RegAL unifies active and semi-supervised learning for medical image segmentation
Boosts Dice scores by 5%+ in ultra-low annotation regimes with only 2 labeled volumes.
Deep Dive
Researchers propose RegAL, a unified active semi-supervised framework for medical image segmentation. It couples sample acquisition with unlabeled data utilization via a shared topology-aware Pareto optimization that evaluates voxel uncertainty, feature diversity, and topological consistency. Across BraTS 2021, dHCP, and ProstateX, RegAL outperforms state-of-the-art baselines in Dice, ASD, and HD95 metrics under extreme annotation scarcity, remaining stable with few labeled volumes.
Key Points
- RegAL uses a shared Pareto optimization to couple active learning sample selection with semi-supervised training, avoiding objective mismatch.
- It evaluates images on voxel uncertainty, feature diversity, and a novel topological consistency metric for selecting anatomically informative edge cases.
- Outperforms state-of-the-art methods on BraTS 2021, dHCP, and ProstateX, with stable performance using only 2 labeled volumes.
Why It Matters
Enables accurate medical image segmentation with minimal labeled data, reducing annotation costs and accelerating clinical AI deployment.