MRI Reconstruction Algorithm Based On Dual-domain Encode-decode Network

Magnetic resonance imaging(MRI)has the characteristics of non-ionizing-radiation,non-radiation,and high soft tissue contrast,and has significant advantages compared with other imaging methods in the diagnosis of many diseases.The MRI equipment fills a column(row)of K-domain data through phase encoding every time until complete K-domain data is obtained by scanning,and then performs inverse Fourier transform on it to obtain the MRI image.However,the time of collecting complete Kdomain data and the slow imaging speed are bottlenecks in the development of MRI technology.Collecting a part of the K-domain data at random or at equal intervals is an important way to reduce the acquisition time of MRI equipment.Because the undersampled data does not satisfy the Nyquist theorem,MRI images will produce serious artifacts.Reconstruction of encoder-decoder networks for under-sampled data is a mainstream method to reduce image artifacts and improve image quality.At present,the under-sampled reconstruction algorithms based on encoder-decoder networks have problems such as difficulty in extracting effective features from under-sampled data,and serious loss of phase encoding information of K-domain data.In order to solve the above problems,two magnetic resonance image reconstruction algorithms with spacefrequency dual-domain feature extraction as the core are proposed for the cases of Kdomain under-sampling data as input and dual-domain under-sampled data as input.The main work,innovations and contributions are as follows:(1)Aiming at the situation that K-domain under-sampled data is used as the network input,an MRI reconstruction algorithm based on space-frequency dual-domain serial alternating convolution in the network is proposed.The algorithm builds a multiscale image supervision module.At the same time,the algorithm takes advantage of the serial promotion of spatial and frequency domain features.It solves the problem of difficult extraction of effective features of under-sampled data and serious loss of K-domain phase encoding information,and achieves high-quality under-sampled magnetic resonance image reconstruction.Experimental results on FastMRI,a large magnetic resonance image database released by Facebook and New York University,show the effectiveness of the proposed method.(2)In order to further improve the reconstruction quality,a parallel MRI reconstruction algorithm based on space-frequency dual-domain encoder-decoder network is proposed.The algorithm takes K-domain under-sampling data and spatial-domain under-sampled data as the input of the network.Then the algorithm adopts a dual-domain network cascaded image reconstruction architecture,which realizes the parallel and deep coupling of spatial and frequency domain features,and exerts the improvement effect of frequency domain features on spatial features.It solves the problem of insufficient utilization of dual-domain features.The experiment results in the FastMRI knee single-coil and multi-coil cases show that the proposed method significantly improves the image reconstruction quality.In summary,the proposed method plays the complementary role of the space-frequency dual-domain features in the encoder-decoder reconstruction network and its cascade network,and improves the quality of MRI image reconstruction.