Application Of Improved Convolutional Neural Network In The Recognition Of Pneumonia Chest Radiographs

At present,the diagnosis method of pneumonia mainly relies on radiologists to make artificial judgments on chest radiographs,but artificial judgments require higher professional standards and they are easily affected by the doctor’s personal subjective,consciousness and physical and mental state,especially in the hospital environment.At the same time,due to the characteristics of the two-dimensional projection of chest radiographs,various organs in the body overlap,which can easily lead to misjudgment.Therefore,using an application with high accuracy and automatic diagnosis to assist doctors has an important application scenario.This paper builds a pneumonia recognition model for chest radiograph images based on the latest related technologies of convolutional neural networks.The specific work is as follows:(1)Pneumonia image recognition based on improved transfer learning model and voting classifierFirst,this article optimizes the migration learning model and builds a feature classification layer.In this classification layer,a global average pooling layer is used instead of the commonly used fully connected layer,1×1-size convolutional dimensionality reduction,and group normalization algorithms that are not limited by batch size are used.This paper uses this structure to replace part of the original transfer learning structure to output the classification results,using two large-scale transfer learning models as the main body,using different learning rates and the Adamax optimizer for training.This paper compares the former with the original transfer learning network model,and the accuracy has been improved.Secondly,this paper also constructs a voting classifier based on probability,which combines the above models and the improved Dense Net-201 model.After using this classifier,the accuracy rate reaches98.37% through experiments,which is higher than that of a single model.Compared with the commonly used basic models and other paper models,the accuracy rates are increased by 1.98-13.25% respectively,which proves the effectiveness of the method in this paper.(2)Pneumonia image recognition based on improved Inception V1 network and SENet networkFirst of all,in order to achieve a balance between the amount of model parameters and accuracy,this paper made a lot of improvements on the basis of Inception V1 and SENet network to build a shallow Se In＿CNN model,compared with the original Inception V1 module,added a new parallel channel,and expanded The receptive field was improved,and the activation function was improved.Compared with the original SENet module,a fully connected layer is added,the layer normalization algorithm is embedded,and the activation function is optimized.At the end of this paper,multiple evaluation indicators are used to prove the effectiveness of the model in this paper.At the same time,the model in this paper is compared with multiple models,with fewer parameters and higher accuracy.Secondly,this article takes the direction of further reducing the amount of parameters and is inspired by the layer jump connection.Based on the former,the Se In Res＿CNN model is constructed.Although the accuracy of the model is slightly lower than the former,the parameter amount is only 20% of the former.Proved that the model is effective.