Magnetic resonance imaging (MRI) is a vital medical imaging modality, but its
development has been limited by prolonged scanning time. Deep learning
(DL)-based methods, which build neural networks to reconstruct MR images from
undersampled raw data, can reliably address this problem. Among these methods,
model-driven DL methods incorporate different prior knowledge into deep
networks, thereby narrowing the solution space and achieving better results.
However, the complementarity among different prior knowledge has not been
thoroughly explored. Most of the existing model-driven networks simply stack
unrolled cascades to mimic iterative solution steps, which are inefficient and
their performances are suboptimal. To optimize the conventional network
structure, we propose a collaborative model-driven network. In the network,
each unrolled cascade comprised three parts: model-driven subnetworks,
attention modules, and correction modules. The attention modules can learn to
enhance the areas of expertise for each subnetwork, and the correction modules
can compensate for new errors caused by the attention modules. The optimized
intermediate results are fed into the next cascade for better convergence.
Experimental results on multiple sequences showed significant improvements in
the final results without additional computational complexity. Moreover, the
proposed model-driven network design strategy can be easily applied to other
model-driven methods to improve their performances.

通过深度学习和模型驱动网络，使用注意力模块和修正模块来实现磁共振图像重建，并在多个序列上取得显著改善的结果，同时减少了计算复杂度。