Deep Learning-based Invertible Coil Compression For MRI

Magnetic Resonance Imaging(MRI)is a versatile and noninvasive technique for medical imaging.However,the long scanning time due to the large amount of data required in the MRI process has limited the further development.In order to improve the imaging speed,many rapid imaging methods are proposed,and parallel imaging is one of the most effective methods.However,with the receiving coil channels increasing,huge data storage and complex mathematical calculation have become an urgent problem to be solved.Coil compression technology can alleviate this problem very well.The coil compression method is to compress the data of multiple channels into fewer virtual coils,thereby eliminating data redundancy caused by excessive coils.However,the current research regarding coil compression is based on the traditional algorithm.The disadvantage of this algorithm is that the compression time is long and the calculation amount is large,especially for large datasets.In order to solve these problems,this paper proposes a reversible magnetic resonance coil compression technology based on deep learning.This study takes advantage of deep learning networks to combine deep learning with invertible coil compression for magnetic resonance imaging.In this study,the encoding-decoding network is selected as the medium for data compression.The output of encoding subnetwork is the compressed coils and the decoding subnetwork is further designed to get the invertible coils.The main advantage of this paper is that three optimization loss functions(invertible loss,compressibility loss and SOS retention loss)are proposed for training to better retain the edge information of the image and realize reversible compression.In addition,the complex images are preprocessed,and stack the real part and the virtual part according to the column as the input of network to make better use of data information in this work.In this study,not only lossless compression can be achieved,but more importantly,we have achieved invertible recovery,and the recovery result exceeds 40 d B,which is basically completely recoverable.This paper has done a lot of simulation experiments,and compared with the current more classical coil compression algorithms.All experimental results show that the proposed method is superior to the classical coil compression algorithm in terms of vision and performance.