Learn to Adapt

Illustrative figure by Shadi Albarqouni

To build domain-agnostic models that are generalizable to a different domain, i.e., scanners, we have investigated three directions; First, Style Transfer, where the style/color of the source domain is transferred to match the target one. Such style transfer is performed in the high-dimensional image space using adversarial learning, as shown in our papers on Histology Imaging (Lahiani et al. 2019a, Lahiani et al. 2019b, Shaban et al. 2019). Second, Domain Adaptation, where the distance between the features of the source and target domains are minimized. Such distance can be optimized in a supervised fashion, i.e., class aware, using angular cosine distance as shown in our paper on MS Lesion Segmentation in MR Imaging (Baur et al. 2017), or in an unsupervised way, i.e., class agnostic, using adversarial learning as explained in our article on Left atrium Segmentation in Ultrasound Imaging (Degel et al. 2018). Yet, another exciting direction that has been recently investigated in our paper (Lahiani et al. 2019c) is to disentangle the feature that is responsible for the style and color from the one responsible for the semantics.

Baur et al. 2017, Degel et al. 2018, and Shaban et al. 2019

Lahiani et al. 2019c


  • Eldad Klaiman, Roche Diagnostics GmbH
  • Georg Schummers and Matthias Friedrichs, TOMTEC Imaging Systems GmbH
Shadi Albarqouni
Shadi Albarqouni
Professor of Computational Medical Imaging Research at University of Bonn | AI Young Investigator Group Leader at Helmholtz AI