Reducing Domain Gap in Frequency and Spatial Domain for Cross-Modality Domain Adaptation on Medical Image Segmentation

Unsupervised domain adaptation (UDA) aims to learn a model trained on source and performs well unlabeled target domain. In medical image segmentation field, most existing UDA methods depend adversarial learning address the gap between different modalities, which is ineffective due its complicated training process. this paper, we propose simple yet effective method based frequency spatial transfer under multi-teacher distillation framework. domain, first introduce non-subsampled contourlet transform for identifying domain-invariant domain-variant components (DIFs DVFs), then keep DIFs unchanged while replacing DVFs of images with that narrow gap. batch momentum update-based histogram matching strategy reduce style bias. Experiments two commonly used cross-modality datasets show our proposed achieves superior performance compared state-of-the-art methods.