Research Of Chest X-ray Image Classification And Segmentation Based On Deep Learning

The worldwide outbreak of COVID-19 pandemics has increased the bueden on the diagnosis of pneumonia and related lung diseases.With the advent of deep learning technology,it has become a mainstream development to make use of computerized diagnosis system assisting physicians in clinical diagnosis.In this study,a deep learning algorithm was implemented to classify lung disease and segment the lung parenchyma in Chest X-ray images,to improve the diagnostic accuracy and efficiency in clinical practice.The main contributions of this work include the following two aspects:1)For the study of Chest X-ray image classification,this paper revised and improved a convolutional neural network model based on the Inception-Res Net-V2 architecture by introducing a CBAM module of attention mechanism.The weight coefficients of different tasks were re-calibrated in the light of the importance of the features to achieve optimal allocation of network computing resources.Meanwhile,feature fusion layer was designed to perform nonlinear fusion of features extracted from the network.To mitigate the issues of low image quality and limited number of labeled data samples in the medical image data set,transfer learning and data enhancement methods were also applied to further improve the performance of the proposed model.2)For the segmentation of the Chest X-ray image,this paper used the encoder-decoder neural network architecture of the U-net network to achieve semantic segmentation of the lung parenchyma in Chest X-ray images.In the encoder part,we used the residual structure in the coding layers and replaced the second standard convolution with an dilated convolution to improve the receptive field range of the model.Furthermore,to enhance the feature extraction ability of the framework,we added a light-weighted attention module,SENet,to the skip connections of the residual structure to strengthen the network’s utilization of effective features and suppress the influence of irrelevant features on the model.In the decoder part,drawing on the idea of dense connection structure,the direct connections between different layers were utilized to enhance the connections between layers in the decoding module,and the model is accelerated by adding batch normalization layers and improving activation functions.This improves the convergence speed and segmentation efficiency of the model.In this thesis work,the classification of Chest X-ray images was implemented and tested using the Chest-Xray dataset,while the segmentation of the lung parenchyma was completed using the Montgomery dataset.The results from extensive testing have demonstrated that the proposed deep learning method can efficiently carry out the classification and segmentation of the Chest X-ray images.