Coronary Angiography Segmentation Based On Deep Learning

In recent years,cardiovascular disease has become one of the most lethal diseases in human beings.It has the characteristics of high morbidity,high disability and high mortality.In our country,the morbidity and mortality of cardiovascular diseases are also on the rise.Therefore,we propose that by combining computer technology with medical treatment,we can extract precise vascular structures from vascular images with more efficient image processing methods,and reduce the manual interaction in diagnosis,reduce the dependence on doctor,and improve the efficiency of disease diagnosis.It also provides assistant strategies for subsequent medical diagnosis based on coronary angiography,such as calcification detection,stenosis detection and so on.At present,the performance of traditional image segmentation methods in medical images are not good enough,and also the research of coronary angiography based on deep learning is just a beginning.In this context,we propose a deep neural network method for coronal angiography image segmentation.The main work of this thesis are as the following parts:Firstly we use traditional algorithms to segment coronary angiographic images,and successfully implement some of the mainstream classical segmentation algorithms,including:Otsu algorithm,GrowCut algorithm.Based on the results,we analyze the performance and shortcomings of these algorithms in this task;Secondly we established the database of cardiac coronary angiography from scratch,including data sources,data desensitization,design and realization of annotation tools,the data were classified according to different pathological changes in medicine,and then subdivided into 7 positions including left and right coronary artery,more than 11,000 rough labeled data and 9,551 fine labeled data were obtained;Thirdly after studying the relevant knowledge and core ideas of convolutional neural network and deep learning,we design an algorithm used for coronal angiography image segmentation based on deep learning,including image preprocessing,network selection and specific structure,training and testing methods,we have successfully implemented the experiments of classical network FCN and PSPNet in the task of segmentation.In the experiment,we used PSPNet and achieved excellent results in every view of angiography.And used three evaluation indicators:accuracy,recall and F1 score to analysis and discuss the results.The results also proved the advantages and effectiveness of PSPNet network based on deep learning in image segmentation tasks of coronary angiography.