Non-Cartesian Under-sampling Reconstruction Of Magnetic Resonance Imaging Based On Convolutional Network Model

Magnetic resonance imaging(MRI)is a kind of imaging technology based on the principle of Nuclear Magnetic Resonance,and is an important means of disease diagnosis and clinical research.However,MRI has been faced with the bottleneck problems of slow scanning speed and long imaging time since its birth.Therefore,accelerating the speed of magnetic resonance imaging has always been an urgent problem in the field of magnetic resonance.Non-Cartesian sampling in k-space is one of the most important methods for accelerating the scanning speed of MRI.However,the data sampled by the non-Cartesian cannot be used for the fast Fourier transform directly,so that the image reconstruction is more complex and difficult to carry out the subsequent quantitative analysis.At present,deep learning has a wide range of applications in MRI reconstruction.The reconstruction method based on deep learning has achieved better results than the traditional algorithm.However,there are few studies on deep learning reconstruction for non-Cartesian sampling.Therefore,this paper conducts the following research on the reconstruction of radial sampling in non-Cartesian sampling.1.This paper proposes a non-Cartesian sampling MRI reconstruction network based on a cascaded structure.The cascaded network is extended to the non-Cartesian case by adding a non-uniform fast Fourier transform layer.We add the attention mechanism to U-Net to improve the denoising ability of the sub-denoiser.The experimental results on the dataset provided by the cooperative organization show that this method has achieved a reconstruction effect that is superior to that of the comparison algorithm.2.This paper proposes a segmentation task-driven MRI frequency-domain reconstruction network.It combines the segmentation task with the image reconstruction task.Besides,a dynamic weighted loss strategy is developed to optimize the total network loss function.And,an iterative teacher supervision strategy is designed to reduce the interference for the downstream task from the upstream task.The experimental results on the dataset provided by the cooperative organization show that the method has achieved better segmentation and reconstruction performance than the comparison algorithm.