Research On Feature Extraction And Classification Of X-Ray Lung Image

According to the World Health Organization report,the number of children who die from pneumonia in the world is higher than that caused by other diseases.Therefore,pneumonia is consistenly estimated as the leading causes of child death.Without timely diagnosis and treatment,it will have a great impact on the patient and even endanger the patient's life.Diagnosis through X-ray images is one of the more important criteria for judging whether a patient has pneumonia.However,the number of professional radiologists in many areas is relatively small.It is of great significance to assist the doctor to diagnose through the X-ray image processing and recognition technology of the lungs of the computer,which can help the doctor to diagnose more quickly,thereby improving the diagnosis efficiency and reducing the delay of diagnosis.In order to solve this problem,this paper studied the lung X-ray images,and mainly carried out the following three aspects:data analysis and preprocessing of lung X-ray images,traditional feature extraction and classification of lung images,feature extraction and classification of lung image based deep learning method.The data analysis and preprocessing of the lung image were mainly to count the basic indicators of all lung images,and to calculate the histogram,and to preprocess the data according to the characteristics of the analysis,and lung image histogram equalization was performed to improve the contrast of the image,thereby facilitating the subsequent image segmentation operation.After the image enhancement work was completed,the paper determined the optimal threshold by Otsu's method to threshold the enhanced image,and extracted the image containing only the lung region.The extraction and classification of the traditional features of lung images were mainly based on the HOG(Histogram of Oriented Gradient)features,LBP(Local Binary Pattern)features and GLCM(Gray level co-occurrence matrix)features of the images.At the same time,this paper combined multiple features to form fusion features,and used SVM(Support Vector Machine)and Random Forest to classify features.The results showed that the method of feature fusion is better than the single feature as a classification feature.The feature extraction and classification of lung image based on deep learning method was mainly through the use of transfer learning ideas,The convolutional neural network with good classification effect in natural images was applied to pneumonia recognition through transfer learning.The feature fine-tuning layer was used to achieve the purpose of transfer learning,and the classification accuracy of the three networks was compared.It was found that the GoogleNet Inceptionv3 network had the best effect after the transfer learning of this paper,and the feature extraction effect of the convolutional neural network after transfer learning was verified.At the same time,in order to make the interpretability analysis of the deep learning model,the model was interpreted by the class activation map method and the occlusion method,and the analysis results were visualized.