Research On Diagnosis Method Of Pulmonary Nodules Based Generative Adversarial Networks

At present,the CT data of patients shows an exponential growth,which causes a serious burden to the doctors in the radiology department.The high intensity of the doctor’s reading work is easy to cause the fact of misdiagnosis.In order to improve the accuracy of the diagnosis of pulmonary nodules,computeraided diagnosis is particularly important.Computer-aided diagnosis technology extracts different types of features(such as color,texture,etc.)of pulmonary nodules in different ways,and integrates the feature data to comprehensively evaluate the lung nodules to provide a reference for the doctor’s decision.However,for the CT data with complex tissue structure,only using traditional methods to extract shallow features,the diagnosis results are often inaccurate.Deep learning extracts distinguishing features of data through convolution operations to further analyze the data.Generative adversarial network includes generative network and discriminant network.These two structures make the generative network output ideal result through continuous confrontation.This article mainly combines the convolutional neural network and the generative adversarial network to study the computer-aided diagnosis system of lung nodules.The main innovations made for the research content in this article are as follows.(1)The use of deep learning technology depends on massive medical data.Due to the lack of lung nodule data,this paper proposes a multi-label GAN to expand pulmonary nodules dataset.This network innovatively proposes to use various depth features of lung nodules as label input generators,so that the generator can generate ideal samples according to the feature labels.The generative network uses feature fusion and deconvolution techniques to map lowdimensional vectors to high-dimensional data.The discriminant network uses convolution and pooling operations to extract deep features of training samples and samples output by the generator.The discriminant network output 1 or 0 to judge the training data and generated data of the network.At the same time,the training data and the generated data are input into the classifier,and the classifier performs the correct label classification on the input data.These three structures are performed simultaneously and iteratively,so that the output contains the distinguishing features of lung nodules but is not completely consistent with the input.Compared with the existing data enhancement technology,this method has better peak signal-noise ratio and structural similarity,and is a more reliable data enhancement technology.The enhanced diversity data provides a rich data source for the diagnosis of benign and malignant pulmonary nodules.(2)At present,most mature segmentation networks require post-processing technology to achieve global correlation,which is time-consuming and inefficient.In order to solve this problem,this paper proposes U-GAN network to segment pulmonary nodules.This method innovatively combines the generative adversarial network with the deep convolutional network.The generator uses UNet framework which is suitable for medical image segmentation and is a mature network currently,and uses its unique symmetric path to segment the lung nodules.Convolution and pooling operations are used to extract features of different resolution levels from top to bottom,using data-dependent upsampling methods and feature fusion techniques to recover detail information for lowresolution data,and finally output the segmentation result map.The discriminant network discriminates the paired input images(referring to the original image and the ground truth or the ground truth and the segmentation result image)through extracting depth features.The output of the discriminator is returned to the generative network,using the gradient descent method to guide the training direction of the generated network.Experiments show that the method achieves a higher mean intersection over union,about 88.14%.This way provides a certain reference to diagnosis of lung nodules.