Uncertainty-guided gradual learning for one-shot segmentation

In this work, Uncertainty-Guided Gradual Learning (UGGL) is proposed for one-shot segmentation of slice-stacks, where only a single labeled slice is available. UGGL extends regular Gradual Learning (GL) with ensemble-based consensus for adaptive sample selection and entropy-based uncertainty weighting during student training. A k-head teacher ensemble is trained on the single ground-truth slice to generate pseudo labels (PLs), uncertainty, and consensus maps. UGGL then selects neighboring slices with high mean consensus (excluding boundary regions) and iteratively trains a student ensemble, prioritizing reliable PLs to prevent error propagation. The evaluation was conducted on the Calgary-Campinas MRI dataset for brain segmentation and compared UGGL against GL with fixed step rates and guided teacher–student (TS) training. Trained from one labeled slice per volume, UGGL improved mean Intersection over Union (IoU) by approximately 0.10 over the one-slice baseline, reducing the remaining gap to fully supervised training to approximately 0.02 IoU, and achieved higher precision and recall with lower variance than both GL and TS. Cross-scanner experiments further indicated superior generalization relative to GL and TS, with the strongest gains on challenging slices. Qualitative 2D and 3D results showed cleaner masks and reduced boundary errors. These properties make UGGL a broadly applicable method for one-shot segmentation in real-world settings.