Deep Network Cascade for Dynamic Cardiac MRI Reconstruction With Motion Feature Incorporation and the Fourier Neural Attention

Magnetic resonance imaging (MRI) provides a radiation-free and non-invasive tool for clinical diagnosis. However, it suffers from a prohibitively long acquisition process for many applications. Compressed sensing (CS) methods have been used for reconstruction from under-sampled data in accelerated acquisitions. Although effective in practice, the image quality can be limited by the expressiveness of handcrafted signal priors such as sparsity. Dynamic MRI requires high spatial and temporal resolution, which makes CS to be more difficult to recover the data taken within a short scanning time. In this paper, we explore to solve the challenging inverse problem by introducing an optimization-inspired deep leaning framework to recover dynamic MRI images. A novel mask-guided motion feature incorporation (Mask-MFI) scheme is proposed to benefit the recovery of the dynamic content, and a spatio-temporal Fourier neural block (ST-FNB) is designed to improve the reconstruction performance by leveraging the redundancies in spatial and temporal domains in a computation and parameter efficient manner. The comparative experiments demonstrate that the proposed framework outperforms other state-of-the-art methods at a range of accelerations both qualitatively and quantitatively. Ablation studies confirm the effectiveness of model components. Moreover, the adaptability and generalization capacity of the introduced method are also validated, which demonstrates the potential of the application of our proposed approach to other reconstruction models to boost their performance.