A class-aware multi-stage UDA framework for prostate zonal segmentation

Unsupervised domain adaptation (UDA) aims to solve the lack of annotation in a new dataset which has non-independent identity distribution compare with training data. It has the potential to help the annotation process in medical image segmentation. Existing self-training based UDA approaches utilize the pseudo labels as ground truth labels for domain adaptation, whereas the generated pseudo labels inevitably introducing the noise when training the model for the target domain, which make the training process unstable and the model is difficult to converge. In the meanwhile, most of the methods ignore the class imbalanced problem. To tackle the issue, we propose a class-aware multi-stage unsupervised domain adaptation framework for prostate zonal segmentation task. We devise a class-specific knowledge guidance strategy for training a better pseudo labels generation model. Extensive experimental results show the effectiveness of our approach against existing state-of-the-art approaches on the UDA problem of prostate zonal segmentation benchmark.