Weight-K-Space Generative Model For Magnetic Resonance Imaging Reconstruction

Magnetic resonance imaging(MRI)is a non-invasive and non-ionizing imaging technique,which is widely applied for clinical diagnosis and medicine research.MRI has better soft tissue contrasts than many other medical imaging modalities.However,MRI has a relatively slow acquisition procedure due to underlying physical and physiological constraints and has negative implications with regard to image quality,patient discomfort.A major way to decrease MRI acquisition time is to decrease the amount of data acquired.Many techniques have been developed to reconstruct the desired image from undersampled measured data to achieve accelerated MRI.Recently,deep learning based parallel imaging has made great progresses in recent years to accelerate MRI.Nevertheless,it still has some limitations,such as the robustness and flexibility of existing methods have great deficiency.Most approaches have so far trained the network in the image domain and only few works directly train in the original K-space domain.Using a generative model to directly interpolate the undersampled data in K-space can avoid the distortion or disappearance of detailed structures caused by the Zero-filled reconstruction.In this work,we propose a scorebased generative model to explore K-space prior information for flexible calibrationfree PI reconstruction,coined weight-K-space generative model(WKGM).The main research work is as follows:(1)WKGM is a K-space domain unsupervised generative model for training and reconstruction.The weighting technology and high-dimensional space augmentation design are efficiently incorporated to process the K-space data for model training and reconstruction.The use of weighting technology can suppress the low frequency information of K-space data and improve the high frequency information.The generative model can learn more sufficient edge information and details.By using channel replication technology,the low-dimensional weighted data is transformed into high-dimensional spatial data.It enables the generative network to more fully learn the prior information of the weighted high-dimensional data,thereby exhibiting good reconstruction performance and generalization ability.The specific performance is that image reconstruction can achieve good and stable results under with varying sampling patterns and acceleration factors.(2)Compared with supervised deep learning technology,unsupervised MRI reconstruction technology avoids a large amount of label data,but the reconstruction quality needs to be improved.WKGM is flexible and thus can be synergistically combined with various traditional k-space PI models(such as SAKE)in the iterative reconstruction process.The proposed generative model performs parallel reconstruction in a coil-by-coil manner,combined with the traditional K-space PI algorithms to makes full use of the correlation between multi-coil data to realize calibration-free parallel imaging,thereby further improving the reconstruction performance of magnetic resonance images.