A Source-Free Segmentation Quality Estimation of a Model Adapted to a New Domain

We address the challenge of estimating segmentation quality in source-free domain adaptation (SFDA), where access to labeled source data is restricted and no labels are available in the target domain. Our approach leverages the segmentations produced by the source model as pseudo-labels and refines them to construct enhanced surrogates of the unknown ground truth. We show that standard metrics such as Dice and ASSD, when computed with these pseudo-labels, closely approximate the true segmentation quality of the adapted network. Evaluated on cardiac MRI and prostate MRI benchmarks, our method computes segmentation quality estimation that is more accurate than methods requiring labeled source data. These results demonstrate that pseudo-labels, commonly used for adaptation, can also provide a simple and effective basis for post-hoc segmentation quality estimation.