Research On Lung Nodule Malignancy Classification Based On Unsupervised Learning

Lung cancer is a medical difficult problem for the human beings to solve,and the extremely high morbidity and mortality rates caused by lung cancer have done tremendous harm to people’s health.Academic studies have shown that in order to effectively treat lung cancer and improve the survival rate of patients,computer-aided diagnosis(CAD)technology can be used to assist doctors in the diagnosis of benign and malignant pulmonary nodules.This method significantly improves the accuracy of early lung cancer diagnosis and the working efficiency of doctors.With the rapid development of artificial intelligence,computer-aided diagnosis technology based on deep learning has shown strong vitality and obvious advantages in the field of medical diagnosis.Aiming at the auxiliary diagnosis of benign and malignant pulmonary nodules,this thesis has analyzed the workflow and the advantages/disadvantages of the traditional radiomics-based classification methods and the main deep learning-based classification methods.And then we have found that the existing methodological ideas of the classification are to use supervised learning to process data sets with complete labels.However,medical image data has certain particularities.Due to the high cost of obtaining labels,medical data sets are scarce.According to this problem,this thesis designs an unsupervised deep learning model to classify benign and malignant lung nodules.This method introduces the idea of abnormal detection into the research on the diagnosis of benign and malignant lung nodules.It only needs the images of benign lung nodules as the training data of the model.By using the combination of generative adversarial network and auto-encoder to simulate the data distribution of benign lung nodules images,the mapping of benign lung nodules images is established to latent space.Through calculating the image reconstruction loss and the discriminator loss of the GAN network,the benign and malignant scores of lung nodules are obtained.The higher the score is,the higher the degree of malignancy is.This method breaks through the limitations of the common lung nodule classification methods based on deep learning,which that require large-scale labeled data sets,and avoids potential overfitting problems.The experimental results of the public data set LIDC-IDRI show that compared with other supervised deep learning methods,the method proposed in this thesis achieves a good classification effect when only benign lung nodules images are used for training.The use of unsupervised learning to complete the classification of benign and malignant lung nodules provides with new ideas for the future studies.In addition,on the basis of the above-mentioned model,this thesis proposes a strategy that we can use the integration of multiple discriminators to improve the model structure of the generative adversarial network.By weakening the abilities of discriminator,the generator can receive more positive feedback of gradient information.Therefore,this way can improve the abilities of generator to learn the data distribution of benign lung nodule images.The result of the experiment shows that this method is not only beneficial to improving the classification effect of benign and malignant lung nodules of the model,but also shortening the training time of the model.