A context-aware deep network for automated characterization of coronary artery plaque from coronary CT angiography

Abstract

Abstract Background Identifying various types of plaque, including calcified, non-calcified, and mixed, can furnish crucial insights into a patient's susceptibility to cardiovascular disease and aid in making treatment decisions. Manual quantification of coronary artery plaque on coronary CT angiography (CCTA) poses several challenges, including subjectivity, time consumption, and potential measurement errors. Purpose Our objective was to develop a context-aware deep network (CADN) that could automatically evaluate coronary artery plaque in a sizable, multi-center prospective CCTA cohort. Methods The APOLLO study is a multi-center study involving CCTA scans of 5,000 Asian patients. At the present stage of this study, CCTA analysis was performed on a subset of 474 patients (138 females; mean age of 58 ± 11 years), of whom 76% were Chinese, 17% were Indian/Malay, and 7% were of other ethnicities. Additionally, 53% had hypertension, 20% had diabetes, 70% had hyperlipidemia, and 27% were obese. A total of 2,983 lesions in 1,422 arteries were analyzed, including 473 non-calcified, 1,612 mixed, and 898 calcified plaques. The CADN combines a 3D convolutional neural network with a Transformer and incorporates contextual information along with local features (Figure 1). The network was trained and tested in 379 and 95 patients, respectively. Results The CADN method showed a high level accuracy in detecting plaque, achieving a lesion-level area under the receiver operating characteristic curve of 0.87. Furthermore, the CADN method demonstrated a lesion-level diagnostic accuracy, F1 score, and intra-class correlation coefficient of 0.74, 0.71, and 0.90, respectively, in characterizing plaque into categories of no plaque, non-calcified, mixed, and calcified. When analyzing its performance in each coronary artery, the CADN method was able to accurately characterize the plaque into multiple categories in the LAD, LCX, and RCA with accuracies of 0.70, 0.73, and 0.77, respectively. Conclusion The context-aware deep network is a novel approach that demonstrates good diagnostic performance in detecting and characterizing coronary artery plaque.Figure 1.CADN for coronary plaque.