Synopsis

In this presentation, we will look into machine learning-based reconstruction and observations made on other imaging modalities than MR. In particular, we can sub-divide reconstruction methods into purely data-driven, analytically inspired, and optimization-inspired. We find that also from a theoretical point of view, embedding of domain knowledge is beneficial. During the presentation, we will discuss further the benefits and risks of these common approaches. In the end, we will give an outlook on future perspectives and potential enablers in the field.

Abstract

In this presentation, we will look into machine learning-based reconstruction and observations made on other imaging modalities than MR. In particular, we can sub-divide reconstruction methods into purely data-driven, analytically inspired, and optimization-inspired. An example of a direct data-driven method is AUTOMAP [1] that is also used in MRI. The analytically motivated methods, map traditional one-pass reconstruction approaches onto machine learning methods such as introduced by Würfl et al. [3]. Optimization-based methods follow the ideas by Kobler, Hammernik, and Pock to unroll the optimization program onto a feed-forward network as shown in MRI [3] and CT [4]. Also from a theoretical point of view, the embedding of domain knowledge is beneficial as it is able to reduce maximal error bounds of the training problem [5]. During the presentation, we will discuss further the benefits and risks of these common approaches [6,7]. In the end, we will give an outlook on future perspectives and potential enablers in the field.

Acknowledgements

The research leading to these results has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (ERC Grant No. 810316).

References

[1] Zhu, Bo, et al. "Image reconstruction by domain-transform manifold learning." Nature 555.7697 (2018): 487-492.

[2] Würfl, Tobias, et al. "Deep learning computed tomography: Learning projection-domain weights from image domain in limited angle problems." IEEE transactions on medical imaging37.6 (2018): 1454-1463.

[3] Kobler, Erich, et al. "Variational networks: connecting variational methods and deep learning." German conference on pattern recognition. Springer, Cham, 2017.

[4] Hammernik, Kerstin, et al. "A deep learning architecture for limited-angle computed tomography reconstruction." Bildverarbeitung für die Medizin 2017. Springer Vieweg, Berlin, Heidelberg, 2017. 92-97.

[5] Maier, Andreas K., et al. "Learning with known operators reduces maximum error bounds." Nature machine intelligence 1.8 (2019): 373-380.

[6] Huang, Yixing, et al. "Some investigations on robustness of deep learning in limited angle tomography." International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2018.

[7] Gottschling, Nina M., et al. "The troublesome kernel: why deep learning for inverse problems is typically unstable." arXiv preprint arXiv:2001.01258 (2020).