Study On Classification Of Diseases And Localization Of Lesions Based On Chest X-ray

With the continuous development of science and technology and people’s more attention to their own health problems,the early screening and diagnosis of chest diseases has become a hot research issue at present.However,due to the small number of professional radiologists in China and the heavy workload,we urgently need the help of computer-aided diagnosis system to complete the preliminary screening and diagnosis of chest diseases.With the continuous development of deep learning,the classification and detection of chest diseases have become an important research task in this field.The objective of this task is to classify all the lesions in the chest medical images and detect the location of the lesions to achieve accurate localization of the lesions.Detection and localization of chest diseases can effectively assist doctors in clinical diagnosis,so as to improve diagnosis efficiency and reduce misdiagnosis rate.In this thesis,aiming at the problems of poor automaticity and low accuracy of traditional chest disease classification and location algorithms,a method of disease classification and lesion location based on chest X-ray was proposed,and tested on the open data set,and achieved good results.The specific research contents and contributions of this thesis are as follows:(1)A multi-label disease classification network was proposed.Aiming at the problems of poor feature extraction and low average accuracy of traditional chest auxiliary diagnosis system in chest X-ray disease classification,a multi-level classification network combining attention mechanism and label correlation was proposed.Firstly,the experimental data was preprocessed,which was mainly divided into X-ray image preprocessing and X-ray label preprocessing.In the image preprocessing,we first observed the sample distribution of various diseases in the data set,analyzed the division way and proportion of the data set,and then scale and normalize the image.Then sort out the sample with the annotation and bind the annotation coordinates to the picture.Augment the data as necessary by data enhancement methods.In the pretreatment of labels,we used one-hot encoding form to process all sample labels,and used word2 vec method to encode disease labels to represent label characteristics.The training of network was divided into two stages.In stage 1,attention mechanism was introduced and a two-branch feature extraction network was constructed to achieve comprehensive feature extraction.In stage 2,the graph convolutional neural network was used to model the label correlation and combined with the feature extraction results of stage 1,so as to realize the task of multi-label classification of chest X-ray diseases.The experimental results show that the average AUC of various diseases in the Chest X-ray14 dataset reaches 0.846.(2)A chest disease localization network was proposed.Aiming at the problems of small sample size and complex characteristics of chest disease localization,a novel chest disease localization method based on composite backbone network and attention alignment was proposed.The network structure designed in this thesis consists of three parts: a composite backbone feature extraction network,an attention alignment module and an object detection network.First of all,the data is preprocessed,mainly aiming at the problem of uneven sample distribution,the data is enlarged,so that the sample distribution approaches equilibrium.In the part of network structure,the deep semantic features and shallow semantic features are taken into account by the composite backbone network to improve the feature extraction ability.Then the global and local attention feature images are dot-product and aligned in the attention alignment module to synchronize them to the final feature image.Finally,the aligned feature map is sent into the target detection network to obtain the final disease detection result.The experimental results show that this method has higher accuracy than other methods of chest disease detection and positioning,and has achieved significant improvement in the general evaluation indicators.