Generative Adversarial Networks For Multi-weighted MRI

Multi-weighted Magnetic Resource（MR）images can provide more reference for accurate diagnosis of the disease,and Magnetic Resonance Imaging（MRI）method can only achieve a weighted method in an imaging process,which not only restricts the reference conditions for disease,but also increases the discomfort and medical expenses of the patients when collecting different weighted MR images.Medical image analysis and computer-aided diagnosis based on deep learning are gradually becoming the solution for accurate medical diagnosis.The emergence of deep learning solves the problem of high-dimensional feature learning.We use deep learning method to innovatively convert T₂-weighted MR images into PD-weighted MR images,thereby obtaining different weighted MR images in an imaging process.The main research work and contributions of the thesis are as follows:（1）Researching MRI method and analysing the PD weighted components contained in T₂-weighted MR images from the tissue relaxation process,which provides theoretical conditions for the conversion of T₂-weighted MR images into PD-weighted MR images.Through the study of the deep learning method,it is proved that this conversion has the technical conditions for implementation.（2）Proposing Preserving-Texture Generative Adversarial Networks（PTGAN）,using improved U-Net model as generator model of PTGAN,and deep convolutional neural network is used as a discriminator model of PTGAN.The adversarial training of generator model and discriminator model realizes the weighted conversion of T₂ and PD.In improved U-Net model,the convolutional layer is used to replace pooling layer to achieve higher-dimensional feature extraction,and the batch normalization layer is added to reduce data differences and further increase the depth of the network.（3）Using least squares loss as the basic loss of PTGAN to reduce decision boundary distance and improve the stability of model training,and adding L2 loss,frequency loss and mean square error in generator model to ensure that the texture of the structure is unchanged during the conversion in spatial and frequency domains.Using a variety of data expansion methods to increase data diversity,designing four network structures to compare with PTGAN,and testing PTGAN models using a variety of MR images.Experiments show that the PTGAN model can convert T₂-weighted images into high-quality PD-weighted MR images.Compared with the acquired weighted images,the converted MR image has a structural similarity of 0.971 and a peak signal-to-noise ratio of 32.944dB.In addition,each conversion reaches 48.4ms under a separate CPU,and only about 4ms under a separate GPU,which can quickly provide more reference information for disease diagnosis in an imaging process.