| Coronary atherosclerotic heart disease,referred to as Coronary heart disease,is the disease with the highest mortality in the world.Coronary heart disease is a heart disease caused by the deposition of lipids,cholesterol,calcium and other substances in the coronary arteries to form atherosclerotic plaques that cause coronary stenosis or obstruction,resulting in myocardial ischemia or necrosis.The rupture of atherosclerotic plaques can lead to thrombosis,triggering a series of acute cardiovascular events such as myocardial infarction.The purpose of coronary heart disease treatment is to control the formation of plaque and prevent plaque rupture,so early identification and classification of plaque is the premise of interventional treatment of coronary heart disease.Imaging techniques commonly used in the diagnosis of coronary plaque can be divided into invasive and non-invasive methods.Intravascular ultrasound is the gold standard for the diagnosis of atherosclerotic plaque.Computed Tomography Angiography(CTA)is a widely used clinical detection method and has been applied to the clinical diagnosis of coronary artery disease.Compared with intravascular ultrasound,CTA has the advantages of noninvasiveness,rapidity and accuracy.For the task of establishing coronary CTA plaque database,this thesis introduces the centerline extraction and cross-section generation algorithm,and designs a plaque labeling software based on cross-section algorithm for plaque labeling.The data source of coronary CTA plaque database is the General Hospital of the People’s Liberation Army,and the data set is coronary CTA data.In this thesis,the Mimics software is used to manually segment the coronary arteries,and the segmented coronary arteries are used to extract the centerline of the blood vessels.To reduce the difficulty of plaque labeling,the original data is transformed into a multi-planar reconstruction(MPR)image by generating the cross-section of the blood vessel.In this thesis,the vessel centerline is extracted from the segmented vessel data by thinning algorithm,and calcification artifact correction is performed on the vessels with calcified plaques.The cross-section generation algorithm is implemented by calculating the mapping relationship between the cross-section of blood vessels and the coordinates of the original CTA space matrix.The proposed cross-section generation algorithm adds rotation around the z-axis on the basis of the original algorithm,which improves the problem of rotation dislocation between cross sections.The cross-section generation algorithm is implemented on MATLAB and Mevislab platform respectively.Two labeling software are designed and implemented in this thesis,they are cross-section category labeling software and plaque labeling software,which respectively realize cross-section category labeling and pixel-level plaque labeling,providing software support for plaque labeling.For the task of automatic detection of multi-category plaques,this thesis uses the radionomics method to achieve multi-class plaque detection.The basic process of radiomics method is as follows:(1)image collection,(2)Region of interest(ROI)acquisition,(3)feature extraction,(4)feature selection,(5)classification and prediction.Since the label portion of the self-built CTA plaque database has not been completed,the Rotterdam coronary artery evaluate dataset is used in this task.The data in Rotterdam include coronary CTA data,centerline coordinates and plaque label based on centerline coordinates.In this thesis,the cross-section generation algorithm is used to convert the Rotterdam data into labeled cross-section data(normal cross-section,non-calcified plaque cross-section,calcified plaque cross-section,mixed plaque cross-section).Due to the uneven number of various types of plaques,this thesis has tried three methods to deal with the imbalance of sample categories,but the results are not good.According to the rule that the radius of vessel lumens is concentrated around 4 pixels,ROI is set as a circular region with a radius of 4.Feature extraction is implemented by Pyradiomics module,which extracted a total of 774 radionomic features.In this thesis,de-collinearity feature selection,Relief-F and random forest feature selection methods are used,and the random forest feature selection method has the best effect.In this thesis,a neural network model composed of two-layer full connection layer,dropout layer and softmax layer is proposed,it and random forest are used as two classifiers to classify the characteristics of radiomics.The performance of the model is evaluated by the macro average F1 score.The optimal model is the algorithm model combined with the neural network after random forest feature selection,and its accuracy is 0.93,macro average F1 score is 0.7.The work of this thesis is summarized as follows:(1)center line extraction based on refinement algorithm is introduced,and the algorithm of cross-section generation is improved;(2)based on the cross-section generation algorithm,two kinds of plaque labeling software are designed and implemented to provide software technical support for plaque labeling in the coronary CTA plaque database;(3)based on the plaque category label data,the multi-category plaque detection task is realized by using the radiomics method. |
