Research On Fast Reconstruction Algorithm For High Fidelity Magnetic Resonance Imaging

Over the past decade,magnetic resonance(MR)imaging has become increasingly popular in radiology and medicine thanks to its advantages of being non-radiative,having high spatial resolution,and providing superior soft tissue contrast.However,in the pro-cess of MR imaging,the physical conditions often make the conventional scanning signal acquisition time-consuming,which would cause certain difficulties for the examination of coma-critically ill patients,restless patients,and children.In addition,the need for long periods of data collection and breath-holding,as well as the patient’s unconscious or voluntary movement,often result in critical information loss or motion artifacts.With the advent of big data and the development of deep learning,it has become particularly important to learn meaningful information from large-scale medical image data through artificial intelligence technologies to overcome the shortcomings caused by physical hardware.Therefore,applying deep learning for medical imaging has become a hot research topic.However,the current deep learning-based MR image reconstruc-tion methods either i)treat the complex MR as two independent channels and neglects the relationship between them;or ii)take amplitude images as the input and lose thep-hase information.Thus,these method do not make good use of the information and suffer from serious information loss.In addition,in clinical diagnosis,it is often to scan multiple modalities at the same time for the lesions observation,and the correlation between these different modalities has not been effectively explored.In order to solve the above prob-lems,the solutions to the fast MR imaging problem can be summed up in two directions:front-end fast reconstruction and back-end super-resolution enhancement.From these two perspectives,we propose an efficient multi-frequency complex convolution operation,a transformer-based MR imaging network,a multi-modal-based fusion scheme,and a multi-task joint optimization framework to solve the problems of accelerated and high-resolution MR imaging.The main research contents and contributions are summarized as follows.(1)From the perspective of fast front-end reconstruction,we propose the Dual-Octave Network(DONet)for fast MR image reconstruction.The proposed DONet can learn multi-scale spatial-frequency features from both the real and imaginary components of MR data.More specifically,our DONet consists of a series of Dual-Octave convolutions(Dual-Oct Conv),which are connected in a dense manner for better reuse of features.In each Dual-Oct Conv,the input feature maps and convolutional kernels are first split into two components(i.e.,real and imaginary)and then divided into four groups according to their spatial frequencies.Then,our Dual-Oct Conv conducts intra-group information updating and inter-group information exchange to aggregate the contextual information across different groups.DONet encourages information interaction and fusion between the real and imaginary components at various spatial frequencies to achieve richer rep-resentational capacity.The dense connections between the real and imaginary groups in each Dual-Oct Conv make the propagation of features more efficient through feature reuse.DONet enlarges the receptive field by learning multiple spatial-frequency features of both the real and imaginary components.Extensive experiments on two popular datasets(i.e.,clinical knee and fast MRI),under different undersampling patterns and acceleration fac-tors,demonstrate the superiority of our model in accelerated parallel MR image recon-struction.(2)Convolutional neural networks are limited in their ability to capture long-distance dependencies due to their inherent locality.To this end,we propose a multi-modal trans-former(MTrans),which is capable of transferring multi-scale features from the target modality to the auxiliary modality for accelerated MR imaging.To capture deep multi-modal information,our MTrans utilizes an improved multi-head attention mechanism called the“cross attention module”,which absorbs features from the auxiliary modal-ity that contribute to the target modality.Our MTrans uses an improved transformer for multi-modal MR imaging,affording more global information compared with existing con-volutional neural network-based methods.A new cross-attention module is proposed to exploit the useful information in each modality at different scales.The small patch in the target modality aims to keep more fine details,while the large patch in the auxiliary modal-ity aims to obtain high-level context features from the larger region and supplement the target modality effectively.We evaluate MTrans with various accelerated multi-modal MR imaging tasks,e.g.,MR image reconstruction and super-resolution,where MTrans outperforms state-of-the-art methods on fast MRI and real-world clinical datasets.(3)From the perspective of back-end super-resolution enhancement,under the guid-ance of the corresponding auxiliary modality,we try to provide additional anatomical information to the target modality image for the super-resolution enhancement.How-ever,current multi-contrast super-resolution(SR)methods tend to concatenate different contrasts directly,ignoring their relationships in different clues,e.g.,in the foreground and background.In this study,we propose a separable attention network(comprising a foreground priority attention and a background separation attention),named SANet.Our SANet could explore the areas of foreground and background in the“forward”and“re-verse”directions with the help of the auxiliary contrast while learning clearer anatomical structure and edge information for the SR of a target-contrast MR image.SANet is the first model to explore a separable attention mechanism that uses the auxiliary contrast to predict the foreground and background regions,diverting more attention to refining any uncertain details between these regions and correcting the fine areas in the reconstructed results.A multi-stage integration module is proposed to learn the response of multi-contrast fusion at multiple stages,get the dependency between the fused representations,and boost their representation ability.Extensive experiments with various state-of-the-art multi-contrast SR methods on fast MRI and clinical in vivo datasets demonstrate the superiority of our model.(4)MR image reconstruction and super-resolution are two crucial techniques in MR imaging.Current methods are designed to perform these tasks separately,ignoring the cor-relations between them.In this work,we propose an end-to-end task transformer network(T~2Net)for joint MRI reconstruction and super-resolution,which allows representations and feature transmission to be shared between multiple tasks to achieve higher-quality,super-resolved,and motion-artifact-free images from highly undersampled and degener-ated MRI data.Our framework combines both reconstruction and super-resolution,di-vided into two sub-branches,whose features are expressed as queries and keys.Specifi-cally,we encourage joint feature learning between the two tasks,thereby transferring ac-curate task information.We first use two separate convolutional neural network branches to extract task-specific features.Then,a task transformer module is designed to embed and synthesize the relevance between the two tasks.Experimental results show that our multi-task model significantly outperforms advanced sequential methods,both quantita-tively and qualitatively.