Siamese Neural Network And The Application In Speckle Tracking Echocardiographic Technique

With the rapid development of deep learning,deep neural networks have shown their talents in various fields,especially those related to image processing.In the field of target tracking,target tracking methods based on deep learning have been occupied by classification networks and siamese neural networks.Siamese neural network can perform single-sample learning and has natural advantages in the field of image block matching.Applying it to target tracking can match the target quickly and accurately,simplify the model while ensuring accuracy,and speed up the tracking.Echocardiography uses the unique physical characteristics of ultrasound to obtain periodic motion information of heart tissue structures such as ventricles,myocardium,endocardium and epicardium,and form heart tissue motion images.It is widely used in the clinical diagnosis and treatment of various heart diseases.Speckle in echocardiogram is not only noise,but also contains a lot of tissue structure information,reflecting the physical characteristics of heart tissue structure.Speckle Tracking Echocardiography(STE)uses the stability and uniqueness of speckles to automatically identify and track their spatial motions,and obtain parameters such as the displacement,velocity,and deformation of the myocardium.Furthermore,the strain and strain rate of the myocardium can be calculated,which has important clinical value in the evaluation of coronary heart disease and other heart diseases.As the number of patients with coronary heart disease in my country has increased significantly,rapid and accurate speckle tracking of echocardiography is of great significance for the prevention and treatment of coronary heart disease and other heart diseases.The most representative target tracking model based on the twin neural network is SiamFC,which applies a fully convolutional twin neural network to target tracking.Subsequent SiamRPN,SiamRPN++,SiamMask,etc.integrated ResNet-50,RPN,multi-feature fusion and other technologies on this basis,and achieved remarkable results.The above model is designed for real life scenes,and has good effects in real life scenes.However,the echocardiogram is very different from the images in real life scenes.One is that the echocardiogram is a grayscale image,while the image in the real life scene is a color image;the other is that there are a lot of speckles in the echocardiogram.It interferes with the real heart contour,and at the same time speckle contains information such as myocardium and blood.This makes it impossible to accurately identify the myocardial tissue in a specific region when the above model is applied to STE,and the recognition results in adjacent frames are very different,so that effective speckle tracking cannot be performed.Therefore,based on the SiamMask model,this paper proposes a new model that can perform speckle tracking on echocardiography—STESiamMask,which uses fixed-scale candidate frames to track speckle in specific areas of myocardial tissue.The STESiamMask model uses the target frame of the myocardial radial scale to generate a training set and train it.At the same time,it optimizes the model in terms of template image,search image,loss function,mask threshold,etc.,and performs well in indicators such as IoU and center position error.This paper also proposes a method to realize STE based on the STESiamMask model.First,preprocess the template frame of echocardiography,cut out the template area with the characteristic points of the myocardium as the center,and then resize the template area to generate the template image.The size of the search area is calculated according to the scaling relationship between the template area and the template image and the size of the search image.The search area is cut out in the tracking frame with the myocardial feature point of the template frame as the center,and the search area is resized to generate the search image.The template image and the search image are processed by the STESiamMask model to obtain the best tracking target frame,and the center point of the target frame is taken as the myocardial feature point for speckle tracking.This paper uses the echocardiographic simulation data set to carry out the application experiment of the STESiamMask model and the speckle tracking method based on the STESiamMask model.According to the experiment,the longitudinal strain of the myocardium is calculated and compared with the standard strain.The results show the validity and effectiveness of the model and method in this paper.Good application performance.In addition,the speckle tracking method based on the STESiamMask model proposed in this article has reached a processing speed of 5～6fps on Ubuntu18.04,GTX 1050Ti,and Intel i3-2120 CPU@3.30GHz,which basically meets the needs of clinical applications.