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

Computer-aided diagnosis(CAD)technology can improve the efficiency of the doctor’s diagnosis and reduce misdiagnosis rate,has become an important research field in medical image.At the same time,the success of deep learning technology has greatly improved the diagnostic effect of CAD and further promoted the research and development of CAD.However,the effect of deep learning model is affected by the size and quality of the training datasets.And due to the particularity of medical images,medical image datasets often have the problems of small amount of data,class imbalance and low image resolution.Therefore,how to effectively perform data augmentation on medical image datasets to generate multi-category and high-quality large-scale medical images is an urgent problem to be solved.In order to improve the above situation,this thesis has carried out the following three studies based on the generation network technology:Aiming at the problem of small amount of data in medical image datasets,this thesis studies the data augmentation of small-scale medical image datasets,and proposes a generative adversarial network model based on feature loss suitable for small-scale medical image datasets.The model uses feature extraction network to extract multi-level features of the image and introduces feature loss so that the model can fully learn the multi-scale features of the original image.In this thesis,experiments are carried out on several fundus image datasets,and the experimental results prove that the proposed model can generate high-quality fundus images even with small training sets.Finally,the model is used to expand the data of the vascular segmentation experiment,and it can be found that the performance of the segmentation model has been improved,which proves that the fundus images generated by the model have practical application value,and can effectively augment the datasets of medical image.Aiming at the problem of class imbalance in medical image datasets,this thesis studies the generation of multi-class medical images,and proposes an end-to-end generative adversarial network model based on domain encoder and mapping network.The model uses the domain encoder and mapping network to generate the domain code of the specified domain image,and the generator uses the original image and the domain code as the input information to generate the image of the specified domain.In this thesis,experiments are carried out on the internal fundus dataset.And by comparing with multiple models,it is proved that the model can generate fundus images of different modalities and different classes on the premise of retaining the features such as blood vessels of the original image,and can effectively augment multi-class data of medical image datasets.Aiming at the problem of low resolution of medical image datasets,this thesis studies the generation of high-resolution medical images,and proposes a generative adversarial network model based on residual structure and attention mechanism to enhance the quality of medical images.The model adds channel attention mechanism and spatial attention mechanism to the generator to retain the details of the original image.In this thesis,experiments are carried out on fundus image datasets,lung CT image dataset and brain MRI image dataset.By comparing with multiple models,it is proved that the model can generate high-quality super-resolution images.