Stage-adaptive sample selection cuts annotation needs in 3d medical segmentation
Less Is More in the Long Tail: Stage-Adaptive Sample Selection for Annotation-Efficient Dense Prediction
Computer Vision and Pattern Recognition
Summary
Training deep learning models for detailed image tasks usually requires lots of expert-labeled data, which is expensive and slow to collect. The authors present SASS, a method that carefully picks which samples to label by adapting to how the model learns and focusing on underrepresented categories. Tested on a large set of 3D medical images covering many body parts, SASS matches almost full performance while using far fewer labeled samples. Interestingly, for some difficult cases, using less labeled data actually led to better results than using everything, suggesting smarter sample selection matters more than quantity.
What this means in practice
- •For medical imaging teams: Reduce expert labeling effort for 3D scans by selecting the most informative samples to annotate for organ and tissue segmentation.
- •For industrial quality control teams: Improve labeling efficiency in volumetric inspection tasks by stage-adaptive sample selection that prioritizes difficult or rare defect types.
Authors
Xiaofei Du, Lei Zhang, Shuyu Yan, Manning Wang, Zhijian Song
Abstract
Deep learning performance generally improves with increasing training data, yet this scaling is fundamentally constrained by annotation cost in large-scale dense prediction tasks with long-tailed category distributions, where pixel- or voxel-level annotation is prohibitively expensive. We propose SASS (Stage-Adaptive Sample Selection), a stage-adaptive data-selection framework for pool-based active learning in long-tailed dense prediction. SASS combines three components: label-free self-supervised gradient scoring, prior-guided category rebalancing with validation-driven feedback, and stage-adaptive acquisition aligned with model training dynamics. This design avoids candidate ground-truth masks during gradient scoring while making acquisition responsive to long-tail imbalance and evolving representations. We evaluate SASS on a multimodal 3D medical segmentation testbed comprising over 100,000 samples spanning 108 anatomical structures. SASS recovers 98.3% of full-dataset performance with a 40% training-pool annotation budget, outperforming BADGE by 5.1 percentage points. Moreover, SASS exhibits a statistically supported less-is-more pattern, surpassing full-dataset training at the Hard-group level and, at the structure level, for the pancreas and gallbladder. More broadly, SASS shows that annotation-efficient learning depends not only on which samples are selected, but also on how the annotation budget is distributed across categories and when model-derived scores begin to guide selection.