Research On Pneumonia Detection Algorithm Based On Deep Learning

Pneumonia is a common respiratory infectious disease.In recent years,due to various factors such as air pollution,drug abuse and the outbreak of rare pathogens,the pneumonia fatality rate has risen significantly,which has attracted great attention from all walks of life.Early diagnosis of pneumonia can help reduce mortality.X-ray imaging-based diagnosis is one of the most important methods for pneumonia diagnosis worldwide.However,due to the rapid growth of image data and the serious shortage of radiologists,it is of great significance to seek intelligent diagnosis methods for auxiliary diagnosis.At the same time,with the development of artificial intelligence,the application of deep learning in medical images shows great advantages.Based on the above problems,this thesis applies the target detection technology of deep learning to the detection of pneumonia images.The main contents of this thesis are as follows :(1)A CLAHELS image enhancement algorithm is proposed by combining contrast limited adaptive histogram equalization with image sharpening.While enhancing the brightness and contrast of the chest X-ray image,the contour information of the lung edge is effectively extracted,which improves the indistinguishable phenomenon between the lesion tissue and the normal tissue in the pneumonia image in the X-ray image,and improves the image quality.(2)Aiming at the serious imbalance of positive and negative samples in the original chest X-ray data set,the Focal Loss designed in the one-stage Retina Net network can better solve the problem of sample imbalance in training,and a pneumonia detection algorithm based on improved RetinaNet is proposed.In the feature extraction step of the algorithm,the conv3＿x and conv4＿x layers of Res Net50 are added to the SE module to enhance the feature extraction ability;In the step of constructing the FPN,according to the distribution of the bounding box of the pneumonia lesion area on the pneumonia image dataset,the P2 feature layer is added to the FPN module to detect the pneumonia lesion area of small targets,and the K-means clustering algorithm is used to optimize the Anchor for pneumonia lesions;In the bounding box classification and regression steps,a global loss function is added after C5,and the SE module is added after P3～P6 to further improve the effect of the pneumonia detection model.The experimental results show that the improved algorithm can preliminarily realize the classification and detection of pneumonia images.(3)In order to further improve the detection accuracy,a pneumonia detection algorithm based on the improved two-stage model Faster R-CNN is proposed,and the network designed in(2)is integrated to improve the detection effect of multiple pneumonia target regions in chest X-ray.Firstly,the experimental data is selected and expanded;Secondly,the backbone network Res Net50 is improved,its conv5 is added to the detection network to improve the performance of the classifier,and a combination of two types of hole residual blocks is introduced to generate a feature fusion model,which comprehensively utilizes shallow and deep features to better extract image features;Then the non-maximum suppression algorithm is improved;Finally,the model fusion of the pneumonia detection model based on Faster R-CNN and the pneumonia detection model based on Retina Net is carried out,and targeted post-processing is performed on the prediction frame,which further improves the accuracy of the localization of the lesion area,and completes the pneumonia detection.In summary,this thesis improves the quality of pneumonia images through the CLAHELS image enhancement algorithm;Two improved target detection algorithms are proposed,and then performs model fusion design to achieve and verify the feasibility of chest X-ray pneumonia detection and improve detection accuracy.