Magnetic resonance imaging reconstruction based on geometric distillation and feature adaptation

Although existing deep learning-based compressed sensing magnetic resonance imaging methods have achieved good results, the interpretability of these methods is still challenging, and the transition from theoretical analysis to network design is not natural enough. To solve the above problems, this paper proposes a deep dual-domain geometry distillation feature adaptive network (DDGD-FANet). The deep unfolded network iteratively unfolds the magnetic resonance imaging reconstruction optimization problem into three sub-parts: data consistency module, dual-domain geometry distillation module and adaptive network module. It can not only compensate for the lost contextual information of the reconstructed image and restore more texture details, but also remove global artifacts to further improve the reconstruction effect. Experiments were conducted using three different sampling modes on public datasets. The results show that DDGD-FANet achieved higher peak signal-to-noise ratio and structural similarity index under the three sampling modes. Under the Cartesian CS ratio, the peak signal-to-noise ratio was improved by 3.34 10%dB and 3.34 dB respectively compared with the three models of ISTA-Net+, FISTA-Net and DGDN.5.01 dB、4.81 dB