[Paper Review] Automatic Coronary Artery Plaque Quantification and CAD-RADS Prediction using Mesh Priors

Rudolf L. M. van Herten, Nils Hampe+8 more|arXiv (Cornell University)|Oct 17, 2023

Cardiac Imaging and Diagnostics4 citations

TL;DR

This paper proposes a deep learning method that directly generates topologically valid coronary artery lumen and plaque surface meshes from CCTA scans using a centerline prior, enabling automated, accurate plaque quantification and CAD-RADS categorization. The approach achieves high intraclass correlation (0.98 for calcified plaque) and linearly weighted kappa of 0.75 for patient-level CAD-RADS scoring on a hold-out test set.

ABSTRACT

Coronary artery disease (CAD) remains the leading cause of death worldwide. Patients with suspected CAD undergo coronary CT angiography (CCTA) to evaluate the risk of cardiovascular events and determine the treatment. Clinical analysis of coronary arteries in CCTA comprises the identification of atherosclerotic plaque, as well as the grading of any coronary artery stenosis typically obtained through the CAD-Reporting and Data System (CAD-RADS). This requires analysis of the coronary lumen and plaque. While voxel-wise segmentation is a commonly used approach in various segmentation tasks, it does not guarantee topologically plausible shapes. To address this, in this work, we propose to directly infer surface meshes for coronary artery lumen and plaque based on a centerline prior and use it in the downstream task of CAD-RADS scoring. The method is developed and evaluated using a total of 2407 CCTA scans. Our method achieved lesion-wise volume intraclass correlation coefficients of 0.98, 0.79, and 0.85 for calcified, non-calcified, and total plaque volume respectively. Patient-level CAD-RADS categorization was evaluated on a representative hold-out test set of 300 scans, for which the achieved linearly weighted kappa ($κ$) was 0.75. CAD-RADS categorization on the set of 658 scans from another hospital and scanner led to a $κ$ of 0.71. The results demonstrate that direct inference of coronary artery meshes for lumen and plaque is feasible, and allows for the automated prediction of routinely performed CAD-RADS categorization.

Motivation & Objective

• To address the limitations of voxel-wise segmentation in coronary artery analysis, which often fails to produce anatomically plausible shapes.
• To automate the clinical workflow of plaque quantification and stenosis grading in CCTA by directly inferring surface meshes for lumen and plaque.
• To improve reproducibility and reduce inter-observer variability in CAD-RADS categorization by replacing manual or semi-automated assessment with a deep learning-based mesh generation framework.
• To ensure topological correctness of lumen and plaque structures through the use of a centerline prior, avoiding common artifacts in traditional segmentation.
• To evaluate the method on diverse data from multiple scanners and institutions, demonstrating robustness across different acquisition protocols and patient populations.

Proposed method

• The method uses a 1D convolutional neural network (1D ResNet) to process cylindrical data derived from centerline-traced coronary arteries, enabling spatial modeling along the vessel path.
• Surface meshes for the coronary lumen and plaque (calcified, non-calcified, and total) are directly predicted from the network output, ensuring topologically valid and anatomically plausible shapes.
• A centerline prior is used to guide the network in generating consistent, continuous meshes along the coronary artery tree, with mesh vertices aligned to the vessel path.
• The network is trained end-to-end on 2407 CCTA scans from multiple vendors and scanner types, with ground truth from expert annotations.
• CAD-RADS categorization is performed by analyzing the maximum stenosis severity across all coronary arteries, derived from the predicted mesh-based lumen and plaque volumes.
• Failure cases due to centerline extraction errors (e.g., occlusions or abnormal ostia) are handled via exclusion criteria, with potential for future interactive refinement using manual seed points.

Experimental results

Research questions

• RQ1 Can direct mesh generation from CCTA data using a centerline prior achieve accurate and topologically valid segmentation of coronary lumen and plaque?
• RQ2 How does the proposed mesh-based method compare to traditional voxel-wise segmentation in terms of plaque volume quantification accuracy?
• RQ3 To what extent can the method automate CAD-RADS categorization with high agreement against expert readings?
• RQ4 How robust is the method to variations in scanner type, reconstruction parameters, and patient anatomy across multi-center data?
• RQ5 What are the main failure modes of the method, particularly in cases of total occlusion or blooming artifacts?

Key findings

• The method achieved an intraclass correlation coefficient (ICC) of 0.98 for calcified plaque volume, 0.79 for non-calcified plaque, and 0.85 for total plaque volume, indicating high reproducibility.
• Patient-level CAD-RADS categorization reached a linearly weighted kappa (κ) of 0.75 on a representative hold-out test set of 300 scans, indicating substantial agreement with expert readings.
• On an external test set of 658 scans from a different hospital and scanner, the method achieved a κ of 0.71, demonstrating robustness across scanner and institutional differences.
• Only 3.5% of scans were misclassified by more than one CAD-RADS category compared to the reference standard, indicating strong reliability.
• The method failed in only 2 out of 2407 scans due to unrecognized abnormal ostia, suggesting high reliability in centerline extraction.
• A 1D Vision Transformer performed worse (κ = 0.63) than the 1D ResNet, suggesting that local, rather than global, dependencies are more relevant for stenosis prediction.