Research On Multi-View Echocardiograms Algorithm For Congenital Heart Disease Based On Deep Learning

Congenital heart disease(CHD)is the most common birth defect and the leading cause of neonatal death in China.Clinical diagnosis can be based on two-dimensional echocardiography.Existing methods based on imaging data of congenital heart disease attempt to achieve diagnostic classification through manual diagnosis and machine learning-based methods.However,these methods cannot perform deep extraction of feature information.In this paper,through the improvement and combination of the convolutional neural network and the visual Transformer module,considering the effectiveness of multi-view data compared with singleview data and the factors that small medical data sets do not support model tuning with a large number of parameters,a lightweight multi-view model is constructed.The joint-view model enables efficient feature extraction for multi-view echocardiograms and has achieved some success in diagnostic tasks.First of all,in view of the characteristics of echocardiography,this paper uses the MobileViT module to effectively learn the view data.Through the disassembly and analysis of the module,it is explained that it can effectively learn the local part of the view when it has all the receptive fields of the feature map.Feature information and long-distance dependencies.Secondly,aiming at the relatively complex problem of Transformer in this module,by comparing the computational complexity of separable self-attention mechanism and multi-head attention mechanism,its effectiveness in realizing module lightweight is illustrated.Again,considering the loss of feature information that may be caused by the convolution module during the downsampling process,the Sim-DSDC module is designed,and calculations prove that it can expand the receptive field and extract more information while reducing weight.After that,this paper analyzes the shortcomings of the ReLU activation function.With the help of the smooth approximation function,the effectiveness of the improved activation function in retaining feature information is proved by calculating the gradient of the maximum function approximation function.Finally,a comparative experiment was carried out between the proposed joint model and the baseline model for the collected congenital heart disease data.In order to prevent the unbalanced data classes from affecting the diagnosis,this paper performed reasonable data expansion on the collected data.Experimental results illustrate the effectiveness of the joint model in multi-view joint diagnosis of congenital heart disease.