Motivation:
Framework:
![](https://img-blog.csdnimg.cn/direct/dea7b6b5b2374d9aaba713cd17f905d0.png)
![](https://img-blog.csdnimg.cn/direct/be882f4e181c4f53bad4ec2971be2bc5.png)
![](https://img-blog.csdnimg.cn/direct/829bd86c96f944ad83089d896350c5f7.png)
![](https://img-blog.csdnimg.cn/direct/d87211c31d124157a6ba826080029e73.png)
![](https://img-blog.csdnimg.cn/direct/8a3ab4ec2228473cb227c78612da97f2.png)
![](https://img-blog.csdnimg.cn/direct/7d999972488c4e4ebd5f5d8c6168a184.png)
Experiment:
Limitation and future work
The proposed MHAN has some limitations. First, in practice, this multi-task network may need to deal with different masks and scaling factors than in the pre-training. Due to the limited memory of MR Scanners, it is not cost-effective to train a model for every case as in the previous section. Second, our method is explored based on single-coil data, which is still a certain distance from practical applications since commercial MR Scanner usually acquires multi-coil data. Third,
the medical data captured by MR Scanner is in 3D format. However, our model mainly targets the reconstruction of 2D images, which inevitably loses the spatial correlation information. In the future, we will devote ourselves to improving the generalization ability of our model and achieving reconstruction of arbitrary under-sampled patterns and SR of continuous scaling factors by fine-
tuning some hyperparameters. In addition, a typical MRI protocol comprises a set of pulse sequences to provide multi-contrast images (e.g., T1, T2, PD, FSPD), which have inter-modality consistent and modality-specific information. It is a worthy try to restore the target modality from its low-resolution under-sampled counterpart with the guidance of the auxiliary modality.