Recent advances in machine learning have led to the introduction of accurate algorithms for several medical applications such as registration, image classification, and segmentation. However, despite the high performance of those methodologies (mostly on controlled conditions ), most of them struggle to generalize to unseen scenarios (e.g data coming from different scanners, acquisition parameters, modality, or even demographic differences). This is a critical problem that limits their applicability and safe implementation in clinical practice and therefore reduces their impact.
Domain adaptation aims to eliminate this lack of generalization and mitigate the underperformance of methods when faced with data coming from a different domain (uncontrolled conditions).
Domain adaptation can be supervised by taking a few labeled samples from the data we wanted to adapt (target domain), and using them to learn relevant information. This solution still limits its practical use as it requires getting manually annotated samples which could be not always available.
On the other hand, unsupervised domain adaptation only requires a set of unlabeled samples from the target domain. Although the performance could be a bit lower compared to the supervised ones, getting unlabeled data is more feasible which makes unsupervised methods more applicable in clinical practice.
Orbes-Arteaga, M., Varsavsky, T., Sudre, C.H., Eaton-Rosen, Z., Haddow, L.J., Sørensen, L., Nielsen, M., Pai, A., Ourselin, S., Modat, M. and Nachev, P., 2019. In Domain Adaptation and Representation Transfer and Medical Image Learning with Less Labels and Imperfect Data (pp. 54-62). Springer, Cham.