk and Aldo J. Vázquez Mézquita and Federico Biavati and Ivana Išgum and Marc Dewey},
  title = {{Roadmap on the use of artificial intelligence for imaging of vulnerable atherosclerotic plaque in coronary arteries}},
  journal = {Nature Reviews Cardiology},
  year = {2023},
  doi = {10.1038/s41569-023-00900-3},
}
Figure