Segmentation And Classification Of Lung Nodule Images Based On Generative Adversarial Networks

In recent years,the mortality rate of lung cancer has increased significantly,seriously affecting the normal life of human beings.Lung cancer mainly presents in the early stages as lung nodules,which vary in shape and size.Manual outlining and screening is required for clinical diagnosis and treatment,and doctors rely on their expertise and clinical experience when screening for nodules.However,with a large number of patients,the examination can produce multiple dimensions of lung tomography(CT)images,which not only increases the workload of the doctor but also makes it easier for errors to occur,resulting in patients not being treated on schedule and missing the best time for treatment.In the process of Computer Aided Diagnosis(CAD),the detection of the location of lung nodules,their segmentation and classification as benign or malignant are important steps in the diagnosis.Traditional methods of segmenting lung nodules are unable to accurately segment nodules and are poorly shared across multiple types of nodules.With the emergence of intelligent technologies,more and more research has found that applying deep learning methods to the medical field has farreaching implications,and deep learning methods require a large amount of data to support them.In this study,a deep learning correlation network model is used to accurately segment lung nodules with the aid of computer-aided diagnosis,and a network structure with high accuracy and applicability to various types of nodules is designed.In addition,a method of classifying benign and malignant lung nodules with enhancement of the data is proposed for the situation where medical image data is small.The main research components of the thesis are as follows:(1)In this paper,an improved Deep Convolutional Generative Adversarial Networdks(DCGAN)model is proposed for use in the segmentation task of lung nodules.The improved DCGAN model mainly replaces the generator model with the V-Net model,which is popular in medical image processing tasks as the proposed V-Net model is mainly for 3D images.The biggest advantage of the V-Net model is the addition of a residual learning mechanism at each layer,which effectively alleviates the gradient disappearance of the model during training,and the whole network uses the Parametric Rectified Linear Unit(Parametric Rectified Linear Unit(PRe Lu)activation function is used for the whole network.In this model,the convolutional layer is used instead of the pooling layer for the downsampling operation,and the Dice loss function is used in the generator to solve the problem of the difference between foreground and background pixels of the image,so that the accurate segmentation of the 3D image is achieved.In this experiment,the accuracy of the discriminator in classifying the real image and the generator-generated image is evaluated using the Binary Cross Entropy Loss(BCE Loss)function.To obtain better results,the experiments used the publicly available Lung Nodule Analysis 16(LUNA16)dataset,which was subjected to a series of pre-processing processes before the data was used for training,testing and validation in a ratio of 8:1:1 respectively.The final result of 98.2% accuracy and 81.9% precision was achieved in the lung nodule segmentation task,which is an improvement over the baseline model V-Net.(2)In this paper,the DCGAN model was used to augment the publicly available LIDCIDRI dataset in response to the difficulty of obtaining medical data.The original data,traditional data augmentation methods and DCGAN-based data augmentation methods were used in the training of a benign and malignant classification model for lung nodules,and a study on the benign and malignant classification task of lung nodules was realized.The LIDC-IDRI dataset was first pre-processed to a certain extent,and the processed data were used in an 8:2ratio for training and testing respectively.The experimental results revealed that the evaluation index values of the DCGAN-based data enhancement for benign and malignant classification of lung nodules were higher than those of the original data and the traditional data enhancement method.