CSCS framework solves cold-start problem for 3D medical image segmentation
First active learning method that selects initial volumes without any labeled data
A major bottleneck in 3D medical image segmentation is the need for extensive manual annotations. Active learning aims to reduce this burden by selecting informative samples, but most methods assume an initial labeled set exists—leaving the cold-start problem unresolved. Now, a team led by Rémi Hattat and colleagues proposes CSCS (Curriculum-Stratified Cold-Start), a framework that selects the first volumes from a fully unlabeled pool without any pre-trained task-specific model.
CSCS uses two self-supervised, label-free signals: local typicality (how representative a sample is in embedding space) and reconstruction-based uncertainty (a proxy for difficulty). These are combined through a weighted geometric score, with weights determined by a closed-form pacing rule based on the annotation budget and a new pool-level statistic called Difficulty-Coverage Ratio. Tested on four 3D medical benchmarks (BraTS, FeTA, Spleen, and an in-house fetal MRI dataset) using nnU-Net, CSCS consistently outperforms baselines, especially in low- to mid-annotation regimes. The method adapts sample selection to the geometry of the unlabeled pool, making active learning more robust from the very start.
- CSCS is the first cold-start active learning method for 3D medical segmentation that requires no initial labeled data
- Combines two self-supervised signals: local typicality and reconstruction-based uncertainty
- Achieves strongest performance gains on low-to-mid annotation budgets across BraTS, FeTA, Spleen, and fetal MRI datasets
Why It Matters
Reduces the manual annotation burden for 3D medical imaging, accelerating AI adoption in clinical workflows.