Medical Image Analysis And Application Of Pneumoconiosis Based On Deep Learning

Pneumoconiosis is systemic disease of diffuse pulmonary fibrosis caused by long-term inhalation of industrial dust,which currently the disease with the highest inci-dence rate among occupational diseases in China.The diagnosis of pneumoconiosis is mainly based on the doctor's analysis and judgment of the lung medical image,while computer-aided medical image analysis technology has been widely used in modern medical care.In recent years,due to the great breakthrough of deep learning,medical image analysis based on deep learning is becoming more and more popular.As pneu-moconiosis is a high incidence of occupational disease in China,there are still some problems to be studied in the medical image analysis of pneumoconiosis:one is the low accuracy of pneumoconiosis recognition based on artificial feature algorithm,the other is the lack of interpretability of algorithm classification results.To solve these problems,identification of pneumoconiosis is studied in this paper the contributions are as follows:Firstly,to solve that artificial feature method has low classification accuracy,a hierarchical classification method is proposed for the classification of pneumoconio-sis.The hierarchical classification method is divided into two stages:our method first determines whether the sample is sick,and then subdivides the categories of "sick"samples.Two points are considered in the process of method design:one is that the single-level base classifier should be as accurate as possible,the other one is that the final classification result should be as accurate as possible when the all levels are combined.For solving the first problem,we compare the results of various deep con-volutional neural network classifications,including the indicators before and after they added cost-sensitive methods,and select the best one.For solving the second problem,a variety of cut points values are compared with the final classification index results in the experiment.The effectiveness of the hierarchical classification method is proved by our experiments.The experimental results show that the hierarchical classification method is more effective than the deep learning end-to-end multi-classification method.The macro-recall is increased from 0.7723 to 0.9162,and our method distinguish sam-ples of stage 3 better.Secondly,for solving that diagnosis of pneumoconiosis needs to be explained re-liably,we base on the deep learning interpretability technology to find the suspected disease areas giving the classification basis,besides,using the fine-grained features of these areas to improve the classification accuracy.Two feature visualization methods are studied to explore the state of high-dimensional abstract features of pneumoconio-sis.First of all,Gradient-weighted Class Activation Mapping(Grad-CAM)method is used to obtain the classification weights and high-dimensional features of the network feature extraction layer.The weights and features are used to find pathological areas that may be nodules,fibrils,dust spots and so on.These areas are highlighted by a heat map,which is the basis for classification,called the "discriminant area".Then,based on the discovery of discriminative areas,Selective Convolutional Descriptor Aggrega-tion method(SCDA)is used to screen the "fine-grained pathological features"from high dimensional convolution feature in the case of only the label of the period.Fine-grained pathological features refer to high-dimensional features of image pathological regions such as suspected dust spots.We further explore these features,including di-mensionality reduction for visualization observation,and train several machine learn-ing classifier with them.The results shows that macro-recall increased from 0.7723 to 0.8857 in the end-to-end method,and macro-recall raise from 0.9162 to 0.9292 in the hierarchical classification method.The effectiveness and rationality of the method is verified by our experiments.Finally,the pneumoconiosis recognition system is designed and implemented ac-cording to the requirements of the pneumoconiosis recognition project I participated in during my master's degree.The system provides services for classification of pneu-moconiosis and location of suspected disease areas,and the core algorithms of these two services are hierarchical classification methods and discriminative region discov-ery methods in interpretable techniques.The system can assist medical practitioners to diagnose pneumoconiosis faster and reduce the occurrence of missed diagnosis and misdiagnosis by double check.