Research On Dynamic Segmentation Of Left Ventricle In Ultrasound Video And Method Of Automatic Calculation Of Ejection Fraction

Accurate evaluation of cardiac function is essential for the diagnosis of cardiovascular disease and the screening of cardiotoxicity.Echocardiography can visually display the structure of the heart and large blood vessels and blood flow,and can quantitatively measure the parameters of cardiac function evaluation,at the same time,it has the advantages of fast image production,low price and no ionizing radiation.Therefore,echocardiography is an important way to diagnose cardiovascular disease,and the two-dimensional echocardiography is the most commonly used in echocardiography.The calculation of the left ventricular ejection fraction is an important step in the diagnosis of cardiovascular diseases using two-dimensional echocardiography.In the traditional method,this process requires the doctor to manually outline the contour of the left ventricular area in the relevant section of the two-dimensional echocardiography.Therefore,the difference in the professional level of doctors will lead to the problems of low measurement efficiency and accuracy,and poor measurement repeatability.At the same time,the previous deep learning model for the segmentation of the left ventricle is mostly the study of artificially annotated highdefinition static images of MRI or CT,rather than using actual echocardiographic videos.Therefore,it is currently difficult to segment the left ventricle in cardiac ultrasound video,on the one hand,ultrasound video has lower image quality than other medical images,on the other hand,the complex heart structure makes it difficult to identify the left ventricle.Regarding the above-mentioned problems,this article has launched a research on the dynamic segmentation of the left ventricle in ultrasound video and the method of automatic calculation of ejection fraction.This article firstly proposes a lightweight model of echocardiographic left ventricle segmentation based on the improved DeepLabV3+ network for segmentation of the left ventricle region of the apical fourchamber view in two-dimensional echocardiography.The model uses weakly supervised learning to learn the left ventricular end-diastolic and end-systolic images of the left ventricular region in the training set video to predict the left ventricular region of other frames in the video,the method of weakly supervised learning reduces the workload of manually labeling data sets.At the same time,in order to enable the algorithm to recognize the left ventricular region in the video in real time and enable it to be deployed on the handheld B-ultrasound device,this article replaces the original backbone network in the DeepLabV3+ network with MobileNetV2 with fewer parameters and calculations to reduce the size of the model.The experimental results show that the model proposed in this article outperforms the segmentation results of FCN＿ReNet50 and DeepLabV3＿ResNet50 models with a Dice similarity coefficient of about 0.92 and an IOU of 0.85.At the same time,the model in this article consumes less training time and takes up less memory.Finally,in order to realize the automatic calculation of left ventricular ejection fraction,this article simplifies the calculation process of ejection fraction.First,extract the contour of the left ventricle from the result of the left ventricular region segmentation,then process the contour and calculate the length of the corresponding structure,and finally directly use the single-sided Simpson method to calculate the left ventricular ejection fraction.The results show that,compared with the actual results,the average absolute error of the left ventricular ejection fraction calculated by the method in this article is 14.1%.