Automatic Coronary Artery Segmentation And Plaque Classification Of Coronay CTA Based On Deep Learning And Radiomics Method

Purpose:To explore the efficacy of AI(artificial intelligence)in coronary artery vascular segmentation,plaque identification and classification,and luminal stenosis grading diagnosis of coronary CTA(computed tomography angiography).Methods: A total of 127,763 coronary artery CTA images of 505 patients were collected from5 different data centers from January 2018 to December 2018,including 261 males and 244 females,with an average age of 52.5±17.6 years.All included patients underwent coronary artery CTA examination because of clinical diagnosis of suspected coronary atherosclerotic disease.All images were processed by the big data center of PLA general hospital.In the process of analysis,the patients’ s ID(identity)number,name,gender,age,birth date and other general information were hidden from the images.All images were standardized before labeling.After image preprocessing,two independent experienced radiologists with over 5years of experience in cardiovascular imaging manually delineated the coronary artery and labeled the plaques on CTA images using anin-housesoftware developed based on MATLAB2017 a.The labeling-modifying pattern was adopted,in which one doctor label the images and the other doctor modified the results.To validate the reproducibility of the plaque labeling,an another radiologist with 20-year-experience randomly selected 50 patients and then conducted the plaque labeling.In this study,coronary atherosclerotic plaques were divided into three categories: calcified plaque,non-calcified plaque and mixed plaque.The 2.5D multi-plane and Hessian maps were used as the adjustment of Mask-RCNN(region convolutional neural network).The adjusted Mask-RCNN was used to segment the coronary vessel and detect candidate plaques.Postprocessing was applied to connect the broken end of the vessel,extract the three-dimensional surface image and the center line of the coronary artery.The origin of coronary artery was detected by centroid tracking and multi-plane reconstruction algorithm.The cross sections of candidate plaques were re-sliced using the multi-plane resecting algorithm.Radiomics and morphological features were extracted from each image.Traditional classifiers(LDA,SVM,GBDT and k NN)and deep learning-based classifiers(VGG16,Res Net50,Xception and Dense Net)were used to classify plaques,and the classification performance of classifiers was tested on different datasets.For the measurement of vascular lumen stenosis,the degree of occlusion is calculated and classified according to CAD-RADS(coronary artery disease reporting and data system).The blocking degree is the cross-sectional area of the plaque divided by the corresponding lumen cross-sectional area(%),and the stenosis degree of the plaque is defined by the minimum blocking degree.Statistically,intra-reader agreement was evaluated using weighted kappa statistics;recall,precision,accuracy and dice were used to evaluate the ability of adjusted Mask-RCNN to segment coronary vessels of both training set and testing set,while recall,precision,accuracy were used to detect plaque candidates of both data sets;accuracy,sensitivity,specificity,positive predictive value and negative predictive value were used to evaluate the ability of the adjusted Mask-RCNN to detect the origin of coronary arteries of the two data sets;the ICC(interclass correlation coefficient)and F test were used to reduce the radiomics and morphological features;true positive,true negative,false positive,false negative,accuracy,specificity,sensitivity,positive predictive value and negative predictive value and ROC(receiver operatorcharacteristic)curve were used to evaluate the classification performance of the two models;accuracy,specificity,sensitivity,positive predictive value and negative predictive value were used to evaluate the results of stenosis degree.All statistical results were evaluated using manual annotation as the gold standard.Results: The weighted Kappa value of the two groups of doctors for the consistency of different plaque markers was 0.848(95%CI[confidence interval]:0.744-0.952);the results of recall,precision,accuracy and dice of adjusted Mask-RCNN for coronary vascular tree segmentation in the training set and testing set were 95.3%,85.7%,82.2%,83.5% and 96.9%,87.0%,80.8%,83.0%respectively;the results of recall,precision and accuracy of plaque candidates detection in the both data sets were 100%,84.6%,79.3% and 100%,83.9%,77.9%respectively;accuracy,sensitivity,specificity,positive predictive value and negative predictive value of the coronary origin detection in both data sets were96.7%,94.2%,98.9%,98.4%,87.6% and 94.5%,92.6%,97.3%,96.9%,90.6% respectively.20 radiomics features and 6 morphological features were selected by ICC and F tests;Dense Net showed the best performance among all classifiers the area under curve was 0.9137,and true positive,false positive,true negative,false negative,accuracy,specificity,sensitivity,positive predictive value and negative predictive value were 983,90,550,88,89.6%,91.8%,85.9%,91.6% and86.21%,respectively.The accuracy,specificity,sensitivity,positive predictive value and negative predictive value of the stenosis calculation were 85.6%,83.1%,90.5%,82.2% and86.7% respectively.Conclusions: The frame work of using the adjusted mask-RCNN for coronary artery extraction,and radiomics and morphological features with a variety of classifiers for plaque detection and classification can greatly reduce image reading time for radiologists and provide good coronary plaque detection and classification accuracy,showing potential in future clinical use.