Revisiting Fourier Domain Adaptation for Medical Image Segmentation

Domain adaptation plays a critical role in medical image segmentation, addressing variations stemming from diverse machine types and specifications. In this paper, we propose a new Fourier Style Transfer (FST)-based approach tailored specifically for this purpose. Our method begins by employing similarity measurement to select an appropriate target image, thereby mitigating domain gaps. Subsequently, we integrate a segmentation label matching strategy to emphasize significant regions within the target image, effectively diminishing the influence of noisy features. Leveraging a source-free domain adaptation architecture that incorporates entropy minimization and class-ratio priority, our framework demonstrates robust performance in challenging medical segmentation tasks. Simulation outcomes underscore the efficacy of our approach over existing methods, as evidenced by superior results achieved on two widely used medical datasets.