Research On Detection Algorithom Of Pneumonia Lesions In CXR Images Based On Deep Learning

Chest X-ray(CXR)images are an important basis for the diagnosis of pneumonia.Diagnosis of pneumonia usually requires professional doctors or radiologists to interpret and diagnose CXR images.Manual interpretation of CXR images takes a lot of time and labor,and is easily affected by subjective factors leading to misdiagnosis.Therefore,automatic detection of pneumonia lesions from CXR images has important practical value.In this thesis,deep learning techniques are used to study the detection methods of pneumonia lesions in CXR images.The main research work is as follows:1.Aiming at the low accuracy of the pneumonia lesion detection algorithm in CXR images,a detection algorithm of pneumonia lesions based on Faster R-CNN was proposed.The algorithm adds a cavity convolution in the feature extraction network and adjusts the downsampling factor,so that the feature image retains more details of the pneumonia lesion and has a larger spatial receptive field.The algorithm firstly extracts the feature image of CXR image by CNN network;secondly,extracts the foreground candidate region containing pneumonia lesions by RPN network,and uses ROI Pooling to determine the candidate region size;finally,the candidate region is tested for pneumonia lesions.Through training and testing on the RSNA pneumonia dataset,and comparing with common detection algorithms,the effectiveness and superiority of the proposed algorithm are illustrated by the results.2.In order to further improve the accuracy of pneumonia detection,a method for detecting pneumonia lesions based on improved Retina Net network was proposed.The method firstly enhances the CXR image contrast by using the CLAHE algorithm,and uses the K-means algorithm to calculate the candidate region aspect ratio according to the annotation information.Then,the feature image of CXR is extracted by CNN network.The CNN network highlights the feature channel containing the information of pneumonia lesions in the feature image through the feature channel attention mechanism,and suppresses the feature channel containing a lot of noise information.Finally,candidate regions are generated based on the pre-set candidate frame parameters,and regions containing pneumonia lesions are selected by the classification/regression sub-network.The results show that compared with the common detection network,the detection method of pneumonia lesion based on improved Retina Net network can effectively improve the detection accuracy of pneumonia lesions,and the AP value can reach 82.52%.When multi-model joint detection,the detection accuracy can be further improved,and the AP value can reach 89.08%.