Pathology AI models read whole-slide images: scans of tissue with ten billion pixels or more, far too large to fit on a GPU at full resolution. Fast systems first look at a cheap low-magnification view and then zoom in only on regions that look suspicious. This works until the disease is a micrometastasis a few cells wide. The low-resolution view blurs it away, the system never zooms in, and the error cannot be recovered later. Time-series search faced the same problem decades ago and solved it without heuristics. To find the best match for a pattern in a very long signal, you compute a cheap lower bound on the true distance, and you skip a candidate only when that bound proves it cannot beat the best match found so far. Candidates are skipped only when the bound certifies they cannot win, and on real data most candidates are skipped (Fast Subsequence Search on Time Series Data; Efficient Retrieval of Similar Time Sequences Using DFT).
We propose carrying this "bound, sort, stop" logic into slide processing. A low-resolution scout network is trained to upper-bound the score the expensive high-resolution model would give each tissue patch, not to guess that score. The bound is calibrated so that the rate of wrongly skipping a decisive patch stays below a user-chosen $\epsilon$. Patches are then processed in descending order of bound, and processing stops once no remaining patch can enter the current top-$k$. Two consequences follow. Compute adapts to each slide: a slide with an obvious tumor ends early, while an ambiguous slide keeps searching. The miss rate on annotated lesions also becomes a quantity we can measure and control instead of an unobserved side effect. Skipped patches are not thrown away. Following the soft-sparsification result from time series, they are folded into a single cheap background summary.
Core Claim
Hierarchical zooming such as Differentiable Zooming for Multiple Instance Learning on Whole-Slide Images (ZoomMIL) picks a fixed number of patches per magnification with a differentiable top-$K$ and reports 40–50x fewer FLOPs. Its budget is fixed per slide, and it has no notion of how likely a discarded region was to matter. Scout networks (Dong et al.) and normal-tissue atlases (Yasir) have the same property. These papers motivate the proposal: they establish both the savings and the "needle in a haystack" failure.
Our claim is narrower and testable. At matched average compute, a scout trained as a calibrated upper bound and paired with best-so-far early stopping will skip far fewer ground-truth lesion patches than a scout trained as a predictor with a threshold. That difference should show up as higher slide-level sensitivity on micrometastasis-only slides.
The stopping rule works as follows. Let $s(x_i)$ be the high-resolution instance score of patch $i$ (for example, its attention logit in the slide classifier), let $u(\tilde x_i)$ be the scout's output on the low-resolution view $\tilde x_i$, and let $s_{(k)}$ be the $k$-th best high-resolution score found so far. Patches are visited in decreasing order of $u$, and the search stops at the first patch for which
$$u(\tilde x_i) + q_\epsilon < s_{(k)},$$
where $q_\epsilon$ is an offset fitted on held-out slides by conformal risk control. The offset is chosen so that the expected fraction of true top-$k$ patches left unvisited is at most $\epsilon$. If $u + q_\epsilon$ really bounds $s$, the stopping rule returns the exact top-$k$. The only approximation is the calibrated failure rate of the bound. The same idea explains the UCR-suite speedups: ordering candidates by a tight bound makes the best-so-far threshold rise fast, and a high threshold prunes more. CNN Kernels Can Be the Best Shapelets contributes the observation that pattern-matching distances reduce to convolution plus pooling. That observation is why a small convolutional scout is a natural family for learning such bounds, rather than an arbitrary heuristic.
Research Program
The first component is the bound-trained scout. It is a small CNN on 5x thumbnails, trained against the high-resolution scorer's outputs with an asymmetric pinball loss at a high quantile. Underestimating a patch's score is penalized much more heavily than overestimating it. The result is a tight upper envelope rather than a mean predictor. The design question is how tight this bound can be made for small lesions, because a loose bound is correct but prunes nothing. We will measure tightness directly as the distribution of $u + q_\epsilon - s$ on lesion versus benign patches.
The second component is soft background fusion. Learning Soft Sparse Shapes for Efficient Time-Series Classification (SoftShape) showed across 128 UCR datasets that fusing low-scoring subsequences into one context token beats discarding them. We supply method here, not evidence about pathology, and that paper is univariate 1D. We transfer the mechanism: unvisited patches contribute through a single token built from their scout features, which cost nothing extra. This keeps the microenvironment context that Snuffy and CAMIL argue matters.
The third component handles a moving target, and it is where curriculum learning and the data-selection literature act as constraints. During training the high-resolution scorer changes, so the scout's target drifts. Instruction-tuning selection methods such as SPICE and CPQS-Tuning choose one static subset up front. CPQS additionally reports that subsets chosen for one model transfer poorly to another. That constrains us: a scout fitted to an early scorer is fitted to the wrong model. We therefore refit the scout and recalibrate $q_\epsilon$ at fixed intervals.
There is a sharper risk. Training only on patches the scout keeps can create a self-reinforcing loop, a patch-level version of the "easy-A" bias that MHIM-MIL counters by masking the highest-attention patches. Easy-first curricula such as Self-Paced Learning for Latent Variable Models and A Task-centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula pull the other way. That theory also warns that a curriculum fails if the starting model cannot solve the easiest stage. We resolve the tension empirically with one schedule. Training begins with exhaustive processing on a small slide subset, which gives a feasible initialization. We then enable pruning with a large exploration fraction, in which a random sample of below-bound patches is still processed at high resolution. That fraction is annealed down while $\epsilon$ stays fixed. The random exploration also supplies unbiased calibration data for the bound.
Baselines
The naive reference is exhaustive processing: every tissue patch at 20x, frozen features, attention-based aggregation, with no pruning. It sets the accuracy ceiling for the same aggregator and the compute cost that all savings are measured against. A second naive reference is uniform random patch subsampling at the same average patch budget as our method. It shows how much of any gain comes from choosing patches at all.
The strongest existing efficiency method is ZoomMIL, which is the closest competitor on this exact question: accuracy under a reduced high-resolution budget. The file has no reported number for ZoomMIL on CAMELYON16, so we will retrain it ourselves on identical features, tissue masks, and splits. We will sweep its per-level $K$ so that its average FLOPs matches ours at several budgets, and give it the same hyperparameter search budget.
For absolute accuracy, the strongest reported CAMELYON16 slide-level result in the file is Snuffy's exhaustive variant at 0.987 AUC (Snuffy). MHIM-MIL reports 96.49% AUC and CAMIL 95.9% AUC with localization Dice 0.515. We plug our pruning in front of the MHIM-MIL and CAMIL aggregators, where the file gives numbers to reproduce. Beating the baselines means two things. At matched FLOPs, our method must beat ZoomMIL on sensitivity for micrometastasis slides and have a lower lesion-patch skip rate. At no more than one fifth of exhaustive FLOPs, it must stay within one AUC point of the same aggregator run exhaustively.
Three ablations isolate the key ingredient. The first uses the identical scout trained with ordinary regression loss and a conformal threshold. This is the decisive ablation, because it tests whether training for a bound matters or whether calibration alone does the work. The second removes early stopping and uses a fixed top-$K$. The third removes the soft background token and discards unvisited patches.
We also include the obvious cross-field baseline: gradient-conflict-penalized selection in the style of SPICE, applied to patches using per-patch diagonal Fisher scores. It requires high-resolution features and gradients for every patch, so it cannot save inference compute. We run it only as a training-time selector, to check whether it chooses more useful training patches than our scout.
Why not just…?
Why not just lower ZoomMIL's or a scout network's threshold until nothing is missed? This is the main threat. The regression-scout-plus-conformal ablation is exactly this alternative made rigorous. If it matches our bound-trained scout on lesion skip rate at equal FLOPs, the asymmetric bound training adds nothing, and the contribution reduces to per-slide adaptive stopping plus calibration. We would report that. Our prediction is that a mean-predicting scout systematically underestimates the rare, tiny lesions, which are exactly the cases where the regression target is dominated by benign neighbors. A calibrated offset large enough to cover them then forces processing of most benign tissue as well. A quantile-trained scout places its slack only where uncertainty is high.
Why not just subsample patches at random, or cluster embeddings and keep centroids? Random sampling at a 10–20% budget should miss a lesion a few patches in size most of the time, simply because of how few patches the lesion occupies. Our random baseline at matched budget quantifies this. The poor localization of permutation-invariant models, with TransMIL at Dice 0.103 versus CAMIL at 0.515, points the same way. Clustering keeps representative tissue, but by construction a micrometastasis is not representative. The data-selection literature (Peiqi Wang et al.) also argues that surface-embedding similarity misses functional redundancy.
Why not just extract features for every patch once and accept the cost? For training on a fixed research cohort this is reasonable, and our exhaustive reference does exactly that. The target here is deployment and retraining with fine-tuned backbones, where feature extraction at 20x across the whole slide dominates cost and has to be repeated. DiPalma et al. use distillation to cut compute 64x, but their student sees only low resolution and has no mechanism to revisit high resolution. Our method keeps high-resolution evidence for the patches that could matter. If exhaustive processing turns out to be cheap relative to aggregation in some setting, the proposal's value shrinks there, and our wall-clock measurements will show it.
Why not just apply SPICE-style coreset selection to patches? Scoring gradient conflict needs high-resolution computation on every patch. That removes the saving at the moment it matters, and it answers a different question: which patches to train on, not which patches to look at. We still run it as a training-time comparison.
Proposed Experiments
The main testbed is CAMELYON16. Its pixel-level lesion annotations let us compute the lesion-patch skip rate, and we will stratify test slides by largest lesion size so that micrometastasis-only slides form their own group. We report slide AUC, sensitivity on the micrometastasis stratum, Dice, FROC (NCRF's reported 0.8096 is a localization reference: Cancer Metastasis Detection With Neural Conditional Random Field), the realized versus target $\epsilon$, and per-slide FLOPs and wall-clock. The plot of accuracy against compute at several budgets is the headline result.
Calibration assumes held-out slides resemble test slides. We therefore calibrate on some CAMELYON17 centers and test on others; CAMIL reports 88.1% AUC on CAMELYON17. This tests whether the miss-rate guarantee survives a scanner and staining shift. We expect it to degrade, and we will measure by how much. TCGA-NSCLC serves as a control where the signal is diffuse rather than needle-like. There, early stopping should terminate quickly and match exhaustive accuracy, and the gains over ZoomMIL should shrink.
Three checks decide the outcome. The first is whether realized lesion skip rates track $\epsilon$ in-distribution. The second is whether the bound-trained scout beats the regression-scout ablation on the micrometastasis stratum at matched FLOPs. The third is whether per-slide compute correlates with slide ambiguity rather than staying flat. A final curriculum check compares three regimes by attention entropy and micrometastasis sensitivity: exhaustive warm-up with annealed exploration, pruning from the first step, and MHIM-style high-attention masking combined with our pruning. The comparison shows whether the self-reinforcing easy-A loop appears and whether exploration prevents it.
References
- Fast Subsequence Search on Time Series Data
- Efficient Retrieval of Similar Time Sequences Using DFT
- Differentiable Zooming for Multiple Instance Learning on Whole-Slide Images
- Fast and Accurate Gigapixel Pathological Image Classification with Hierarchical Distillation Multi-Instance Learning
- Creating an atlas of normal tissue for pruning WSI patching through anomaly detection
- CNN Kernels Can Be the Best Shapelets
- Learning Soft Sparse Shapes for Efficient Time-Series Classification
- Snuffy: Efficient Whole Slide Image Classifier
- CAMIL: Context-Aware Multiple Instance Learning for Cancer Detection and Subtyping in Whole Slide Images
- SPICE: Submodular Penalized Information–Conflict Selection for Efficient Large Language Model Training
- CPQS-Tuning: A Model Self-Perception-Based Data Filtering Algorithm for Efficient Instruction Fine-Tuning
- Multiple Instance Learning Framework with Masked Hard Instance Mining for Whole Slide Image Classification
- Self-Paced Learning for Latent Variable Models
- A Task-centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula
- Diversity Measurement and Subset Selection for Instruction Tuning Datasets
- Resolution-based distillation for efficient histology image classification
- Cancer Metastasis Detection With Neural Conditional Random Field