Research On Deep Learning Algorithms Of Artery Segmentation In Medical Images

Recently,with the development of machine learning and deep learning techniques,the subject of computer vision achieves rapid development.Medical image analysis,as one of the traditional branch of computer vision,is essentially meaningful for medical diagnosis.This paper introduces the application of deep learning in different medical images.This different applications can be used to solve medical problems in real clinical scenario.Since the category of medical image analysis is broad which generally includes classification,segmentation,detection and registration.In this paper,we mainly study the deep learning in medical image segmentation to assist doctor completing medical image analysis and diagnosis in a short time.In this paper,two scenarios in which the coronary artery segmentation in X-ray angiographic image and middle cerebral artery segmentation in B-mode ultrasound image are analysed.By segmenting the coronary arteries in medical images and examining them,important information about abnormal narrowing and plaque,which are main causes to these diseases,can be found.Manual segmentation of coronary arteries are time consuming and dependent on the observer,which makes the need of automatic segmentation techniques apparent.However,the segmentation of coronary arteries in X-ray angiographic images is a challenging task due to the existence of artifacts and low image quality.By segmenting the fetal middle cerebral artery,the standard slice of B-mode ultrasound image can be obtained and this further provides the suitable location for measuring blood parameters in fetal middle cerebral artery.Traditional ways of obtaining standard slice of B-mode image highly depend on the expert doctor.Moreover,the color Doppler is needed for detecting the location of fetal middle cerebral artery and then the measurement of blood parameters can be completed.This produce is nontrivial in which many manually operations exist.In this paper,two deep learning algorithms are proposed.These two algorithms are used for segmenting coronary artery in X-ray angiographic image and fetal middle cerebral artery in B-mode ultrasound image.These two algorithms make it possible for automatically detecting the level of vascular occlusion and blood parameter measurement respectively.The main contributions in this paper are:1.In this paper,we design and realize a simple yet effective domain discrepancy guided data augmentation algorithm.By synthesizing virtual images to reduce the discrepancy between source domain(eye fundus image)and target domain(X-ray angiograms).Differ from other data augmentation methods,the method proposed here can not only synthesize virtual images but also the corresponding labels.We can start unsupervised learning by using these synthesized images and labels.2.In this paper,we design and realize a method for automatically detecting the standard slice of B-mode ultrasound image and automatically locate the suitable position for measuring blood parameters in fetal middle cerebral artery.By using this method,many unnecessary manually operations can be avoided to further improve the efficiency of diagnosis.At the same time,the method can be used to assist junior doctor to learn the procedure of diagnosis.3.Extensive experiments show that the methods proposed in this article can effectively reduce the discrepancy between source domain and target domain to further improve the performance of domain adaptation methods.Moreover,experimental results show that our method is quite effective for automatically obtain the standard B-mode slice and providing effective location for blood parameters measuring.