Research On Key Technologies Of Magnetic Resonance Images Denoising Based On Deep Learning

Magnetic Resonance Imaging technology has the advantages of safety and noninvasive.It is one of the indispensable means of clinical medical auxiliary diagnosis.It is widely used to obtain the information of human anatomical structure and physiological function.In recent years,MRI technology has greatly improved in imaging speed and resolution,but it is still inevitably affected by noise in the acquisition process.Most of the existing MR denoising methods are based on the undersampling noise and ignore the signal sampling noise.Considering the noise in the process of signal acquisition and under sampling at the same time,this paper constructs the MR noise simulation model?A parallel denoising network in K space and image domain is proposed? A denoising model of MR image based on multi strategy learning is proposed? Aiming at 3D MR image,a 3D MR denoising model based on Generative Adversarial Networks is proposed.The main research contents are divided into three parts:（1）A MR image denoising network based on parallel learning in K space and image domain is proposed.Firstly,a noise simulation model is constructed to simulate the noise generated in the MR imaging process.The simulated Rician noise is firstly added to the fully sampled MR image,and then the mask is used for under sampling to obtain the final simulated noisy MR image.Then a parallel network is constructed to learn both k-space and image domain information,and the learned information is fused.The parallel learning network can deal with mixed noise effectively,and can also get better denoising effect for single noise.（2）A multi strategy learning MR image denoising model is proposed.The model can learn k-space and image domain,content and noise at the same time.An optimization model is proposed to reduce the training cost without affecting the denoising performance.Experiments show that satisfactory denoising performance can be achieved,and the image structure and texture information can be preserved.（3）A 3D MR denoising model based on Generative Adversarial Networks is proposed.Since most of MR clinical data are 3D sequence images,this paper proposes MSGANs（including MS-WGAN and MS-Patch GAN）based on 3D convolution and GAN.A Depthwise Separable Convolution module is introduced to reduce the computational cost of the model.Experiments show that MS-GANs can obtain satisfactory denoising performance and retain the structure information of the image well.