Constrained Domain Adaptation for Image Segmentation

Domain Adaption tasks have recently attracted substantial attention in computer vision as they improve the transferability of deep network models from a source to target domain with different characteristics. A large body state-of-the-art domain-adaptation methods was developed for image classification purposes, which may be inadequate segmentation tasks. We propose adapt networks constrained formulation, embeds domain-invariant prior knowledge about regions. Such take form anatomical information, instance, structure size or shape, can known priori learned samples via an auxiliary task. Our general formulation imposes inequality constraints on predictions unlabeled weakly labeled samples, thereby matching implicitly prediction statistics and domains, permitted uncertainty knowledge. Furthermore, our easily integrate weak annotations data, such image-level tags. address ensuing optimization problem differentiable penalties, fully suited conventional stochastic gradient descent approaches. Unlike common two-step adversarial training, is based single network, simplifies adaptation, while improving training quality. Comparison adaptation reveals considerably better performance model two challenging Particularly, it consistently yields gain 1-4% Dice across architectures datasets. results also show robustness imprecision The versatility novel approach readily used various problems, code available publicly.