Coronary Artery Plaque Intelligent Segmentation And Recognition Based On Optical Coherence Tomography Images

With the increasing aged tendency of population,cardiovascular diseases have become the first disease in the incidence and mortality of all diseases in China.Currently,imaging diagnostic techniques for cardiovascular diseases include CT,intravascular ultrasound(IVUS),and optical coherence tomography(OCT).Because OCT images far exceed the resolution of CT and IVUS images(the resolution OCT is about 10 ?m),they are widely used in judging coronary atherosclerotic plaques.However,the current diagnosis of plaque using OCT images still requires manual segmentation by a physician.Therefore,automatic segmentation and recognition of plaques have great clinical significance for the diagnosis of coronary atherosclerosis.In this paper,combining the characteristics of OCT images,artificial intelligence method is introduced.After the completion of plaque segmentation,the characteristics of plaque are expanded,and the intelligent identification of fibrosis,calcification,and plaque is completed.The main research contents of this article are as follows:Firstly,by combining K-means clustering algorithm with GraphCuts algorithm,the proposed method can realize the segmentation of coronary atherosclerosis multiple plaque in this paper.For the traditional method of segmenting,only one kind of plaque or dividing multiple patches is the not high accuracy.The method can optimize plaque area's edge and realize multi-area segmentation.It can achieve segmentation of coronary OCT images--fibrotic plaques,calcified plaques,and lipid plaques,accurately.By calculating the Jaccard coefficient,it is found that the boundary feature information of the plaque is well preserved,and the segmentation accuracy reaches 82.5%.Secondly,geometric features are added and it could accomplish multi-pattem regional feature extraction.Based on non-geometric features(such as FV,PCA,HOG and LBP),we combine them with two geometric features(BF and TF)to characterize the plaque.The experiments show that the geometric feature improved the accuracy of plaque recognition to 96.8%.Thirdly,a patch intelligent identification method based on improved SVM classifier is proposed.For automated classification of the plaque,a hard example mining(HEM)strategy is introduced to train SVM and improve the effectiveness of training data.To assess our method,we investigated how individual feature contributes to the overall classification result,especially which feature most affects the targeted atherosclerotic plaque classification.Data-sets from 20 OCT pullbacks are used to train and test our algorithm.The overall classification accuracy reaches 96.8%,and calcific,lipid-rich and fibrous plaque are 94%,97.2%,99.2% respectively.The work of this paper can greatly reduce the time spent by doctors in segmenting and identifying plaques,reduce the subjectivity and diversity among different doctors,and assist clinicians in the diagnosis and treatment of coronary heart disease.