Automatic Myopathy Recognition Based On Digital Pathological Images

Myopathy is a kind of widespread disease.Due to the different parts and nature of the lesions,the clinical manifestations are variable,and the diagnosis is difficult.Pathological analysis is the gold standard of myopathy diagnosis,however,as the pathological manifestations of myopathy are complex,the analysis and diagnosis work often costs a lot of time and energy,meanwhile the number of experienced pathological analysis experts is very small,as a result,patients often can not get timely diagnosis and treatment.With the development of digital technology on pathological images,the muscle pathological images can be saved,transmitted and shared.On this basis,the application of machine learning,especially deep learning technology,in the analysis of digital pathological images is expected to greatly improve the efficiency and accuracy of myopathy analysis.In this thesis,the application of machine learning technology applied to automatic classification of myopathy pictures is studied as follows:Firstly,12 types of muscle pathological sections are sorted out,and the sections are magnified by 100 times one by one to generate digital pathological pictures stored in the computer.These images are cut,selected,standardized and enhanced.At last,a data set containing 192,000 pictures is produced for the training and testing of the subsequent classification models.Secondly,traditional machine learning method is used to extract texture,color,shape and HOG features of muscle pathological images,and feature fusion is carried out.In this thesis,an enhanced correlation-based feature selection(ECFS)method is adopted: the heuristic method is used to calculate the correlation between the feature category and the feature matrix,so that the features is sorted,then the best feature subset is obtained by using the first priority search to traverse all the features in order.The reduced features are input into a multi-classification SVM model for the training.ECFS improves the training efficiency and makes the trained model achieve an 83.30% classification accuracy.Thirdly,this thesis focuses on the implementation of convolutional neural networks for automatic classification of myopathy images.Using transfer learning method,convolution layer of pre training networks on Image Net is used as image feature extractor.On this basis,aiming at both global features and local detail features that must be taken into account,a multi-scale classification network(MCN)based on feature pyramid is proposed: At first,the input image is processed by multiple rounds of downsampling-convolution-upsampling to generate several feature maps with the same size but representing different scale feature information.Then,the feature pyramid strategy is used to fuse these feature maps,and the loss function is used to optimize the parameters of feature fusion.The experimental results show that compared with the single scale network model,the classification accuracy of MCN is higher,reaching 93.30%.According to the fact that the major categories of myopathy always contain many subcategories,combined with the realities of application,a progressive classification method is proposed.The main body of the method is a cluster composed of a parent network for major category recognition and several sub networks for subcategory recognition.According to the idea of divide-and-conquer,the multi-classification task is simplified,with the improved speed and accuracy.Finally,this thesis designed and developed an auxiliary diagnosis system of myopathy,including four functional modules: image analysis module,intelligent diagnosis module and models of medical records input and query,which can assist doctors in pathological analysis and myopathy diagnosis,and realize the basic functions such as the storage and access of patients’ information.