Automated comprehensive evaluation of coronary artery plaque in IVOCT using deep learning

Abstract

The process of manually characterizing and quantifying coronary artery plaque tissue in intravascular optical coherence tomography (IVOCT) images is both time-consuming and subjective. We have developed a deep learning-based semantic segmentation model (EDA-UNet) designed specifically for characterizing and quantifying coronary artery plaque tissue in IVOCT images. IVOCT images from two centers were utilized as the internal dataset for model training and internal testing. Images from another independent center employing IVOCT were used for external testing. The Dice coefficients for fibrous plaque, calcified plaque, and lipid plaque in external tests were 0.8282, 0.7408, and 0.7052 respectively. The model demonstrated strong correlation and consistency with the ground truth in the quantitative analysis of calcification scores and the identification of thin-cap fibroatheroma (TCFA). The median duration for each callback analysis was 18 s. EDA-UNet model serves as an efficient and accurate technological tool for plaque characterization and quantification.

Keywords: Artificial intelligence; Cardiovascular medicine.

Graphical abstract

Figure 1

Challenging and failure cases of automatic IVOCT image segmentation (A–H) Images of the model successfully segmenting challenging cases. (A and B) represent before and after segmentation of residual blood images caused by incomplete rinsing. (C and D) represent coronary artery dissection images before and after segmentation. (E and F) represent branch vessel images before and after segmentation. (G and H) represent extremely eccentric catheter images before and after segmentation. (I–L) Images of cases in which the model fails to perform accurate segmentation. (I and J) represent microvessel images before and after segmentation. (K and L) represent images of branch vessel that do not intersect the lumen before and after segmentation. The microvessel has been erroneously segmented as calcified plaque (white arrow). The branch vessel that does not intersect the lumen has been incorrectly segmented into calcified plaque (red arrow).

Figure 2

Ablation experiments for the proposed model MIoU, mean intersection over union.

Figure 3

Correlation and consistency between the maximum lipid angle and minimum fibrous cap thickness between the proposed model and ground truth Correlation of maximum lipid angle (A), Bland-Altman analysis of maximum lipid angle (C), correlation of minimum fibrous cap thickness (B), and Bland-Altman analysis of minimum fibrous cap thickness (D).

Figure 4

Confusion matrix for comparing the proposed model-identified TCFA with the ground truth (n = 90) TCFA, thin-cap fibroatheroma.

Figure 5

Correlation and consistency of calcification angle, calcification thickness, calcification depth, and calcification area between the proposed model and ground truth Correlation of calcification angle (A), Bland-Altman analysis of calcification angle thickness (E), correlation of calcification thickness (B), Bland-Altman analysis of calcification thickness (F), correlation of calcification depth (C), Bland-Altman analysis of calcification depth (G), and correlation of calcification area (D), and Bland-Altman analysis of calcification area (H).

Figure 6

Confusion matrix comparing the proposed model-predicted calcification scores with the ground truth (n = 87) The grading of the calcification score (0, 1, 2, 3, 4) in the graph is usually used to indicate different degrees of calcification, from no calcification to severe calcification, with larger numbers indicating more severe calcification. And the calcium score was defined as follows: 1 point for a maximum calcium thickness >0.5 mm, 2 points for a maximum calcium angle >180°, and 1 point for a calcium length >5 mm.

Figure 7

Flowchart of the study design for the development and evaluation of the EDA-UNet model

Figure 8

Performance comparison of the proposed model with state-of-the-art network models The calcified plaques in the yellow boxes of the (B), (E), and (F) rows exhibit a high degree of visual similarity to the background, complicating their differentiation and extraction. In the yellow boxes of the (A) and (C) rows, the outer boundaries of the lipid plaques are difficult to determine, with punctate fibrous plaques and calcified plaques observed inside the lipid plaques, further complicating plaque identification and segmentation. In the yellow box in the (D) row, punctate lipid plaques are scattered within the fibrous plaques, challenging the accurate extraction of intravascular tissue.