Towards Practical Algorithm Selection for Unsupervised Domain Adaptation in Medical Imaging

This paper proposes a label-free criterion for jointly selecting an unsupervised domain adaptation algorithm and its hyperparameters when deployment-domain labels are unavailable. Given candidate models trained by multiple UDA methods and hyperparameter settings, the method first uses several label-free signals to nominate one model per algorithm. It then aggregates those nominees across algorithms into an agreement reference for each unlabeled target sample. The candidate with the highest agreement to this reference is selected. Experiments cover four brain MRI datasets, four chest X-ray datasets, and seven clinically relevant transfer scenarios. The abstract reports better selection performance than competing methods across different algorithm pools.