Research On Magnetic Resonance Parameter Mapping Combining Physical Model And Deep Learning

Magnetic resonance imaging(MRI)is a multi-contrast imaging modality without radiation,which has been widely applied in the diagnosis and assessment of disease.Magnetic resonance(MR)parameter mapping is a quantitative MRI technique,which can reveal the spatial distribution of the biophysical parameters in tissue,such as transverse relaxation time(T2)and effective transverse relaxation time(T2*),and provide valuable quantitative information regarding the diagnosis of diseases.Rapid and accurate reconstruction of parameter maps is vital for the clinical promotion and application of MR parameter mapping.Conventional model-based MR parameter mapping algorithms are interpretable,stable,and generalizable.However,these methods commonly use handcrafted and sophisticated priors to construct optimization problems,and the iterative optimization process is time-consuming.The rise of deep learning in the recent years has shown great potential for MR parameter mapping.Deep learning-based MR parameter mapping methods are efficient and high-performance.However,the supervised training of deep neural network requires ground-truth parameter maps.Because of the effect of noise and other factors,the ground-truth or high-quality parameter maps are probably inaccessible.Besides,general deep neural networks are typically difficult to interpret and lack generalization ability to MR data obtained by using different acquisition protocols.In order to make full use of the advantages of the above two kinds of methods,this study investigates the MR parameter mapping methods that combine physical model and deep neural network.The contents of this dissertation mainly include the following aspects:(1)A model-guided self-supervised deep network has been proposed for(=1/T2*)parameter mapping of iron-loaded liver.The proposed network employs a noise correction model and an improved total variation model to construct a loss function that are used to train the network.The self-supervised training of the network only needs clinical T2*-weighted images and does not require ground-truth R2*parameter maps.Meanwhile,the proposed method can correct the bias introduced by the noise and enable accurate and efficient R2*mapping of iron-loaded liver.(2)A model-based MR R2(=1/T2)parameter mapping network has been proposed.The proposed method combines physical model-based data consistency term and deep learning-based regularization term to construct the optimization problem.The alternating direction method of multipliers is used to solve the optimization problem and then unrolled to construct the proposed network.The architecture of the network is based on conventional iterative optimization algorithm and has better interpretability.The variation in acquisition parameters can be addressed by the data fidelity component of the network,which improves the generalizability of the network.The proposed method was tested on R2 mapping of brain.The results showed that the proposed network can achieve rapid and accurate R2 mapping even when the settings of echo time were substantially different from that of training data.