Active learning beats full-dataset training on the hardest anatomical structures
A paper posted on 22 September 2026 tests an active-learning method called SASS on a pool of 100,357 candidate samples spanning 108 anatomical structures, built by stacking ten public segmentation datasets into one long-tailed benchmark. At a 40% annotation budget, SASS recovers 98.3% of the Dice score a model gets from labeling everything. On the 18 hardest structures, it does not just approach that ceiling, it clears it: 0.622 Dice against 0.610 for the fully labeled pool, a gap the authors report as statistically significant (p=0.011).
What changed, and when
On 22 September 2026, Xiaofei Du, Lei Zhang, Shuyu Yan, Manning Wang and Zhijian Song posted Less Is More in the Long Tail: Stage-Adaptive Sample Selection for Annotation-Efficient Dense Prediction (arXiv:2609.25850). The paper's test case is dense prediction on 3D medical images, where the annotation cost per sample is high (a radiologist segmenting a volume, not clicking a bounding box) and the class distribution is long-tailed: a handful of structures like ribs and vertebrae appear in nearly every scan, while structures like the gallbladder or pancreas appear rarely and get proportionally little training signal even when the whole dataset is labeled.
Most active-learning work claims to match full-dataset performance at a fraction of the labeling cost. This paper's distinguishing claim is narrower and stronger: on the classes that are already starved for data, a well-targeted subset can outperform using everything, because "everything" is itself skewed toward the common classes.
The evidence table
The benchmark pools ten public datasets, TotalSegmentator, AMOS, BraTS21, KiTS, MMWHS, CrossMoDA, FLARE22, WORD, AbdomenCT-1K and VerSe, into 100,357 image-category units across 108 structures, split into 70,351 training candidates, 15,003 validation samples and 15,003 held-out test samples. All methods below select from the same 70,351-candidate training pool at a 40% budget and are scored on the same held-out test set, three runs each.
| Method | What it does | Overall Dice |
|---|---|---|
| Random | uniform sampling from the pool | 0.614 ± 0.003 |
| Entropy | uncertainty-based selection | 0.628 ± 0.003 |
| Coreset | diversity-based selection | 0.624 ± 0.003 |
| BADGE | gradient-embedding active learning | 0.637 ± 0.003 |
| LESS | prior long-tailed active-learning baseline | 0.638 ± 0.003 |
| LESS+Organ | LESS with an added organ-frequency prior | 0.658 ± 0.003 |
| SASS | this paper: gradient scoring + rebalancing + staged acquisition | 0.688 ± 0.003 |
| Full dataset (100%) | every candidate labeled | 0.700 ± 0.002 |
Source: arXiv:2609.25850, Table I, epoch 300, mean ± standard deviation over 3 runs. A variant with class-balanced sampling on the full dataset ("Full+CB") reaches 0.702 ± 0.003, not shown.
The per-group and per-structure breakdown is where the paper's real claim sits, and it is not in the headline table. On the 18-structure Hard group, SASS scores 0.622 ± 0.005 against 0.610 ± 0.005 for the full dataset, a difference of 0.012 (95% CI 0.003 to 0.021, p=0.011). On two individual structures: pancreas goes from 0.590 ± 0.009 (full dataset) to 0.607 ± 0.010 (SASS), and gallbladder from 0.598 ± 0.009 to 0.614 ± 0.010. Neither per-structure gap carries its own reported p-value, and the intervals are not adjusted for testing multiple structures, which the paper does not claim to correct for.
Why a smaller, chosen subset beats everything
The mechanism is not "fewer but better-chosen samples generalize better" in the usual active-learning sense of avoiding redundant easy examples. It is that the full dataset's class balance is itself the obstacle: ribs and vertebrae recur across nearly every scan, so a model trained on every labeled sample spends most of its gradient steps on structures it already segments well, and the rare classes get a share of training signal proportional to their rarity in the data, not their difficulty.
SASS attacks this in three stages. First, a self-supervised gradient-scoring step ranks unlabeled candidates by how much they would change the model, measured as the gradient norm of a DINO-style teacher-student objective between the current checkpoint and an earlier one, with random projection to keep this cheap at 100,000-candidate scale. Critically, this scoring needs no ground truth, so it runs on the full unlabeled pool. Second, a prior-guided rebalancing step overrides pure gradient ranking with two priors: a scale-aware prior that favors small structures and a coverage-balance prior that penalizes categories already over-selected, with validation feedback boosting any category whose Dice falls below the median, and a hard group-budget cap stopping repetitive groups like ribs and vertebrae from consuming a batch. Third, acquisition is staged over training: early rounds rely on the priors alone, because the model is not yet reliable enough for gradient scoring to mean anything; gradient scoring switches on once validation Dice crosses 0.42, and its weight rises after that, with a forgetting-based curriculum reordering the pool after each expansion to stabilize training.
What it means for anyone budgeting annotation on imbalanced data
"Label everything" is not the ceiling when classes are imbalanced. The usual frame for active learning is a tradeoff: less labeling for a little less accuracy. This result removes that tradeoff for the classes that matter most in a long-tailed task, because full-dataset training is itself biased toward the common classes. Spending less, deliberately, on the structures the data already over-represents is not a compromise, it is a correction.
The gain is concentrated, not general. SASS's overall-score lead over the next-best method (LESS+Organ, 0.658) is three points of Dice, a real but modest margin. Its lead on the Hard group (1.2 points over full-dataset training) is smaller in absolute terms but is the one the paper reports as statistically significant, and it is the one that would not show up at all in an overall average, which is dominated by the common structures that both methods already handle well.
A annotation budget decision needs the per-class breakdown, not the headline number. Comparing selection methods only on overall Dice would have hidden the Hard-group result entirely: SASS's 68.8 overall score is still two points below the full dataset's 70.0. The finding worth acting on is buried in a table most reports would cut for space.
Our judgement: this is one paper's benchmark, assembled by the authors from ten existing public datasets rather than measured on an independent held-out clinical population, and the per-structure confidence intervals are not corrected for testing multiple structures at once. Two individual-structure results (pancreas, gallbladder) out of 108 structures is suggestive, not conclusive, that the long-tail effect generalizes past the Hard group as a whole.
Check it yourself
The ten source datasets are all public. The comparison that matters is not the overall Dice column, it is the Hard-group and per-structure rows.
open https://arxiv.org/abs/2609.25850 # abstract, authors, submission date
open https://arxiv.org/html/2609.25850 # full text: Table I, Section 4 (method), Section 5 (Hard-group and per-structure results)
# the 10 pooled source datasets, all public
# TotalSegmentator, AMOS, BraTS21, KiTS, MMWHS, CrossMoDA, FLARE22, WORD, AbdomenCT-1K, VerSe
python3 - <<'PY'
# recompute the headline comparisons directly from the paper's reported numbers
full, sass = 0.700, 0.688
print("overall recovery at 40%% budget: %.1f%%" % (100 * sass / full))
hard_full, hard_sass = 0.610, 0.622
print("Hard-group gap (SASS minus full): %+.3f Dice" % (hard_sass - hard_full))
pancreas_full, pancreas_sass = 0.590, 0.607
gallbladder_full, gallbladder_sass = 0.598, 0.614
print("pancreas gap: %+.3f, gallbladder gap: %+.3f" % (pancreas_sass - pancreas_full, gallbladder_sass - gallbladder_full))
PY
# overall recovery at 40% budget: 98.3%
# Hard-group gap (SASS minus full): +0.012 Dice
# pancreas gap: +0.017, gallbladder gap: +0.016The overall-recovery figure (98.3%) is a simple ratio and matches the paper's own framing. The Hard-group and per-structure gaps are differences between reported means; the paper's own significance test (p=0.011) applies to the Hard-group comparison, not to either individual structure, which it reports with confidence intervals but no separate p-value.
What would prove this wrong
The benchmark is built by the paper's own authors from public datasets they chose and pooled into one long-tailed testbed, and the Hard-group definition (18 structures) is also theirs. Both choices could favor the method they are introducing without anyone having manipulated results, simply because the testbed was built to showcase a long-tail effect.
A dated prediction: by 31 December 2027, an independent group, not the SASS authors, will run a comparable long-tailed active-learning comparison on a different pooled benchmark (a different set of source datasets, a different Hard-group definition) and report whether a stage-adaptive, rebalancing-based selector still beats full-dataset training on the rarest classes. If independent replications instead show the rare-class gain shrinks to noise once the benchmark and the method are chosen by different people, the generalizable claim here is wrong, whatever this specific benchmark shows.
Sources
- Du, X., Zhang, L., Yan, S., Wang, M., Song, Z. Less Is More in the Long Tail: Stage-Adaptive Sample Selection for Annotation-Efficient Dense Prediction. arXiv:2609.25850, 22 September 2026. Table I, Sections 3 to 5; HTML version.
- Source datasets pooled into the 100,357-candidate benchmark: TotalSegmentator, AMOS, BraTS21, KiTS, MMWHS, CrossMoDA, FLARE22, WORD, AbdomenCT-1K, VerSe, as listed in arXiv:2609.25850 Section 4.
- BLOMEGA. LESS data selection, recounted.
- BLOMEGA. GCcore: pathology coreset selection against a random budget.
- BLOMEGA. Data annotation research: the latest.