Deep maximum a posterior estimator for accelerated MRI reconstruction

Deep learning methodologies have demonstrated their efficacy in addressing MRI reconstruction from a collection of under-sampled measurements captured by multiple coils and yield reconstructions of exceptional quality. Since an acquired MRI volume is composed of many MRI slices, leveraging the correlation information between adjacent slices to assist in the reconstruction of the current slice is an intuitive idea. In this study, based on the maximum a posterior (MAP) estimation framework, we propose a novel MAP-based MRI reconstruction method called MAP-MRINet with adaptive spatial fusion and deep image prior. Specifically, we develop a novel observation model that considers the correlations among adjacent under-sampled slices, enabling integration of slice alignment and image reconstruction. Additionally, we split the image prior into high-frequency (HF) and low-frequency (LF) components based on the Framelet transform to provide more dedicated constraints. We unroll the proposed iterative MAP-based MRI reconstruction algorithm into a deep convolutional network. We conduct experiments on two publicly available real-world MRI datasets at various acceleration rates, and both qualitative and quantitative comparisons outperform current state-of-the-art single-slice and multi-slice reconstruction techniques. The experimental results validate the effectiveness and practicality of the proposed model.