Morphological Analysis Of Echocardiography Based On Deep Learning

In today’s society,heavy work pressure,coupled with unhealthy lifestyle such as irregular diet and lack of exercise,lead to frequent cardiovascular diseases.As an important method for doctors to diagnose heart diseases,cardiac ultrasound can directly display the patient’s cardiac anatomy and symptoms.Based on this research background,how to use a computer to quickly and accurately process cardiac ultrasound has become a major research focus.In recent years,the breakthroughs and developments of deep learning algorithms have greatly promoted changes in various research fields,such as driverless cars,emotion recognition and face recognition.Applying deep learning algorithms to cardiac ultrasound images can assist doctors in diagnosis,giving treatment plans,and calculating medical parameters,which effectively reduce the workload of doctors and is of great significance to computer-aided diagnosis systems.As far as we know,there are few studies on the application of deep learning to echocardiography,and there is currently no research using RetinaNet for left ventricular localization or segmentation networks such as LedNet for left atrium segmentation.Based on this background,this paper analyzes the shape of cardiac ultrasound based on deep learning algorithms to help doctors better diagnose heart disease and give follow-up treatment plans.This paper mainly encountered two challenges during the experiment.One is that there are few open data sets with complete labels,and the amount of data is very small.The second is that most of the cardiac ultrasound images are grayscale images with few features such as grayscale and texture,which cannot be effectively analyzed using traditional method.In order to solve the above problems,this paper has made four main contributions:(1)Three data sets are created.Medical images contain a large amount of patient private information,and there are fewer open data sets with complete labels.Therefore,this paper has created corresponding data sets for different research tasks.The first is to create a cardiac ultrasound view classification data set,which contains 4837 images in six views;the second is a left ventricular localization data set,which contains 3653 images of A2C,A3C,and A4C containing the complete left ventricle;the third is the left atrium inner and outer membrane segmentation data set,which contains 1480 images in three views of A2C,A3C,and A4C.All images in the data set are collected from the hospital on-site,which is more meaningful for clinical treatment.(2)Classification of cardiac ultrasound views based on deep learning.On the cardiac ultrasound view classification data set,four classification networks of AlexNet,ResNet,VGG,and DenseNet are applied to train the six cardiac ultrasound views,and then the classification performance was compared and evaluated.Experimental results show,DenseNet achieved the best results,with an accuracy rate of 99.90%(3)Left ventricular localization based on deep learning.Based on the created left ventricular localization data set,the left ventricular localization task is performed through RetinaNet and the final left ventricular localization result map is obtained.Using mIoU as the evaluation index of the left ventricular localization effect to test the A2C,A3C,and A4C views,the results were 85.83%,79.42%,and 83.87%,respectively.(4)Segmentation of the inner and outer membranes of the left atrium based on deep learning.Based on the created left atrium inner and outer membrane data set,this paper uses FCN,DeepLab,LedNet,PSPNet,DANet,CGNet,DenseASPP,ENet,ESPNet,these 9 segmentation networks to segment the left atrium inner and outer membrane,and fully compare the differences for different heart views,the effect of segmentation network and experiment time.From the perspective of the evaluation index,for the Hausdorff distance,ENet in A3C has the best effect with a result of 3.8939;for ASD,ENet in A4C has the best effect with a result of 0.5491;for Dice coefficient,LedNet in A3C has the best effect with a result of 0.8339;for Pixel Accuracy,LedNet in A3C works best,the result is 0.9934.