Automatic Analysis And Visual Enhancement In Cardiac Medical Images

Worldwide,cardiovascular disease has always been the biggest threat to public health.In China,with a population of 1.4 billion,the mortality rate of cardiovascular disease remains the first among all diseases.In the diagnosis and treatment of cardiovascular diseases,cardiovascular medical images,as the analysis basis,play an important role in the diagnosis and treatment process.The artificial intelligence algorithms based on computer vision have the potential to alleviate the shortage of medical resources and improve the diagnosis and treatment efficiency of cardiologists.From this point of view,this thesis mainly studies the artificial intelligence algorithms related to cardiac coronary angiography and MRI images in cardiac medical images.Specifically,the research on the automatic analysis direction of artificial intelligence is mainly carried out on the cardiac coronary angiography image,including the multi-category segmentation of coronary artery,the automatic extraction of cardiac coronary angiography tree,the automatic calculation of coronary stenosis rate and the automatic prediction of the direction of undeveloped vessels with total occlusion of cardiac coronary arteries;Secondly,it mainly carries out fast reconstruction on NMR images for improving image quality.Automatic artificial intelligence analysis of coronary angiography can provide rapid analysis results and important reference for cardiologists,so as to improve the diagnosis and treatment efficiency of cardiologists significantly.On the other hand,the fast reconstruction of cardiac MRI images effectively improve the quality of under sampled MRI data,which improves the MRI generation speed.High quality images will provide more effective information about pathology,with the aim of increase the diagnostic accuracy of cardiologists.These two studies are aimed at providing more effective medical image information for clinical cardiologists,so as to improve their diagnosis and treatment efficiency.Combined with clinical practice,the research work of this thesis designs the adaptability of relevant artificial intelligence algorithms,with the intention of ensuring their effectiveness.The work in this thesis are fourfold as following:(1)Establishment of large-scale coronary angiography dataset.A dataset based on coronary angiography data is collected,sorted and labeled,which makes up for the shortcomings of the existing public coronary angiography dataset,such as insufficient data volume,few coverage categories and no support for computer application,and lays a foundation for the research of coronary artery segmentation algorithms and lesion detection algorithms.The coronary artery recognition dataset,which is applied to coronary artery segmentation,covers 20 types of vessels in clinical diagnosis,and is capable of locating different angiography vessels accurately;The coronary artery lesion morphology dataset were concentrated on lesion detection.Five different forms of common vascular lesions were marked and located in detail,including stenosis,calcification,dissection,thrombosis and total obstruction.The establishment of the dataset combines the knowledge of many fields such as coronary clinical medicine and computer vision.The coronary artery recognition dataset is the only multi-category dataset of coronary vessels in the world.(2)Multi-category segmentation of coronary artery based on coronary angiography.This thesis includes two parts of work on the problem of multicategory segmentation of coronary artery.The first part proposes a vessel segmentation training process using pseudo labeled data.This training process is able to improve the final results for any segmentation network.In addition,this training process can ensure that the model can produce good results even if the fine-labeled data is insufficient.The second part designs a neural network framework for multi-category segmentation of coronary angiography.Specifically,based on the large-scale pixel level fine labeled samples established in the previous step,we use a conditional generative adversarial network to understand the semantics of vessel segments,learn the characteristics and semantics of different vessel segments in coronary angiography,and finally complete the intelligent recognition of coronary artery segments.(3)The extraction of coronary artery tree and its application in angiography.In coronary angiography,the coronary artery is a tree structure in terms of topology.Based on the segmentation image results generated by the vessel segmentation model trained in the previous study,this thesis uses the geometric method to automatically calculate the position of the root node in the topological tree structure.On this basis,using the method of computer graphics,through the algorithm of vessel thinning and centerline extraction,a vessel tree with more abundant information is constructed,including the position of vessel centerline,the direction of each point of vessel centerline,the vessel diameter of each point,the branch segment of the vessel,etc.After constructing the tree structure of this topology,higher-level coronary artery analysis will be completed on this data structure in combination with the previous vessel segmentation results,such as automatic calculation of the stenosis rate of vessel stenosis and automatic prediction of the direction of non angiographic vessels blocked in the angiography.(4)Fast reconstruction algorithm based on cardiac MRI.At present,in clinical cardiac MRI scanning,due to the movement of the heart and chest in the natural state,patients need to hold their breath for many times during the acquisition of MRI images.Considering the rest time of patients,the conventional cardiac MRI scanning time will be relatively long,so various motion artifacts will be introduced in the imaging process.In order to solve this problem,this thesis combines the characteristics of cardiac MRI images,takes multiple cardiac MRI data slices as the input,accelerates the MRI technology based on the deep convolution neural network in deep learning,and restores the low-quality MRI images with low sampling rate to high-quality MRI images.