Source-Free Domain Adaptation for Medical Image Segmentation via Selectively Updated Mean Teacher

Automated medical image segmentation is valuable for disease diagnosis and prognosis, and it has achieved promising performance with deep neural networks. However, a segmentation model trained on a source dataset may not perform well on a different target dataset when the distribution shift or even modality alteration exists between them. To address this problem, domain adaptation techniques can be applied to train the model with the help of the unannotated target dataset. Often when the target data is available, only a segmentation model trained on the source dataset is provided without the source data, and in this case, source-free domain adaptation (SFDA) is needed. In this work, we focus on the development of SFDA techniques for medical image segmentation, where the given source model is updated based on the target data. Since no annotations are available for the target dataset, we propose to leverage the consistency of predictions on the target data when different perturbations are made, and adopt the mean teacher framework that can effectively exploit the consistency. Moreover, we assume that the update of the entire model in vanilla mean teacher is suboptimal because when no annotated data is available the knowledge learned for segmentation in the source model can be easily forgotten. Therefore, we propose selectively updated mean teacher (SUMT), which seeks to adapt the source model parameters that are sensitive to domain variance and retain the parameters that are invariant to domains. In SUMT, we develop a progressive layer update strategy with channel-wise weight restoration that alleviates forgetting. To evaluate the proposed method, experiments were performed on three datasets, where the source and target data used different modalities for segmentation, or their images were acquired at different sites. The results show that our method improves the segmentation accuracy compared with other SFDA approaches.