Research On Classification Algorithms Of Pulmonary Nodules Detection Based On CT Images

Lung cancer has the highest morbidity and mortality rate in our country.Because the early clinical manifestations of lung cancer are not obvious,most patients are in the middle and late stages of lung cancer when clinical symptoms appear.Even after active treatment such as surgery,radiotherapy and chemotherapy,the survival rate of patients is still low.Therefore,if early screening,early detection and early diagnosis of lung cancer can be carried out,the survival rate of patients can be effectively improved and the quality of life of patients can be improved.The early manifestation of lung cancer is pulmonary nodules,so the detection and diagnosis of pulmonary nodules is of great significance.In recent years,Deep learning has achieved great success in the field of image.Deep learning technology has also been widely used in the diagnosis of lung cancer,mainly focusing on lung CT images.In the detection system of pulmonary nodules,the rapid and accurate detection of pulmonary nodules is a challenging task.At present,most of the existing studies have the problem of high false positive rate.To solve this problem,this paper focuses on the false positive removal algorithm based on ResNet model.The main work of this paper is as follows: Firstly,the analysis of DICOM standard and DICOM file format is introduced.Then,the characteristics of medical CT image and data preprocessing are introduced.The development and basic structure of convolutional neural network are introduced.Four common convolutional neural networks are introduced: AlexNet,VGGNet,GoogLeNet,ResNet,and their structures and characteristics are described respectively.Secondly,the data set,LUNA16 data set used in this experiment is introduced.The distribution of data in the data set is analyzed,and the distance of voxels in the data set is set.Because the number of positive samples in the data set is too small,the data enhancement strategy is adopted to expand the positive samples.Before designing the structure of the network model,this paper first analyses the influence of the size of the local slice of the input image on the performance of the model,and gives the same 3D CNN model of the five main structures used by GorkemPolat.The size of the local slice of the input image of the five models is different.On this basis,this paper designs a ResNet-based network model structure,which is compared with the original 3D CNN model.The ResNet-based network model increases the number of layers of the network,and adjusts the size and number of convolution cores.In order to reduce the parameters of the network and accelerate the training speed of the network model,the global average pooling is used to replace the full connection layer.Then,Batch Normalization and stochastic gradient descent optimization algorithm are used to optimize the network.Then,the method of network training is introduced.Finally,the experimental results of this paper are analyzed.The experimental results show that the algorithm designed in this paper can effectively remove false positives and has high classification accuracy.