Active few-shot segmentation by reinforcing data selection

TL;DR - This paper uses reinforcement learning to jointly select complementary support examples for few-shot medical image segmentation. On cross-institutional pelvic MRI data, it improves adaptation over random selection and existing state-of-the-art methods.

• The agent selects an entire support set from an unlabeled candidate pool to maximize downstream segmentation performance.
• Joint optimization accounts for interactions and complementarity between examples, unlike independent sample-scoring approaches.
• The method targets robust adaptation to new segmentation tasks with very few labeled examples.
• Results emphasize that capturing target-domain variation across the support set is important for effective adaptation.