Research On Congenital Heart Disease Diagnosis Algorithm Based On Weighted Feature Fusion

Congenital heart disease（CHD）is the most common birth defect.Neonatal deaths in China are mostly due to congenital heart disease.Early computer-aided diagnosis attempts to use a single mode for feature extraction,but this does not provide sufficient information and the performance of the model is poor.Congenital heart disease diagnosis based on multi-view data has become a more reliable auxiliary diagnosis technology.This paper will study the problem of data loss in multi-view data,the fusion of multi-channel feature addition fusion strategy of convolutional neural network,and the problem of information redundancy between views in multi-view classification.Specifically summarized as follows:Firstly,for the problem of missing data in multi-view data,Cycle-GAN and Star-GAN can realize domain-to-domain and multi-domain image generation.However,in the multi-view congenital heart disease dataset,there is a one-to-one correspondence between the two views of a sample.Therefore,this paper proposes Sup-Cycle and Sup-Star to solve the problem that Cycle-GAN and Star-GAN cannot achieve one-to-one correspondence between input images and generated images when generating data.Secondly,aiming at the problem of information redundancy in multi-view data,this paper proposes a new deep learning classification model（M-CNN）,which adds multiple pooling layers to the feature extraction module of the model to alleviate the information redundancy in multi-view data.Furthermore,aiming at the problem that the weight distribution between the features of the convolutional neural network addition fusion strategy is too balanced when fusing features,an adaptive addition fusion strategy is proposed.The fusion strategy adds weight parameters on the basis of the addition fusion strategy so that the model can automatically adjust the weight between each view according to the training set.In addition,in view of the fact that the image generated by GAN is more dependent on the sample size of the training data set,this paper uses data enhancement to alleviate the problem of small sample size of the training data.At the same time,this paper adds ₂L norm to the loss function of the classification real model to improve the generalization of the model,and explains why the ₂L norm can improve the generalization of the model from a mathematical point of view.In order to verify the relevant conclusions,the image generation experiment and multi-view image classification experiment in this paper are carried out under the three data sets of gray image and color image of A4 C section,gray image of 5 sections such as A4 C and color image of 5 sections such as A4 C.In the experiment,the gray image of A4 C section and the color image of A4 C section are used as the data set.The experimental results show that the images generated by Sup-Cycle are more similar to the real images than the images generated by Pix2pix and Sup-Star.The results of multi-view classification experiments show that compared with the use of zero matrix or the same type of completion to fill the missing facets,the data set filled with Sup-Cycle and Sup-Star has better classification results in classification experiments.Finally,compared with four multi-view classification models such as Dsc-net and MVCNN,it is found that the classification accuracy of M-CNN using adaptive addition fusion strategy is slightly improved.