Source-free unsupervised adaptive segmentation for knee joint MRI

•To the best of our knowledge, this is the first study of source-free UDA adaptation for multi-tissue knee joint segmentation in which the source and target data were obtained from different MRI pulse sequences and different MRI systems.•We utilized two novel ideas to enhance the quality of the pseudo labels for unlabeled target data: improving the pseudo labels through augmented probability maps and introducing uncertainty into the objective function by estimating the variance of predictions in the cross-pseudo supervision (CPS) mechanism. The augmented probability maps can offer a more comprehensive and continuous representation of the predicted segmentation compared with a single point-wise prediction. By incorporating spatial information and considering the uncertainty associated with each prediction, the augmented probability maps contribute to more accurate and reliable pseudo labels. Furthermore, the proposed uncertainty-aware objective function guides the model to focus on regions where it exhibits higher reliability while being more cautious in uncertain areas.•The proposed simple and effective source-free UDA method combines feature alignment, self-learning, and consistent regularization methods to achieve high-performance knee joint segmentation, surpassing the existing state-of-the-art source-free UDA methods. Batch normalization statistics are used to align task-related features from both the source and target domains, facilitating effective adaptation of the model to the target dataset. A voting strategy is used to iteratively refine pseudo-labels, enabling robust learning from unlabeled target data. The self-learning mechanism enhances the model's ability to generalize and adapt to the target domain data. Additionally, we utilize consistency regularization, introducing data and network perturbations to encourage consistent predictions and prevent overfitting, thereby improving the model's generalization capability. Knee osteoarthritis is a prevalent disease worldwide. The automatic segmentation of knee tissues in magnetic resonance (MR) images has important clinical utility in assessing knee osteoarthritis. Deep learning-based methods show great potential in this application, but they often require a large amount of labeled training data, which is challenging and expensive to acquire. Unsupervised domain adaptation that transfers the learned knowledge from a source labeled dataset to a target unlabeled dataset can be used to address this pro