CMCD-Net:Unsupervised domain adaptation with Contrastive learning for cross-modality and cross-disease brain lesion segmentation

Unsupervised domain adaptation (UDA), as a robust transfer learning strategy that utilizes source domain richly labeled data to solve the target domain unlabeled semantic segmentation, has great potential to tackle the problem of data scarcity and domain shift. However, in the medical domain, most existing methods are limited to image-level constrained alignment by image-to-image translation, and few can simultaneously address the dual domain shift of both modality and lesion discrepancies. The motivation of this paper is to explore the potential of UDA for addressing cross-modal and cross-lesion brain lesion segmentation without generative training. The specific novelty of our proposal mainly includes: (1) proposing a UDA technology for cross-modality and cross-disease brain imaging segmentation, which are more general and urgent in clinical practice; (2) exploring the potential of Fourier-based bi-directional frequency domain adaptation to alleviate imaging protocol discrepancies as an alternative to burdensome generative training; (3) proposing a self-training strategy with category-level prototype contrastive learning to exploit cross-domain similar semantic features. The proposed CMCD-Net exhibited encouraging adaptation performance in two segmentation tasks: multi-modality MRI gliomas, multi-modality MRI glioma to meningioma. In addition, massive ablation experiments validated the effectiveness of each component in our proposal. The code is publicly available at https://github.com/Snow1949/C2_MUDA.