Deep Reinforcement Learning Based Unrolling Network for MRI Reconstruction

Compressed sensing (CS) has been popular in magnetic resonance imaging (MRI) and accelerates the measurement acquisition process by undersampling data in k-space while exploiting sparsity and incoherence to achieve accurate reconstructions. Deep learning methods have recently demonstrated superior performance for MRI reconstruction from undersampled data. However, most of the methods rely on complex models to achieve promising results via deterministic optimization. Alternatively, very recent works utilize deep reinforcement learning (DRL) to restore high-quality images, where a policy is learned to select appropriate actions or tools to progressively refine corrupted images with a simple network model. However, the model capability of DRL-based approaches is, to some extent, limited due to its finite action space. Moreover, most of these DRL methods are physics-free, while it is well-known that the prior concerning the physical forward model is extremely crucial for solving ill-posed inverse problems such as in CS-MRI. Motivated by these challenges, we propose a novel DRL-based unrolling framework by integrating model priors into the intrinsic iterative process of DRL strategy for MRI reconstruction. Thus the capability of the DRL model is significantly enhanced by exploiting the merits of the unrolling scheme with almost no additional computational cost. Extensive experiments demonstrate that our proposed DRL-based unrolling framework achieves superior MRI reconstruction performance compared with the previous baselines.