Multi-contrast MRI Reconstruction Based On Unsupervised Deep Learning In Gradient Domain

In clinical practice,magnetic resonance imaging can generate multiple magnetic resonance images by using different contrast settings,thus providing rich multi-contrast information for clinical diagnosis.However,long scan time increase sensitivity to physical movement,which can reduce the imaging effect.Therefore,as the number of scans increases,there is an urgent need for technologies that can shorten the scanning time to reduce costs and patients’ discomfort without sacrificing image structure information.Compressed sensing can speed up magnetic resonance imaging by reducing the kspace(or Fourier space)measurements obtained directly from the machine.Compressive sensing theory research shows that if the transformed data is sparse,the k-space data with high undersampling can be reconstructed with accurate original data,which greatly reduces the amount of data required for sampling and thus reduces the time spent in scanning.However,if the information obtained is insufficient,the image structure information may be lost or blurred.Therefore,sparse constraints are applied to reconstructed images with additional prior information to improve the quality of image reconstruction.In recent years,with the emergence of deep learning,many researchers have proposed many novel methods for constructing prior information.Therefore,how to effectively use the powerful learning ability of deep learning to extract prior information has become a hot research issue.In this paper,we propose a new method to efficiently reconstruct multi-contrast magnetic resonance Images with similar anatomical structures from these partially-sampled K-space data by using a deep learning unsupervised generation model.The method proposed in this paper consists of two successive stages.In the prior learning stage,that is,the training stage of unsupervised deep learning,the generation model based on fractional estimation is used to train the prior information of the gradient domain from the single contrast image dataset.After the prior information is determined,it is applied to multi-contrast reconstruction under different Settings,such as different number of contrast images and different sampling modes.In the iterative reconstruction process,the data consistency term,gradient graph and group sparsity constraint term are updated iteratively to obtain satisfactory results.Simulation results and experimental results of in vivo magnetic resonance imaging data show that this method can achieve lower reconstruction error and retain clearer image structure than other algorithms.To sum up,this disseratation focuses on the study of unsupervised deep learning for multi-contrast magnetic resonance imaging reconstruction,mining the potential prior information of images and fusing it into the multi-contrast magnetic resonance imaging reconstruction algorithm.The algorithm proposed in this disseratation can effectively provide better fidelity by taking advantage of the structural similarity between different contrast images and the more accurate approximation capability provided by the generation model.