Contrastive image adaptation for acquisition shift reduction in medical imaging

Clément Hognon, Pierre-Henri Conze, Vincent Bourbonne, Olivier Gallinato, Thierry Colin, Vincent Jaouen, Dimitris Visvikis

Abstract: Highlights•Acquisition shift is responsible for lack of robustness of deep models to scanner and protocol changes.•We propose a deep image adaptor to mitigate the acquisition shift, named contrastive image adaptation.•Structural preservation is enforced with a gradient loss while translation is guided by GAN and contrastive loss.•The method requires as little as one 2D image in the target domain.•Experiments show positive impact on segmentation and image to image translation.