Research On MRI Image Super Resolution Algorithm Based On WGAN

Magnetic Resonance Imaging(MRI)is a non-invasive medical imaging technique that provides accurate anatomical structure and lesion information for clinical diagnosis by generating multi-contrast images of the internal structure of the human body.However,due to the long MRI scan time,it is difficult for patients to remain still for a long time,which makes MRI imaging prone to motion artifacts,and it is difficult to obtain highresolution magnetic resonance images clinically.With the rapid development of the computer field,deep learning magnetic resonance image super-resolution technology can effectively improve image clarity,which is of great significance in auxiliary medical research and clinical diagnosis.Generative Adversarial Networks(GAN)have demonstrated superior performance in super-resolution tasks.It obtains clear and realistic data by treating the generative problem as a confrontation and game between the two networks of discriminator and generator,but due to the large required training data set and the difficulty of optimizing the two networks at the same time,GAN faces many super-resolution reconstruction problems such as artifacts in the generated image,difficult convergence,and lack of detail.In order to solve the above problems,based on WGAN(Wasserstein GAN),this paper does the following research work from the perspective of improving network structure and training strategy:(1)In view of the problems of missing diagnosis and misdiagnosis caused by artifacts that are not conducive to the doctor’s interpretation of the patient’s condition,this paper introduces a relatively average discriminator,and adds a coordinate attention module to the discriminator,which can effectively reduce the problem of artifacts in magnetic resonance images by learning important information in the image according to the importance of the feature;(2)Aiming at the problem of difficult training at the network layer,the segmented training strategy is adopted;Moreover,the adversarial loss function of Wasserstein GAN is used to solve the problem that GAN training is not easy to converge;(3)Aiming at the problem that the noise distribution is uneven and the details are not rich due to magnetic resonance imaging.In this paper,dense residual blocks are introduced into the field of magnetic resonance images in the network structure.The generator part of the network adopts a multi-scale residual dense module to ensure that more high-frequency image features are extracted.In order to verify the effectiveness of the proposed algorithm,experiments are carried out on the self-built brain magnetic resonance image dataset,and the public dataset Bra TS-2021 is used for model training and validation,and multi-center big data is used to improve the generalization of the model.The test results show that the peak signal-to-noise ratio,structural similarity and perceptual similarity are 34.6573 d B,0.9355 and 0.033 respectively on the self-built dataset.On the public dataset Bra TS-2021,the peak signal-to-noise ratio,structural similarity and perceived similarity were 34.9404 d B,0.9462 and 0.041,respectively,which were ahead of other classical super-resolution reconstruction algorithms in the control group.At the same time,the visual effect and texture details are significantly improved,which proves that the proposed network algorithm can solve the problem of super-resolution reconstruction of MRI images.