Research On Lung Nodule Detection Algorithm Based On Deep Learning

Malignant lung nodules have a long window before they become lung tumors.Therefore,early screening and treatment of lung nodules is of great significance for improving the quality of life of patients.Traditional early screening and diagnosis work mainly relies on CT images of the lungs.Clinicians diagnose the nature of nodules based on basic features such as color and texture in the images.This method has high requirements for the doctor’s diagnostic level and is likely to cause missed diagnosis.,The problem of misdiagnosis.The lung nodule diagnosis method based on deep learning can dig deeper into the complex recessive representations of lung image data,and can further analyze and process deeper features to make classification and prediction more intelligent.However,in actual lung nodule detection tasks,the neural network model often has problems such as gradient disappearance and gradient explosion,resulting in poor detection results;the network model has a serious overfitting problem during the training process.In response to the above problems,this paper conducts related research on lung nodule detection based on deep learning related methods,and explores lung nodule segmentation methods based on deep learning and benign classification methods of lung nodules.The main work and research of this paper are as follows:(1)The data preprocessing method of the original LIDC-IDIR data set based on image processing technology is studied.Thickness normalization and data enhancement operations were performed on the slice data to make the data format uniform and the data volume expanded;for the lung nodule label data,the expansion and corrosion operations were performed to make the nodule information and background information significantly different,and more Conducive to feature extraction.(2)A RDU-Net deep network model suitable for lung nodule detection is proposed.This paper uses U-Net as the basic network to address the problem of poor segmentation accuracy of lung nodules due to large differences in the size,shape,and visual effects of lung nodules in lung tissues,and add a deep residual unit and a dropout layer.Constructed a lung nodule segmentation method centered on the RDU-Net network structure,which effectively avoided the gradient disappearance and gradient explosion problems in the network training process.The Dropout layer is added to the original network,which effectively improves the over-fitting phenomenon in network training.The experimental results show that the accuracy of the segmentation network reaches0.910 under the Dice index,which verifies the effectiveness of the segmentation model.(3)Constructed a multi-branch mixed model suitable for benign classification of lung nodules.In view of the current problem of unsatisfactory results in the benign classification of lung nodules,this paper uses Alex Net,Google Net and VGG16 as the basis to merge into a multi-branch hybrid model.By fusing the corresponding hierarchical network features in different models,it makes up for the information in the single network learning process.Mining the disadvantages of inadequate,enhance the full use of multi-dimensional image features.The experimental results show that the accuracy of the classification model reaches 93.3% under the ACC index,which verifies the effectiveness of the classification model.