Research On Pneumonia Type Recognition Method Based On Deep Transfer Learning

In the context of the global outbreak of COVID-19,accurate diagnosis of pneumonia types has been become one of the research focuses.In medical imaging assisted diagnosis,the diagnosis of pneumonia is based on X-ray,CT and MRI.X-rays are widely used in remote areas of China due to advantages of high imaging quality and low price.However,for professional radiologists,it is easy to cause visual fatigue after reading a lot of X-rays for a long time,which inevitably leads to misdiagnosis.In recent years,it has become a mainstream trend to use deep learning method for medical image aided diagnosis.But most of the existing deep learning models are based on a large number of training data.Due to the particularity of medical images,it is difficult to have large public data sets.To solve the problem,this paper presents a deep transfer learning model for pneumonia types recognition,which can also get better recognition rate through training small data sets.This paper focuses on how to improve the recognition rate of pneumonia types,and the main research work is as follows:First,based on a large number of literature reading,the field of pneumonia recognition and the latest research status of COVID-19 were reviewed from traditional image processing,machine learning and deep learning methods.This paper summarizes the deficiencies of the existing research in the field of pneumonia recognition,and elaborates the basic theories of deep learning and transfer learning.Secondly,the target domain data set used in this paper is preprocessed.In order to maintain the relative balance of the number of chest X-ray of various categories,digital image processing technology was used to rotate and amplify COVID-19 X-ray,and in order to solve the problems of noise and low contrast in chest X-ray,a variety of filtering algorithms and gamma transform are used to denoise and enhance chest X-ray data sets in the target domain.Thirdly,on the basis of small data sets,this paper presents a deep transfer learning method for pneumonia chest X-ray recognition.After pretraining chestx-ray8,getting the base weight,then the weight is transferred to the DPFT-VGG16 model to train the target domain data set again.This model effectively avoids the problems of over-fitting caused by insufficient data and the low recognition rate of using traditional deep learning methods.DPFT-VGG16 model is proved to be excellent by the relevant experiments and the confusion matrix,the total recognition accuracy reached 93.89%.Then,in view of the problem that traditional scalar convolutional neural network in deep learning and transfer learning can’t capture the spatial information relationship between different features,a deep transfer learning recognition method fusing capsule network for pneumonia chest X-ray recognition was presented.The capsule network composed of vector neurons was fused after the convolutional layer of DPFT-VGG16 model,which was further optimized by using vector neurons and dynamic routing algorithm in the capsule network to find the spatial relationship between different features.The experimental results show that the total accuracy of DPFT-VGGCaps Net model is improved to 94.67%.Finally,the research content of this paper is summarized,and relevant suggestions for the follow-up research are given.