Research And Application Of Pneumonitis Detection Algorithm Based On Deep Learning

In recent years,with the continuous development of hardware and software platform technology in the field of deep learning,deep learning technology can effectively help identify,classify,and quantify existing medical images,and can automatically summarize hierarchical features from the data,Which makes it has begun to slowly penetrate into some areas of medical applications,and has achieved good results.Pneumonia is an extremely important disease in medical imaging.Children become the most susceptible.A large number of patients die every year due to ineffective treatment of serious infections,which has a great impact on society and families.As we known,due to the explosive increase in the amount of medical imaging data and the serious lack of pathologists,tradional diagnostic method of pneumonitis is limited.At the same time,coupled with the difference in doctors’ level,inexperience,and heavy workload in different hospitals,it may lead to inaccurate and untimely diagnosis results and delay the patient treatment.Aiming at the problems of pneumonia detection mentioned above,this thesis carried out research on pneumonitis detection algorithm based on deep learning to abtain an effective new method.Mainly researches include:1.According to the preliminary analysis of the pneumonia characteristics of the lung X-ray image,it is known that the pneumonia detection task belongs to the target detection category.Therefore,this paper selects three typical target detection algorithms Yolov3,Retina Net,and Mask-RCNN for pneumonia detection research.The classification accuracy and regression accuracy of the YOLOV3 algorithm is 84.30% and 70.43,Retina Net algorithm is 86.20% and 76.54%,and Mask-RCNN algorithm is 88.30% and 83.13%.And the detection performance of Mask-RCNN algorithm is optimal.2.For futher improve the detection performance,the advanced Mask-RCNN algorithm is researched.Firstly,the feature fusion of Res Net50 and Res Net101 networks is used to improve the backbone network of Mask-RCNN,which improves the classification accuracy of Mask-RCNN by 2.3 percentage points and the regression accuracy by 7.39 percentage points.Then,using the deep separation convolution idea of the Mobile Net network to replace the standard convolutions in Res Net50 and Res Net101 networks with deep separable convolutions,the improved Mask-RCNN backbone network parameters are reduced by about 42% and the detection speed is reduced by 42 ms.Sencondly,the local features of the pneumonia lesions detected by the algorithm are enhanced to make the pneumonia edge and detail information more prominent,and the local features become clearer,which is more conducive to the pathologist’s diagnosis of pneumonia types.Finally,the improved Mask-RCNN network model was used to detect pneumonia images.After experimental testing,the improved Mask-RCNN algorithm’s classification accuracy and regression accuracy were 90.40% and 90.37%.Ccompared with the classification accuracy before the improvement,the classification rate and regression accuracy were increased by 2.1 and 7.24 percentage points,respectively.3.In order to make the research results operable in practical applications,also use computer software technology to build a system test platform to enable pneumonia detection to achieve a complete closed loop of "front desk submission,model detection,and result display".It can be concluded from the experimental results that the improved Mask-RCNN algorithm in this thesis is feasible and effective for the detection of pneumonia,with high detection performance and good detection effect.Pathologists can use the research rsults in the present work to assist in the development of Intelligent pneumonia diagnostic,improve the diagnostic efficiency and diagnostic accuracy of lung X-ray medical images,and thereby reduce the rate of misdiagnosis.