Research On MRI Image Reconstruction Method Based On Convolutional Neural Network

Magnetic resonance imaging(MRI)technology is widely used in clinical examination due to its advantages of non-ionizing radiation,high resolution of soft tissue and multidirectional imaging,but its disadvantage is that the imaging speed is slow.The multi-coil parallel MRI technology improves the data scanning speed,and the compressed sensing MRI reconstruction improves the image reconstruction quality through a large number of iterative operations.The combination of the two technologies widens the application range of MRI in clinical practice.However,the reconstruction of compressed sensing requires a long reconstruction time to get a good image,and when the acceleration factor is too large,the SNR of the reconstructed image is low.Deep MRI changes the traditional image reconstruction framework,and through the study of a large number of data samples,a good network model is obtained,so as to quickly and accurately predict the new data.We studied the MRI image reconstruction method based on deep convolutional neural network,analyzed and compared the advantages and disadvantages of different network models for MRI reconstruction,and optimized the performance of the depth model.Firstly,the method of deep MRI reconstruction based on single coil is studied.Through the comparative analysis of ADMM-net,spatial U-net and frequency U-net models,the W-net model proposed by Frayne et al was optimized by using the method of transfer learning,and the optimized model was added with data update module.The improved W-net1 model was proposed and anti-noise training was carried out.Then,the method of deep MRI reconstruction based on multi-coil was studied.By analyzing the application of the CNNDC-net model based on the real convolution kernel to the parallel MRI reconstruction,the Complex-CNNDC model based on the complex convolution kernel is derived.The real convolution kernel in CNNDC-net is all replaced by complex convolution kernel to learn more original features in multi-channel MRI data.Finally,the data set was processed by undersampling,and then divided into training set,verification set and test set,and the training of various models was realized in Keras platform.The test results showed that after the transfer learning of the W-net model,the model significantly improved the generalization ability of MRI data such as the brain containing tumor lesions and the knee with relatively simple structure.Moreover,after adding data update layer into the model,the improved model reconstruction is of better quality,and the texture details of lesions and cerebellum are recovered better.Compared with other models and traditional MRI reconstruction algorithm,the improved W-net1 model has the best performance.Compared with the traditional algorithm,the W-net1 model after anti-noise training can reconstruct good MRI images from the test data with different proportion of noise.Therefore,the optimized W-net1 model not only has better performance,but also has stronger anti-noise capability.For the CNNDC-net model,compared with U-net and the traditional parallel reconstruction algorithms,the test results have the highest value of the peak signal-to-noise ratio and the lowest value of the root mean square error.The Complex-CNNDC model with complex convolution kernel can restore the texture details of cerebellum better.By applying the depth model to the single-coil and multi-coil MRI data acquisition modes,we achieved rapid and accurate reconstruction of undersampled MRI data.It is of great significance for the acceleration of clinical MRI scanning and the development of depth imaging system.