| Magnetic Resonance Image(MRI)is one of the important auxiliary diagnostic methods in clinical medicine.It can not only visualize the tissues and organs of interest,but also has advantages that other medical imaging methods do not have,such as no radioactivity,multiparameter acquisition,and high degree of imaging freedom.However,doctors often change the parameters to obtain multiple images with different contrasts to obtain richer details.Since the images with different contrasts are not collected at the same time,acquisition time will increase with the number of contrast images.The long acquisition time is an important factor hindering the development of magnetic resonance imaging.Therefore,it is very important to accelerate the acquisition of magnetic resonance images and to study effective multi-contrast joint reconstruction algorithms,and it is also a hot topic in the field of magnetic resonance imaging.Under-sampling can directly reduce the time for data collection.When the under-sampling rate does not meet the Nyquist sampling criterion,the directly reconstructed image will have serious artifacts.Thereby,efficient reconstruction algorithms are required to obtain images with high signal-to-noise ratio by exploiting a series of prior information.The traditional optimization method established the objective function by looking for effective prior constraints,and then iteratively solved the objective function through numerical algorithms.The commonly used prior constraints in traditional multi-contrast MR image reconstruction optimization methods include joint sparse constraints,joint gradient constraints,etc.However,neither the predefined sparse domain nor the sparse domain learned through dictionary adaptation is too simple to deal with the complex tissue and organ images.In recent years,researchers have tried to use deep learning methods to complete the reconstruction of multi-contrast magnetic resonance images,but the existing deep learning-based methods are all data-driven and do not fully utilize the information of the imaging system and compress the learning space of network,resulting in a large learning space and poor interpretability of the overall network.This thesis proposes multi-contrast image reconstruction method based on statistical properties and joint sparsity and joint group sparsity-based deep learning for multi-contrast MRI reconstruction.The main work includes: 1)Reviewing the main methods of deep learningbased magnetic resonance image reconstruction;2)Proposing a multi-contrast image reconstruction method based on statistical properties and joint sparsity,and completed the derivation of the algorithm and the proof of the convergence conditions.The experimental results show that the proposed method improves 4% compared with joint sparse reconstruction in terms of peak signal-to-noise ratio;3)By combining the advantages of traditional optimization methods and deep learning,joint group sparsity-based deep learning for multicontrast MRI reconstruction is proposed.The network is expanded according to the traditional iterative formula.Compared with the data-driven deep network that needs to learn the mapping from under-sampling to full-sampling,the proposed network model only needs to learn the transformation process of the joint sparse domain.Experimental results show that the proposed method outperforms traditional optimization methods and data-driven deep learning methods,and is robust to non-aligned environments. |
