Fig. 3. Results of prostate segmentation for UDA.

a–d, The four evaluation metrics, DSC (%), RAVD (%), ASSD (mm), and MSSD (mm). In each subfigure, the left and right mappings indicate that the networks are trained without and with the proposed AIDE. D1, D2, and D3 refer to domains 1, 2, and 3. In each mapping, the vertical axis indicates the dataset utilized to train the models and the horizontal axis indicates the dataset utilized to test the models. e–f Example segmentation results when transferring models between domains. GT stands for ground truth, referring to the high-quality annotations. Conven is conventional, indicating that the results are generated by a fully supervised optimization method utilizing the combined dataset (high-quality labels for the source domain training data and model-generated low-quality labels for the target domain training data).