Active few-shot segmentation by reinforcing data selection

2026-07-27 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Mathematics & Computational Sciences, Health & Medical Research · Depth: Expert, long

Summary

A new reinforcement learning (RL) framework has been developed for support-set selection in few-shot medical image segmentation. This framework optimizes support sets jointly, rather than through independent sample scoring, to maximize downstream segmentation performance. Experiments on a cross-institutional pelvic MRI dataset demonstrated significant improvements, with the proposed method achieving an 8.1 percentage point increase over random selection and a 4.5 percentage point increase over DVRL at a support-set size of K=4. The approach consistently outperformed random selection, Task-Based Prioritization (TBP), and Data Valuation using Reinforcement Learning (DVRL) across various support-set sizes and anatomical structures. Ablation studies confirmed that larger candidate pools (e.g., P=32 vs. P=8) further improved performance and that the benefits generalized to other meta-learning algorithms like MAML.

Key takeaway

For Machine Learning Engineers developing few-shot medical image segmentation models, you should implement reinforcement learning for support-set selection. This approach significantly improves adaptation performance by jointly optimizing for complementary examples, rather than relying on individual sample utility. Consider using larger candidate pools for selection, as this can further enhance segmentation accuracy. This method generalizes across meta-learning frameworks like Reptile and MAML, offering a robust strategy for data-scarce medical imaging tasks.

Key insights

Reinforcement learning can optimize few-shot support sets by jointly selecting complementary examples for improved medical image segmentation.

Principles

• Support-set quality is a set-level property, not a sum of individual sample utilities.
• Complementarity among selected examples is crucial for effective adaptation.
• Joint optimization of support sets outperforms individual sample scoring.

Method

A reinforcement learning agent directly predicts a support set from an unlabelled candidate pool. This set adapts a meta-learned segmentation model, with validation performance providing the reward signal to optimize the selection policy.

In practice

• Apply RL for support-set selection in few-shot medical imaging tasks.
• Prioritize complementary examples over individual sample utility.
• Consider larger candidate pools for enhanced support-set identification.

Topics

• Few-shot Learning
• Medical Image Segmentation
• Reinforcement Learning
• Support-set Selection
• Meta-learning
• Pelvic MRI

Code references

• timfkl/active-few-shot-learning

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Active few-shot segmentation by reinforcing data selection — Few-shot learning performance in medical image segmentation significantly improves by optimizing support sets for complementarity using reinforcement learning.
• HPR-SAM: Hierarchical Probabilistic Representation Learning for Prompt-free SAM-based Medical Image Segmentation — HPR-SAM improves prompt-free medical image segmentation by learning hierarchical probabilistic anatomical representations for SAM.
• Have I Solved This Before? Retrieving Similar Segmentation Problems for Evolutionary Learning — Storing abstract system models enables retrieving and refining existing solutions for new segmentation problems.
• SegMix:Shuffle-based Feedback Learning for Semantic Segmentation of Pathology Images — Shuffle-based feedback learning improves weakly supervised semantic segmentation in pathology by adapting to multi-scale image characteristics.
• Re-Evaluating Continual Learning with Few-Shot Adaptation — Few-shot evaluation offers a more comprehensive assessment of continual learning model stability and plasticity than traditional 0-shot methods.
• PSP: Harnessing Position and Shape Priors for Cross-Domain Few-Shot Medical Image Segmentation — Harnessing invariant anatomical position and shape priors enables robust cross-domain medical image segmentation despite texture discrepancies.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.