Wavelet-based spectrum transfer with collaborative learning for unsupervised bidirectional cross-modality domain adaptation on medical image segmentation

Unsupervised cross-modality domain adaptation for medical image segmentation has made great progress with the development of adversarial learning-based methods, but the training of adversarial models is considerably complicated. In this paper, we propose a conceptually simple but effective unsupervised domain adaptation method to achieve adaptation on frequency spectrum components of target and source images decomposed by a novel spectrum transfer strategy. Specifically, we replace the high-frequency components of the source domain images with that of the target domain images for details feature adaptation and adjust the low-frequency components by histogram matching for style adaptation. Besides, we propose multi-direction collaborative learning on both target and source domains to further improve the performance. Experimental results demonstrate that our method significantly outperforms state-of-the-art UDA methods for medical image segmentation on two publicly available datasets (cardiac dataset, and abdominal multi-organ dataset) in both CT to MRI and MRI to CT domain adaptation scenarios