Research On Medical Image Data Augmentation Technology Based On Generative Adversarial Network

Computer-aided diagnosis(CAD)technology has become an important diagnostic method in the medical industry,which can effectively improve the accuracy of diagnosis and reduce missed diagnosis.At present,the mainstream technology of artificial intelligence(AI)based on deep learning has obvious advantages in the field of medical diagnosis.The development of AI has greatly improved the diagnostic effect of CAD technology,but the mainstream technology of deep learning requires a large number of data sets.However,in the field of medical imaging Because of the special nature of medical data,it is difficult to solve the problem of scarce data sets.Therefore,the use of generative adversarial networks to enhance medical image data is expected to be an effective method to solve this problem.This thesis focuses on the generation of adversarial networks and focuses on the research of medical image data augmentation techniques based on the generation of adversarial networks to solve the problem of scarce medical image datasets.The main work is divided into the following parts:1)Construction of medical image dataset.In this paper,lung nodules are selected as the research object.Based on the medical image data set LIDC-IDRI(Pulmonary Nodules Open Data Set),the original image is segmented using threshold-based image segmentation,and then the XML diagnostic annotation information is read.Lung nodules were extracted based on the labeled information to construct the data set required for this paper.2)Based on the Deep Convolutional Generative Adversarial Network(DCGAN),it is combined with the capsule network and WGAN(Wasserstein GAN)to improve it and apply it in the field of medical imaging.The specific method is to use the capsule network to replace the convolutional neural network discriminator in DCGAN,and adjust the digital capsule layer in the capsule network to make it a two-class discriminator that distinguishes between true and false.Refer to the design idea of the loss function in WGAN to set up the improved model Loss function.Design a number of comparative experiments,and use the naked eye to qualitatively judge the effect of the generated images.The experimental results show that the model solves the problems of insufficient diversity and poor quality of the generated medical images to a certain extent.Although the training time is long,the training process is relatively stable.3)Use convolutional neural network to classify benign and malignant medical images,and use the accuracy,sensitivity and specificity of experimental results as evaluation indicators to quantitatively evaluate the quality of generated images.In view of the scarcity of data,a 4-fold cross-validation method was designed to select the optimal classification model as the initial experimental model.At the same time,7 different data sets were designed for testing.Combined with the existing research results at home and abroad,a comprehensive analysis from multiple aspects,Compare experimental results.The classification experiment results of convolutional neural network show that the improved generative model has good performance,and the generated images can effectively improve the accuracy of the classification model and reduce the rate of missed diagnosis.In addition,mixing traditional data augmentation methods and generating adversarial networks can greatly improve the performance of classification models,and is an effective method for medical image data augmentation.