Research On Non-Cartesian Magnetic Resonance Imaging Based On Cascaded Convolutional Neural Network

Magnetic resonance imaging is currently one of the best medical imaging modalities,with advantages such as non-ionizing radiation and high soft tissue contrast.Therefore,it plays an important role in clinical testing of the brain,spine,and knee,among others.However,the data acquisition time for Magnetic resonance imaging is long,and this slow acquisition can cause discomfort for patients and lead to high examination costs.Therefore,fast Magnetic resonance imaging is a major scientific goal in the field of Magnetic resonance imaging.In accelerated Magnetic resonance imaging research,nonCartesian sampling has an absolute advantage and approaches real imaging scenarios.However,the data sampled by non-Cartesian sampling is non-uniformly distributed,and cannot be directly Fourier reconstructed,which leads to complex reconstruction algorithms and low imaging efficiency.In recent years,deep learning-based methods have been widely used and have shown significant advantages over traditional algorithms in accelerated imaging,but there has been less research on non-Cartesian sampling.To address the slow speed of traditional non-Cartesian Magnetic resonance imaging,thesis proposes to apply a fully data-driven imaging method to non-Cartesian sampling conditions.Experimental results show that this method can improve the structural similarity and speed of reconstruction images compared to traditional algorithms.However,the model relies entirely on the dataset during training,and its stability is not strong,while its generalization ability needs to be improved.To address the problem that existing methods cannot simultaneously balance imaging quality and imaging speed,thesis proposes a cascade network,which is a modelbased deep learning network(MoDL)and a non-Cartesian extension of the cross-domain network.The cascade network relies on a priori and is more stable than a fully data-driven imaging-based approach.Meanwhile,in order to improve the high-frequency information extraction capability of the cascade network,thesis proposes to introduce convolutional denoising self-encoder in the Image-net of the cascade network.The experimental results show that this cascade network is effective for both non-Cartesian sampling,radial and spiral,and its relevant image evaluation indexes are better than the existing methods.Also,by comparing the imaging performance of each method under different acceleration factors,it is shown that the present method is still effective for high acceleration factors and can improve the imaging performance.In summary,thesis combines the advantages of model-driven and data-driven methods,and then proposes a cascaded network to simultaneously improve the quality and speed of non-Cartesian Magnetic resonance imaging.Competitive results were achieved under radial and spiral non-Cartesian sampling.This plays a role in promoting the application of non-Cartesian sampling technology in accelerated Magnetic resonance imaging and can provide technical reference for the clinical implementation of fast Magnetic resonance imaging.