Automatic Classification And Parameter Measurement Of Echocardiogram Using Deep Learning

Echocardiogram can be used to visually display and quantitatively measure the important information about anatomical structure,function and hemodynamics of the heart,thus reflecting the cardiac structural characteristics in physiological and pathological conditions,and plays a main role in heart disease diagnosis.The accurate measurement of cardiac parameters is an important component of echocardiographic diagnosis.However,the traditional methods require physicians to manually choose the relevant views and trace the region of interest,which is inefficient,low-precision,non-repeatable and cannot fulfill the needs of practical applications.To solve the above problem,the research is carried out on automatic measurement methods of echocardiographic parameters using deep learning in this paper.Left ventricular ejection fraction(LVEF)is one of the important predictors for evaluating the cardiac systolic function,which is of great significance for the clinical diagnosis of heart failure,the observation of drug and surgical effect,and the judgement of the condition and prognosis.Accordingly,it is selected as the representative to study the automatic measurement methods of echocardiographic parameters.The process is divided into the following three steps.Firstly,the automatic classification methods for echocardiographic video views,which are based on convolutional neural networks,are studied in this thesis.The first step is to parse the echocardiographic videos into static images.Then,eight static views are classified automatically using the Inception V3 and ResNet50 models respectively,of which the ResNet50 model with better accuracy is selected as the standard model of classification.To make use of the inter-frame correlation of the video,the ResNet50+LSTM model is constructed,which is used to extract the time series features of two-dimensional image sequence and realize the classification of echocardiographic video views.The test results show that the average classification accuracy of ResNet50+LSTM model is 0.9790 and 5.82% higher than the ResNet50 model,which meet the requirement for clinical utility.Under the premise of accuracy classification,the automatic segmentation methods of left ventricle on apical four-chamber and apical two-chamber views,which are based on fully convolutional neural networks,are proposed.The VGG19 FCN and U-Net networks are used in this paper.At first,the segmentation basic models VGG19FCN/A and U-Net/A are trained by the labeled static images.Secondly,a semi-supervised learning method is proposed,by fine-tuning the basic models' weights parameters using the echocardiographic video that the first frame labeled,the semi-automatic models VGG19FCN/B and U-Net/B are obtained.The left ventricle can be segment automatically using the two basic models VGG19FCN/A and U-Net/A,which meets the primary requirements of clinical application.The two semi-automatic models VGG19FCN/B and U-Net/B require manual interaction.However,the segmentation accuracy is significantly improved and still perform well in the case of unclear endocardium.The comparison results between the two networks shows that the overall accuracy of U-Net is slightly lower than VGG19 FCN,while the number of parameters is about 1/359 of VGG19 FCN and the processing time is about 1/12.7 of VGG19 FCN.Meanwhile,the edge of the segmented left ventricles based on U-Net model is smoother than VGG19 FCN.Finally,the segmented left ventricles are processed and the left ventricular ejection fraction is calculated using the Simpson's method.Compared with the manual tracing methods,the errors of the left ventricular ejection fraction calculated based on VGG19 FCN and U-Net are 6.29% and 6.59% respectively.In summary,starting from the input of echocardiographic videos,as well as the automatic classification of views,automatic segmentation of left ventricles and automatic measurement of parameter,this paper implements the automatic calculation of left ventricular ejection fraction.The method proposed in this paper is superior to traditional methods and has significant clinical application value.Considering the similarity of the parameter calculation methods,these methods can also be followed in the automatic measurement of other cardiac parameters.