High-Fidelity MRI Reconstruction with the Densely Connected Network Cascade and Feature Residual Data Consistency Priors

Since its advent in the last century, magnetic resonance imaging (MRI) provides a radiation-free diagnosis tool and has revolutionized medical imaging. Compressed sensing (CS) methods leverage the sparsity prior of signals to reconstruct clean images from under-sampled measurements and accelerate the acquisition process. However, it is challenging to reduce strong aliasing artifacts caused by under-sampling and produce high-quality reconstructions with fine details. In this paper, we propose a novel GAN-based framework to recover the under-sampled images, which is characterized by a novel data consistency block and a densely connected network cascade used to improve the model performance in visual inspection and evaluation metrics. The role of each proposed block has been challenged in the ablation study, in terms of reconstruction quality metrics, using texture-rich FastMRI Knee image dataset.