Automated Artificial Intelligence Mapping of Coronary Plaque Calcification: A Comparison with Manual Intravascular Image Analysis

Abstract

Background/Aims: Intravascular imaging during percutaneous coronary intervention (PCI) improves clinical outcomes; however, is dependent on accurate and rapid interpretation of the images generated. This study aimed to compare coronary artery calcification assessment using a novel automated artificial intelligence (AI) software algorithm with manual optical coherence tomography (OCT) image analysis. Methods and Results: A deep neural network based on a UNet-like architecture was developed and trained to identify calcified atherosclerotic plaque from an independent dataset of expert-annotated clinical intravascular OCT pullbacks. The AI network was subsequently validated on previously unseen clinical OCT pullbacks that were manually annotated for plaque calcium and used to quantify clinically relevant calcified plaque characteristics. Correlation and agreement between the expert-annotated images and the model predictions were evaluated. In total, 8259 cross-sectional images comprised the training and internal validation dataset. Pixel-based classification by the AI model performed best to identify calcified plaque (AUC = 0.96), with an overall diagnostic accuracy of 73.3%. During independent external validation, the model correctly identified 934 of the 1248 calcified plaques, corresponding to a diagnostic accuracy of 74.8%. The AI model performed well in assessing the calculated OCT-calcium score (ρ = 0.84; 95% confidence interval [CI], 0.81-0.87, p ≤ 0.001). Conclusions: Implementation of an automated AI software algorithm provides a rapid and efficient method to comprehensively map coronary calcium in intravascular OCT images. With further training and refinement, it is anticipated that the AI machine learning software will continue to improve, enabling new robust tools for clinical OCT calcium detection to better guide PCI procedures.

Keywords: artificial intelligence; calcium; optical coherence tomography.

Figure 1

Schematic of the UNet-like architecture. The contracting (left) path was composed of four convolutional blocks, each comprised of two 3 × 3 2D convolutional layers and an activation function (rectified linear unit/ReLU), followed by a 2D maxpooling layer and a dropout layer connecting adjacent convolutional blocks. The expanding (right) path was composed of a 2D upsampling layer, which matched the corresponding 2D maxpooling layer in the contracting path, followed by a 2D convolutional layer and an activation function (ReLU). In combination with both vertical and horizontal feature bridges, a final segmentation map was produced.

Figure 2

Benefits of artificial intelligence guidance for intravascular imaging optimization during PCI *** including a summary of main results of the study and examples of calcium annotation by the manual reader and the AI software algorithm.

Figure 3

Scatterplots depicting the correlation between the experienced OCT human readers and AI model. (A) Minimum calcium thickness, (B) arc of calcium, and (C) OCT-calcium score. Manual human interpretation is represented on the X-axis, and the AI-model interpretation is represented on the Y-axis.

Figure 4

Manual segmentation vs. automated artificial intelligence analysis of calcified coronary plaques. Representative examples showing (A) calcium score 1 and (B) calcium score 3. The OCT-calcium score is calculated as the total sum of the following characteristics: calcium thickness (1 point for >0.5 mm), maximum arc (2 points for >180°), and length (1 point for >5 mm), with OCT calcium scores of >3 being associated with higher potential for clinical stent under-expansion. The first image of (A,B) depicts the unedited, original clinical OCT frame with the OCT catheter in the center of the lumen (dark black space). The second image depicts the manually annotated clinical OCT frame with calcified plaque highlighted (green) and minimum thickness and arc of calcium reported. The third image depicts the AI algorithm-annotated clinical OCT frame with calcified plaque highlighted (green/blue) and minimum thickness and arc of calcium reported. Ca = calcified plaque.