Abstract：In recent years,since deep learning has been prominent in image processing,researchers have begun to use deep learning in the reconstruction of magnetic resonance imaging(MRI).Existing deep learning models need to rely on large amounts of data for training,and it is difficult to obtain large amounts of data from medical images.Therefore,in order to effectively improve the quality of MR imaging reconstruction,this paper proposes a deep learning MRI reconstruction method that performs well on small data sets.This paper improves the U-Net model,combines the advantages of GoogleLeNet and ResNet,proposes the UGR-Net model,and combines it with the data consistency layer to obtain a cascaded UGR-Net.Under multiple acceleration factors,three under-sampling modes are used to under-sample complex brain data,and the reconstruction performance of the cascaded CNN network,cascaded U-Net model,and newly proposed cascaded UGR-Net reconstruction model make a comparison.The experimental results show that the cascaded UGR-Net is superior to the cascaded CNN model and the cascaded U-Net model in both visual and quantitative indicators.
2022, Vol. 43 
