Synopsis

Deep learning can learn expressive feature extractors from data and allows to solve tasks that are challenging to handle with fixed, hand-crafted models. In MR image reconstruction, we aim at recovering an image from the acquired k-space data which is corrupted by measurement noise. In accelerated imaging, the k-space is undersampled leading to an ill-posed problem. To solve it, the reconstruction is regularized. These regularizations can be learned by neural networks. Reconstructions differ in their input and targeted application. Image enhancement, direct mapping, physics-based unrolling and distribution-based methods for supervised learning will be presented and their advantages and disadvantages discussed.

Summary of main findings

In supervised deep learning for MR image reconstruction, regularization terms are learned and approximated by neural networks. Image enhancement, direct mapping, physics-based unrolling and distribution-based methods will be presented and discussed.

Acknowledgements

No acknowledgement found.

References

No reference found.