Research On MRI Reconstruction Algorithm Based On Deep Learning

Magnetic Resonance Imaging(MRI)is widely used in the fields of treatment detection and medical diagnosis,and there is no ionizing radiation hazard.However,the long scan time has become the main disadvantage of MRI.Therefore,in order to improve the imaging speed while ensuring the quality of MRI,this paper studies the use of deep learning network models to design effective MRI network models,which can improve the reconstruction performance of MR images.The main research contents of this paper are as follows:(1)The original brain data of Calgary-Campinas was used and the data set was normalized.1000 slices of 25 subjects were selected as the training data set,and 400 slices of 10 subjects were used as the test data set.First of all,based on the good performance of the Google Le Net model in the classification task,this paper improves the Google LeNet model,and proposes the S-GLe Net model to be used in the reconstruction of MR images.Secondly,the deep cascading network model uses the data consistency layer for the reconstruction of MRI images for the first time,which image reconstruction quality is good.Therefore,this paper proposes to combine the S-GLe Net model with the data consistency layer as the Cascaded S-GLe Net model.The simulation experiment results show that,in a variety of under-sampling modes,the newly proposed Cascaded SGLe Net model has a great improvement in the reconstruction of MR images compared with the deep cascaded network model and Hybrid-Cascade network model.(2)Since the U-Net neural network performs well in the field of medical image segmentation,this paper proposes to use the cascaded U-Net model in the reconstruction of MR images.The simulation experiment results show that in a variety of under-sampling modes,the cascaded U-Net model can effectively improve the quality of MR image reconstruction.However,there are a large number of parameters need to be learned,and there is excessive blur in reconstructed image.There is still much room for improvement.(3)The new network model proposed in this paper improves the original U-Net model and combines the advantages of the Google Le Net model and the Res Net model,called the UGR-Net model.Then,on the basis of the UGR-Net model,using a variety of ways to deepen the model depth at the same time,the up-sampling method and activation function are changed,the PCK-Net model is proposed.In addition,the two models of UGR-Net and PCK-Net are combined with the data consistency layer,which is called Cascaded UGR-Net model and Cascaded PCK-Net model.Multiple under-sampling modes are used to under-sampling the complex data of the brain,and the reconstruction performance of the newly proposed model is tested.The simulation experiment results show that in a variety of under-sampling modes,the Cascaded UGR-Net and the Cascaded PCK-Net are superior to the compared reconstruction models in terms of visual and quantitative indicators.Especially for the cascaded PCK-Net model,in the case where the sampling mode of the training data set and the test data set are different,the reconstruction effect is still very good.