Research On The Reconstruction Algorithms Of Undersampling MR Image Based On Deep Learning

Magnetic resonance imaging(MRI)has become one of the commonly used auxiliary diagnosis and treatment methods in the clinic because of its non-radiation and high soft tissue contrast characteristics.The acquisition of MRI data is sequential,so more data requires longer scanning time.In order to speed up magnetic resonance imaging,the common method is to collect a small amount of data.However,traditional MR image reconstruction methods has been not suitable for fast MRI.How to reconstruct the high quality MR image which meet the clinical requirement from undersampling data is main consideration of paper.Fast MRI generally uses image reconstruction algorithm based on compressed sensing.This algorithm has certain limitation in capturing image details,as a result the reconstructed image quality is not high.Deep learning has capability of powerful feature extraction,which can overcome the shortcomings of traditional reconstruction methods.However,most of deep learning networks use single-domain reconstruction methods and then ignore the powerful feature learning capability in dual-domain cascades.On the consideration of information loss in the transmission process of traditional networks,firstly,this paper proposes the following MRI image reconstruction models by analyzing the correlation between slices:1)A multi-scale dense nested network(MDNN)model is proposed.MDNN uses the dense connection network and the multi-scale ideology to reconstruct the highquality magnetic resonance images.Firstly,based on the dense connection network,the combination with different dilation rates are placed on the dense connection units;Secondly,transmission channel is found based on the prior information between slices;Finally,the fidelity for K-space data is designed.In the model,dense connection mechanism in network iteration layer is added,which can avoid information loss.2)A cross-domain deep dense dual-domain cascade network(CD-CN)model is proposed.Firstly,the symmetrical multi-scale feature fusion(SMFF)module is designed for extracting the feature information of different scales in wavelet domain;Secondly,the bidirectional multi-channel feature pyramid(BMFP)structure is proposed to strengthen the feature transmission between slices;Finally,the two modules are cross-cascaded.In the model,the same domain is densely connected in the iterative process,which can strengthen the feature fusion between the same domain.Secondly,the two models are test on the Calgary-Campinas and fast MRI datasets.For 12.5% K-space data on Calgary-Campinas dataset,compared with other algorithms,when using MDNN model,the results show that the SSIM of reconstructed images improves 0.96%-3.87%,and the PSNR improves 0.66 d B-2.88 d B;when using CD-CN model,the SSIM of reconstructed images improves 0.86%-2.48%,and the PSNR improves 0.61 d B-2.83 d B.For 12.5% K-space data on fast MRI dataset,compared with other algorithms,the SSIM of reconstructed images using CD-CN improves 1.61%-2.34%,and the PSNR improves 0.85 d B-1.8d B.The experimental results show that the two reconstruction methods proposed in this paper can significantly increase the reconstruction accuracy of MRI images.Thus,the proposed MDNN and CD-CN have advantages in MR image reconstruction tasks.