Research On Pneumonia Detection Method Based On Deep Learning

In recent years,the degree of air pollution has become increasingly serious,resulting in an increase in the incidence of pneumonia,a common disease of the respiratory system.Children are a high-risk group of pneumonia,and in 2015,920,000 children worldwide died of the disease.Early diagnosis of pneumonia can help reduce mortality,but medical diagnosis of pneumonia through chest X-ray is a difficult task,requiring doctors with clinical experience to observe the chest radiograph for a long time,resulting in a lack of medical resources.Early diagnosis of regional pneumonia is very difficult.In recent years,with the gradual maturity of artificial intelligence technology,deep learning has gradually developed in the field of target detection.It is also a popular trend to use deep learning target detection technology to solve medical problems.The existing target detection model is based on a natural image data set,which can be trained using a large number of natural image samples.However,for the detection of pneumonia images,the data samples are obviously inferior to the natural image samples,and the medical images are completely different from the natural images,so the algorithms on the natural images cannot be directly used on the pneumonia images.In response to this problem,this paper designs a deep learning target detection method that is specifically used on pneumonia data.Targeted pretreatment of pneumonia data was performed by analyzing the characteristics of pneumonia chest radiograph data.The existing target detection algorithm is studied.The two-stage single-stage target detection algorithm and the two-stage target detection algorithm for target detection are targeted for pneumonia data respectively.Finally,the two types of algorithms are model-fused to obtain the final Test results.The single-stage model,based on Retinanet's pneumonia detection algorithm,is designed as follows: First,adjust the size of the prediction frame to adapt to pneumonia data,and ensure that the pneumonia data yields effective test results.Second,test different backbone networks to better extract pneumonia.Features,and use the idea of migration learning to initialize the parameters;third,modify the regression branch of the network,combine multi-dimensional image features to improve the detection performance of pneumonia;Fourth,increase the global loss function after the C5 layer of the network structure It is used to classify the global picture information in advance to better extract the characteristics of pneumonia.Fifth,use the focal loss function to classify the test samples,and adjust the internal parameters to apply pneumonia data to prevent imbalance between positive and negative samples and improve pneumonia.Detection performance.The two-stage model,based on Mask-RCNN,is designed as follows: First,the segmentation branch is removed from the Mask R-CNN according to the annotation format of the pneumonia dataset;secondly,the size of the prediction frame in the first phase of the adjustment algorithm ensures pneumonia The data yielded effective test results.Thirdly,different backbone networks were tested to better extract pneumonia characteristics,and the parameters were initialized using the idea of migration learning.Fourthly,a staged training method was proposed,which was phased training.The parameters help to improve the accuracy of the model algorithm.Finally,the model of the pneumonia detection algorithm based on Mask R-CNN and the Retinanet-based pneumonia detection model are combined to improve the accuracy of the pneumonia detection algorithm and complete the detection of pneumonia.The use of the above deep learning target detection technology to classify and detect pneumonia chest radiographs is helpful for preliminary screening and labeling of a large number of pneumonia chest radiographs,improving the work efficiency of doctors,and providing medical assistance to areas with scarce medical resources.