Research On Classification Of Benign And Malignant Lung Nodules Based On Transfer Learning

Today,lung cancer has become one of the malignant tumors with the fastest increase in morbidity and mortality and the greatest threat to human health and life.If lung cancer can be diagnosed and removed at an early stage,it can greatly improve the survival rate of patients.Currently,biopsy is the "gold standard" for clinically determining benign and malignant tumors,but it has considerable limitations.The main driving factor of lung cancer is the size of lung nodules,which are divided into benign and malignant types.The classification of benign and malignant lung nodules in lung CT images is an innovative method for clinical screening and early diagnosis of lung cancer.Computeraided diagnosis technology can assist doctors in the classification of benign and malignant pulmonary nodules,and improve the diagnostic accuracy and screening efficiency.Deep learning has great potential in the study of benign and malignant pulmonary nodules classification,and transfer learning can solve the contradiction between the small scale of medical data and the need for a large amount of training data for deep learning,so it is expected to be widely used in this field.This paper studies and establishes a benign and malignant classification model of lung nodules based on transfer learning.The specific work content is as follows:(1)A method for multi-scale sampling and synthesis of image input is proposed,and a benign and malignant pulmonary nodule classification model based on multi-scale transfer learning is established.This method can make full use of image information at different scales such as the core area of the lung nodule,the overall area,and the peripheral environment to improve the classification accuracy.At the same time,the performance requirements for the front-level lung nodule detection network are also significantly reduced,and only the nodule needs to be detected.The center point of the knot is enough.In the 10-fold cross-validation experiment conducted on the LIDC-IDRI dataset,the average classification accuracy of the proposed multi-scale transfer learning model on the test set reached 90.26%,and the AUC reached 0.956.Compared with other deep learning models,the multi-scale transfer learning model has achieved better results.(2)A method for categorizing benign and malignant lung nodules based on the attention mechanism is proposed.After extracting image features through the residual network Res Net50,this method introduces an attention mechanism to make the model focus on more critical areas.Among them,the attention module is composed of two parts:a bilinear pooling module and a dual attention module.In the course of the experiment,a training method that alternately opens the number of convolutional network layers and changes the learning rate at the same time is also proposed,which can effectively accelerate the convergence speed of training and avoid the problem of local optimality.The experimental results show that this method is superior to other methods.(3)A benign and malignant classification model of lung nodules based on partial adversarial domain adaptation is proposed to solve the problem of mismatch between the source domain and target domain label spaces in migration learning.This paper builds a data set based on the LIDC database and the image data provided by the hospital,uses the deep convolutional neural network as the feature extractor,extracts transferable features through the confrontation training between the feature extraction and the domain discriminator,and adds the non-confrontation domain discrimination at the same time This method is compared with several migration learning algorithms such as DANN,SAN,PDDA and so on.Experimental results show that this method can effectively solve the problem of spatial mismatch between different domains.