Study On Deep Learning Based Classification Of Lung Perfusion Images

SPECT lung perfusion nuclear medicine imaging technology has been widely used in the diagnosis of lung diseases,in clinical treatment is mainly manual image reading,because of the low imaging quality of lung perfusion nuclear medicine images,individual differences and other factors,for doctors to read the efficiency and reliability of disease diagnosis has brought great obstacles.In this thesis,an automatic classification method of SPECT lung perfusion images based on deep learning model was studied on the basis of the preliminary optimization processing of image data,in order to provide an efficient and reliable auxiliary diagnosis of lung perfusion nuclear medical images.The research work of this thesis is mainly reflected in the following aspects:(1)Preprocessing of nuclear medical imaging data of SPECT lung perfusion.The original pulmonary perfusion DICOM file data was converted into gray-scale images using the normalization technology to improve the impact of individual differences and the huge variation range of original data values on the classification effect.In view of the phenomenon that the deep learning model is easy to overfit caused by the small amount of medical image data and the data imbalance,image translation,rotation and transfer learning methods are used to effectively solve the problem.(2)Residual model of pulmonary perfusion image classification with feature fusion.On the basis of in-depth analysis of functional imaging of SPECT lung perfusion nuclear medicine,using the RESNET-50 network as the basic classification model framework,transfer learning was introduced to effectively improve the generalization ability of the classification model,and feature fusion method was introduced to improve the effective use of the deep feature information of the network,so as to improve the accuracy of model recognition.Experimental results show that this classification method can effectively detect lung lesions,and the classification accuracy of 95.5% can be achieved.(3)VGG multi-classification model of pulmonary perfusion images with attention mechanism.In order to further automatically identify the severity of lung disease,the VGG network was used as the basic classification model framework,and the attention mechanism was introduced to improve the ability of the model to extract features from lung perfusion images.The experimental results show that this classification method can effectively detect lung severity,and the four-classification accuracy of 88.1% can be obtained.(4)Based on the above research contents,this thesis designed and implemented an auxiliary diagnosis system integrating nuclear medicine image data management,data preprocessing,classifier selection and automatic recognition of pulmonary perfusion images.In this thesis,a large number of comparative experiments were designed in the classification experiment of SPECT lung perfusion images.The experimental results show that the improved model has a good performance in terms of classification accuracy,sensitivity and specificity,and has a good improvement effect compared with the original model,which proves the applicability of the experimental method.By constructing the prototype classification system of SPECT lung perfusion images,it can assist doctors in the diagnosis of lung diseases.