Label-Free Zoom Curricula for Low-Label Whole-Slide Pathology

A whole-slide pathology image is a scan of a tissue slide at microscope resolution, often billions of pixels. At full magnification it breaks into tens of thousands of small tiles. Transformers relate every tile to every other tile, so their cost grows with the square of the tile count, and they cannot read such a slide directly. A common workaround imitates a pathologist. The model looks at the slide at low magnification, picks a few suspicious regions, and zooms into only those. The weak point is the picking step. The model learns which regions to zoom into from a single diagnosis label per slide. When only a handful of labeled slides exist, the early picks are close to random. The high-magnification part of the model then trains mostly on irrelevant tissue and never gets good enough to correct the picks.

We propose to break this loop with a training curriculum that is itself label-free. Early in training, the model zooms into regions chosen by a self-supervised "surprise" signal: how badly a small network, given the low-magnification view of a region, predicts what that region looks like at high magnification. The zoom budget starts wide and narrows as training proceeds. The choice of where to zoom is gradually handed over from surprise to the model's own label-trained attention. The claim to test is that, at a fixed compute budget per slide, this schedule lets hierarchical zooming models trained on 10–20% of slide labels approach their fully labeled performance. It should also make the zoomed regions cover tumor more reliably than fixed-budget selection.

Core Claim

In low-label hierarchical multiple instance learning (MIL), the zoom decision is the main bottleneck. MIL here means classifying a slide as a "bag" of tile features when only the bag has a label. A curriculum over the zoom decision should fix it. Its difficulty measure is computed from unlabeled slides, and it moves from wide, surprise-driven zooming to narrow, attention-driven zooming.

Three things are new relative to existing work. First, the curriculum is over where and how much to zoom, not over which samples to train on. Second, its difficulty signal needs no labels, which addresses curriculum learning's standing problem of hand-designed difficulty measures. Third, it is evaluated specifically in the low-label regime, where selection is most fragile.

For region $r$ at training step $t$, the zoom score is

$$ \text{score}_t(r) = \alpha_t \, \hat{s}(r) + (1-\alpha_t)\, \hat{a}_t(r). $$

In this formula, $\hat{s}(r)$ is the rank-normalized cross-scale prediction error. It is $\lVert g(z_r^{\text{low}}) - \bar{z}_r^{\text{high}} \rVert^2$, where $z_r^{\text{low}}$ is the region's low-magnification embedding and $\bar{z}_r^{\text{high}}$ is the mean embedding of its high-magnification child tiles. The term $\hat{a}_t(r)$ is the rank-normalized attention the slide classifier currently assigns to the region. The weight $\alpha_t$ decays from 1 to 0. The top $k_t$ regions are zoomed, and $k_t$ shrinks from a wide initial budget to the final one. The attention layer over selected tiles uses Nyström attention, a low-rank approximation that costs time linear in the number of tiles. Together these keep per-slide cost bounded by $k_t$ and roughly constant across methods.

Relation to Prior Work

The closest existing system supplies the architecture we modify. Diagnose Like a Pathologist builds a quadtree over magnifications. It picks the top-$k$ patches by attention at low magnification, zooms into them, and fuses levels with Nyström attention. Its $k$ is fixed throughout training and its selection is purely label-driven from initialization. It was also not evaluated with reduced labels. We keep its hierarchy and change only how selection is trained.

Dual-Curriculum Contrastive MIL is the one prior work that combines curriculum learning with multi-magnification slides. Its first curriculum stage is saliency-guided encoding across 5×, 10× and 20× driven by supervised risk labels, and it targets survival prognosis. Our schedule differs in two ways: it needs no labels to start, and it acts on the zoom budget and selection source rather than on instance encoding. Its reported ablation, where removing the magnification curriculum lowered concordance, motivates the direction. It does not show that the curriculum helps when labels are scarce.

Two papers motivate the label-free selection signal.

  • Magnification Prior shows that pairing the same tissue at different magnifications is a strong self-supervised signal. It matches supervised accuracy on BreakHis with 20% of labels, but only on isolated patches. We use the cross-magnification relation not as a pretraining pair but as a map of where low magnification is uninformative.
  • Boosting WSI Classification filters up to 60% of negative patches. It does so using a negative feature bank built from labeled data, which is exactly what a low-label setting lacks.

CAMIL and SPAN constrain the design.

  • CAMIL combines self-supervised tile features with linear attention over a flat, single-magnification bag.
  • SPAN builds a sparse pyramid over frozen features. It shows that preserving spatial coordinates yields strong segmentation, with a SegCAMELYON Dice of 0.900.
  • Both show that flat or pyramid models with good features are strong. Any gain we claim must survive comparison with them at equal features.

The general case for easy-to-hard ordering comes from Self-Paced Learning. That work trains on low-error samples first to avoid poor early optima. Our "easy" phase is a wide zoom budget: it tolerates bad selection by covering more tissue.

Research Program

Pretraining. We train a tile encoder once on unlabeled slides using contrastive self-supervision, following CAMIL's recipe. The same frozen encoder is used for every method. We also repeat the main comparisons with PLIP features, following SPAN, to test whether the effect survives a stronger encoder.

Surprise predictor. On the same unlabeled slides, we train the small predictor $g$ to regress mean child-tile embeddings from low-magnification region embeddings. This is cheap because it operates on precomputed features.

Before any classifier training: does surprise locate tumor? CAMELYON16 has pixel-level tumor annotations, so we can check this directly. We measure the tumor recall of the top-$k$ surprise regions and compare it with random regions and with low-magnification feature norm. We also inspect failure cases. Tissue folds, pen marks and blur may be highly surprising yet diagnostically irrelevant. If they dominate, we will add a standard artifact-filtering step and report its effect separately.

Classifier training. The hierarchical classifier is trained with the schedule on $\alpha_t$ and $k_t$. We compare linear and step pacing and three decay lengths. Pacing is tuned on a validation split only.

Analysis. Beyond accuracy, we track two quantities during training:

  • the fraction of zoomed regions that overlap annotated tumor;
  • the agreement between surprise-based and attention-based rankings.

The mechanism we propose predicts that label-driven selection alone has poor tumor overlap early at low labels and recovers slowly, while the curriculum reaches high overlap sooner.

Baselines

Naive reference. Attention-based MIL over low-magnification tiles only, with no zooming, using the shared encoder. This measures what zooming buys at all. Beating it is necessary but not interesting.

Strongest hierarchical competitor. The Diagnose Like a Pathologist quadtree model with its fixed $k$ and attention-only selection. It reports CAMELYON16 AUC 0.946 and TCGA-NSCLC AUC 0.921 with full labels. It is the strongest competitor because it is our architecture minus our ingredient. We reimplement it on the shared encoder with its final $k$ equal to ours. It gets the same hyperparameter search budget, the same number of optimization steps, and the same label splits.

Strongest flat competitors. CAMIL reports CAMELYON16 AUC 95.9%, CAMELYON17 AUC 88.1% and TCGA-NSCLC AUC 97.5%. SPAN reports CAMELYON16 AUC 0.957 with PLIP features. Both process all high-magnification tiles, so each gets more compute per slide than our method. We run both on our label splits with the same encoder, and report their cost alongside accuracy. For CAMELYON16 and TCGA-NSCLC, no reported result found on Lacuna for any of these methods at 10% or 20% of slide labels, so all low-label numbers will be produced here.

Ablations isolating the ingredient.

  • A $k$-annealing schedule with attention-only selection ($\alpha_t = 0$), which tests whether the wide-to-narrow budget alone explains any gain.
  • Surprise-only selection with no handover ($\alpha_t = 1$).
  • Random selection with the same $k_t$ schedule.
  • A reversed curriculum, narrow to wide.

What counts as success. At 10% and 20% labels, the curriculum should beat the fixed-$k$ hierarchical model and the $k$-annealing ablation by a margin exceeding the seed-to-seed spread over at least five label-split seeds. It should also stay within the flat baselines' AUC while using a fraction of their tiles per slide.

Why not just…?

Why not just use strong frozen foundation-model features with a flat MIL head? This is the most serious threat. SPAN shows that switching from ResNet50 to PLIP features alone raises CAMELYON16 AUC from 0.939 to 0.957. With features that good, selection may matter little, and a flat head might already be label-efficient. We test this directly by running every comparison with PLIP features at 10% and 20% labels. If flat PLIP models match the curriculum at every label fraction, our claim shrinks to a compute claim: equal accuracy with far fewer tiles per slide. We would report it as such.

Why not just anneal the zoom budget without the surprise signal? A wide-to-narrow $k$ schedule is the cheapest curriculum and might capture most of the benefit by covering more tissue early. It appears as a named ablation. The surprise signal earns its place only if it beats this schedule at low labels, and if its early tumor overlap is measurably higher on CAMELYON16 annotations.

Why not just skip zooming and run linear attention over all high-magnification tiles? CAMIL does this with Nyström attention and reports strong AUCs, so hierarchy is not required for accuracy with full labels. Our argument concerns cost and low-label behavior. We report AUC against tiles per slide, peak memory and wall-clock time. If flat linear attention also wins at low labels for equal cost, the hierarchical premise fails for this task.

Why not just pretrain with a magnification-contrastive objective and stop there? Magnification Prior improves the tile encoder but says nothing about which regions to zoom. It is complementary. We will include a variant whose encoder is trained with cross-magnification pairs, to check whether a better encoder makes the surprise-based selection unnecessary.

Proposed Experiments and Decisive Checks

Datasets and metrics. We use CAMELYON16 for metastasis detection with tumor annotations, CAMELYON17 for multi-center shift, and TCGA-NSCLC for lung cancer subtyping. Label fractions are 10%, 20% and 100%, and the remaining slides serve as unlabeled data for the encoder and the predictor $g$. Metrics are slide AUC and F1. For localization we report heatmap Dice on CAMELYON16, where CAMIL reports 0.515 against CLAM-SB's 0.459. Cost is reported as tiles attended per slide, FLOPs and peak GPU memory.

First decisive check. Before training any classifier, surprise-ranked regions must recover annotated tumor substantially better than random regions at the same budget. If they do not, the method has no basis and we stop.

Second decisive check. At 10% labels on CAMELYON16, where micrometastases make selection hardest, the curriculum must beat both the fixed-$k$ model and the $k$-annealing ablation. It must also show earlier and higher tumor overlap in the zoom set during training.

Third decisive check. The PLIP-feature comparison against flat heads determines whether the result is about label efficiency or only about compute.

Secondary check. A CAMELYON17 held-out-center evaluation tests whether a label-free selection signal generalizes across centers better than label-trained attention. This is plausible, because surprise is computed per slide rather than learned from labels, but it is not guaranteed.