Automated Assessment of Coronary Artery Calcification in IVOCT Based on Deep Learning

Abstract Coronary artery calcification (CAC) is a marker of atherosclerosis, capable of reflecting the severity of coronary artery lesions. However, there is currently a lack of an end-to-end evaluation method that can achieve rapid, accurate, and automated CAC assessment. In this study, a deep learning-based classification model (LFL-Net) was developed to directly extract calcified plaques, along with their angles and thickness, from intravascular optical coherence tomography (IVOCT) images to obtain CAC scores. The internal dataset comprised IVOCT images from 367 patients across two centers, utilized for model training and internal testing. Additionally, IVOCT images from 10 patients from another independent center were used for external testing to validate the model's generalization ability. In external testing, the LFL-Net model achieved an accuracy and recall of 0.7048 and 0.7202, respectively, in the calcified plaque classification task; 0.7893 and 0.8013 in the calcification angle classification task; and 0.7683 and 0.6724 in the calcification thickness classification task. Moreover, in the quantitative analysis of calcification scores, the model demonstrated an accuracy of 0.9192, sensitivity of 0.7990, and specificity of 0.9463. The results indicate that the LFL-Net model performs exceptionally well in handling complex IVOCT image data, offering a stable and accurate technical tool for CAC assessment.