Evolutionary Neural Network For Recognition Of Diffuse Lung Diseases

In recent years,the application of convolutional neural networks in the medical field has received extensive attention from researchers at home and abroad,but they still face two important problems that need to be resolved: First,when faced with complex medical image processing tasks,stacked convolutional block construction The convolutional neural network cannot learn the most distinguishing unique features to classify and recognize complex lung textures,and the existence of irrelevant or redundant features causes the classification performance to fail to meet the medical and clinical requirements.Secondly,the end-to-end model makes deep neural networks generally regarded as unexplainable "black boxes".The lack of comprehensive interpretability makes it difficult to gain the trust of experts,doctors and patients.This paper studies these two issues,proposes an evolutionary neural network(Evolutionary Neural Network,Evo NN)to improve network classification performance,so that it can meet medical clinical requirements,and uses feature visualization and attribution visualization to explain the network’s decision-making process,making the network model The decision-making process is more transparent to humans and enhances the interpretability of deep neural networks.Evo NN is a network structure searched by a network structure search algorithm based on evolutionary algorithm.According to different search targets,the network structure search algorithm is applied to feature filtering and structure separation to obtain the corresponding Evo NN.First,explore in detail how the convolutional neural network recognizes features,and then use the network structure search algorithm to filter out irrelevant or redundant features in the convolutional neural network and retain the most discriminative features,thereby obtaining Evo NN with higher classification performance.Subsequently,the network structure search algorithm is further used to separate the convolutional neural network structure according to the specified category,and then the Evo NN with a simpler structure for classification and recognition of the specified category is obtained.Finally,the feature filtering experiment was performed on the real diffuse lung disease data set.The Evo NN obtained through feature filtering achieved an average classification accuracy of 95.65% and an average F value of 0.9568 in the task of texture classification and recognition of diffuse lung disease,achieving high resolution The best performance for classification and recognition of diffuse lung disease texture on CT images.In addition,the structure separation experiment was carried out on the public CIFAR data set and the real diffuse lung disease data set,and the network structure was separated according to the specified category to become independent Evo NN,and the accuracy rate varied from-0.55%to 0.55%.It belongs to the acceptable range,and the effectiveness of the algorithm is intuitively verified through feature visualization and attribution visualization.