Applying Machine Learning Methods For Coronary Artery Plaque Detection On Computed Tomography Angiography(CTA)

Cardiovascular disease is one of the main causes of human death,and coronary artery disease is the most common cardiovascular disease.The detection of coronary atherosclerotic plaque is of great significance for the early intervention of coronary artery disease.As a noninvasive,low-risk and low-cost imaging modality,coronary computed tomography angiography(CTA)is considered to be an alternative to invasive coronary angiography.It is used for the early screening and prevention of coronary artery disease.However,calcification artifacts,motion artifacts caused by beating heart,etc.will reduce the CTA image quality,resulting in large inter-observer errors in visual assessment.In automatic or semi-automatic methods,threshold-based segmentation method cannot accurately segment coronary plaque,the method that rely on vessel segmentation is difficult to generate a complete coronary artery tree,and the voxel classification method is time-consuming and laborious to obtain the label.In order to solve these existing problems,this paper uses two different machine learning methods to complete automatic detection of coronary plaques based on pre-acquiring the coronary centerlines.The research work is mainly as follows:(1)Detection of multi-class atherosclerotic plaque based on support vector machine(SVM).In view of the difficulty in distinguishing different types of plaques with existing methods,this study uses machine learning methods to identify coronary plaques.Firstly,we retrieved the transverse cross sections along centerlines.Secondly,we used the level set method to roughly segment the blood vessel,obtained the region of interest.Thirdly,we extracted a random radius symmetry feature vector to characterize the plaque image,and used random strategy to augment the training data.Finally,we trained a support vector machine classifier to recognize plaque images.Our method outperforms other classifiers and different feature extraction schemes,with the average precision of 92.6% and average recall of 94.3%.The experimental results show that the method has high accuracy in the recognition of patch images.Compared with the traditional method,this method does not rely on the fine segmentation of the lumen,does not need to mark the complete plaque,and has higher clinical application potential.(2)Detection of coronary plaque based on ladder convolutional neural network.In order to solve the problem that it is difficult to train a robust deep model with a small amount of labeled samples,this paper studies a deep model based on semi-supervised learning and uses a large amount of unlabeled data to assist in the detection of coronary plaque.Firstly,we extracted three-dimensional image patches as training samples along the coronary centerlines.Secondly,labeled and unlabeled samples are sent to the convolutional neural network for training at the same time.The unsupervised learning network learns the reconstruction representation,and the supervised learning network filters the reconstructed information learned,and adds jump connections to make high-level learning more abstract characterization.Finally,the network is trained using a loss function that combines supervised and unsupervised.The accuracy and recall of the network on the test set are 71.88% and 88.46%,respectively,which are 19.5% and 3.84% higher than that of the supervision method.The network uses unlabeled data to improve the accuracy of plaque detection and has better performance on tasks with a small amount of labeled sample data.