Research On MRI Super-resolution Image Reconstruction By Using Attention Mechanism Based On Generative Adversarial Network

Magnetic resonance imaging is a common imaging method in medicine,which can image internal tissue structures non-invasively.It has the advantages of high imaging resolution and no radiation in scanning.However,the resolution of MRI images is limited by many factors such as hardware conditions,signal-to-noise ratio,and scan time.Therefore,the research on the reconstruction algorithm of super-resolution magnetic resonance images has important research significance.Image super-resolution is to reconstruct a high-resolution image from a single low resolution image without changing the hardware conditions.At present,with the widespread application of deep learning technology in the field of computer vision,image super-resolution reconstruction technology based on convolutional neural network has achieved significant results in MR image super-resolution reconstruction.MRI image reconstruction by using deep learning method has become the focus of super-resolution magnetic resonance imaging research.This article mainly focuses on generative adversarial network to reconstruct the magnetic resonance super-resolution image.The specific research content is as follows:(1)In this thesis,a reconstruction framework by using Self-Attention based Super-Resolution Generative Adversarial Networks(SA-SR-GAN)is proposed to generate super resolution MR image from low resolution MR image.Moreover,the Self-Attention mechanism is integrated into Super-Resolution Generative Adversarial Networks(SR-GAN)framework,which is used to calculate the weight parameters of the input features.At the same time,spectral normalization is added to make the discriminator network training process more stable.The network was trained with 403 D images(each 3D image contains 256 slices)and tested with 10 images.The experimental results show that the PSNR and SSIM values of the super-resolution magnetic resonance image generated by the proposed SR-SA-GAN method are higher than the state-of-the-art reconstruction methods.(2)In this thesis,Fused Attentive Generative Adversarial Networks(FA-GAN)is proposed for super-resolution MR image reconstruction.The generative adversarial network is used to generate high-resolution magnetic resonance images from low-resolution magnetic resonance images,and the attention mechanism is integrated into the super-resolution reconstruction method.In the framework of the FA-GAN,a combination of channel attention and self-attention is used to calculate the weight parameters of the input features.Moreover,the spectral normalization process is introduced to make the discriminator network more stable.Experimental results show that the proposed FA-GAN method can further improve the accuracy of the super-resolution image.This thesis combines the attention mechanism and the use of generative adversarial networks to reconstruct high-resolution magnetic resonance images.The algorithm in this thesis is superior to the classical methods in terms of signal-to-noise ratio and structural similarity indicators,while maintaining the stability of model training.