Study On Auxiliary Screening Methods For Pneumoconiosis Based On Chest X-rays

Pneumoconiosis is a lung disease caused by long-term inhalation of productive dust,which has become the occupational disease with the highest incidence rate in China.The formation of pneumoconiosis usually takes several years.At the beginning,the lungs can not clear all dust particles,resulting in lung inflammation and temporary damage to lung tissue.If the lung inflammation is not detected in time,it will eventually lead to the deterioration of the disease,resulting in pulmonary fibrosis,and make the patient die in hypoxia.At present,the diagnosis of pneumoconiosis mainly depends on X-rays,which are checked by experienced doctors with occupational disease diagnosis qualifications.However,the number of doctors in the radiology department is limited.In order to reduce the burden of doctors,reduce the misdiagnosis rate,and accurately grasp the progress of the disease,it is necessary to use artificial intelligence to carry out auxiliary screening,reduce the subjectivity and fatigue of doctors,and improve the efficiency of disease screening.Therefore,this thesis constructs a pneumoconiosis screening model based on the chest X-rays obtained and the image characteristics of pneumoconiosis and the number of the sample.Due to the fewer number of chest X-rays of pneumoconiosis,although the traditional deep convolutional neural network can achieve better performance in the training set,it will cause the problem of overfitting of the model,resulting in poor model performance in the test.According to the characteristics of chest X-ray images of pneumoconiosis,considering the problem of a fewer amount of data,few-shot learning method is used to complete the screening of pneumoconiosis.The backbone network of the pre-training model VGG16 is used to replace the feature extraction part of the few-shot learning screening framework.In addition,the classification module is reconstructed to design a suitable network for pneumoconiosis screening,which can solve the situation that the depth learning model is unavailable due to insufficient data,and can meet the requirements of pneumoconiosis screening.The experimental analysis shows that the few-shot learning screening network constructed in this thesis has better performance than other methods that solve the problem of insufficient data.In the above initial screening model,the backbone network of the pre-trained model VGG16 is used as the feature extraction module.The module structure and parameters depend on the professional knowledge and experience of researchers,and consume a lot of time.Therefore,the neural architecture search method can be used to find a better feature extraction module network structure for pneumoconiosis screening.In this thesis,we combine the few-shot learning method with the differentiable neural architecture search method.The search space is redesigned to add candidate operations such as attention module.Through the improvement of the existing differentiable neural architecture search method,it can be well combined with the few-shot learning method,and the accuracy of the pneumoconiosis screening framework is further improved.In order to further improve the performance of the pneumoconiosis screening model,the self-knowledge distillation method is introduced.The limitation of knowledge distillation is the inefficiency of knowledge transfer,the student model cannot fully learn all the valuable knowledge of the teacher model.In addition,finding a suitable teacher model will consume a lot of experimental resources and time costs.Therefore,the method of self-knowledge distillation enables the model to use its own knowledge to continuously iterate and obtain the best screening model.Experiments show that the final screening model obtained by self-knowledge distillation has the accuracy of 94.5%,recall rate of 92.5% and precision of 91.3%,which can further improve the performance of the whole screening network.In this thesis,a few-shot learning method is used to screen pneumoconiosis using chest X-rays.The neural architecture search method is used to find the appropriate screening model.The self-knowledge distillation method is used to improve the performance of the screening model,which provides a reference for the research of subsequent screening methods for pneumoconiosis.