Robustness to Domain Shifts in MRI for Deep Learning-based Methods: A Review
Summary
Deep learning-based approaches have seen a lot of success in the space of Magnetic Resonance Imaging from segmentation to real-time registration. However, these methods often fail to generalize to different domains, that is to say scans from different scanners, sequences or populations. We review two classes of approaches to counteract these so-called domain shifts in deep learning for MRI data: data harmonization and domain generalization, and discuss their pros and cons. Data harmonization removes domain-specific information from MR images themselves, whereas domain generalization trains a neural network to be robust to a wide variety of input images from different domains. Five papers were found for data harmonization and sixteen papers for domain generalization. Based on these papers, we conclude that both data harmonization and domain generalization are viable for small expected domain shifts. In practice, the extent of domain shifts will often be unknown prior to deployment of the neural network or will be too large. In this more realistic case, we determine that domain generalization offers better generalization capabilities than data harmonization. On the other hand, data harmonization can be used to remove domain-specific information from new data, making it possible to use already trained task networks (e.g. a segmentation network) without having to retrain on the new domain or requiring labelled data from this domain. Both methods therefore have useful applications in different scenarios.