Certified Zoom: Coarse-to-Fine Slide Search With a Guaranteed Miss Rate

Systems that read gigapixel pathology slides save time by looking at a blurry thumbnail first and zooming in only where something looks suspicious. This is how they get hundredfold speedups. It is also how they miss things. A tiny metastasis that is invisible at low magnification never gets a second look, and nothing downstream can bring it back. We propose to treat each "should I zoom here?" decision as a yes/no question at one node of a spatial tree. The thresholds for these questions would be calibrated jointly on held-out slides, so that the whole cascade misses annotated lesions no more than a user-chosen fraction of the time, for example 5%. Within that constraint, the system reads as few high-resolution pixels as possible and stays inside a fixed memory budget.

The core idea fits in two sentences. Zoom-in pipelines currently prune regions using fixed top-$K$ cutoffs or uncalibrated scores, so their false-negative rate is whatever it happens to be. We instead split an explicit miss budget across magnification levels and add a cheap recovery step: confirmed tumor revives its pruned neighbors. The result is a guarantee on missed lesions, not just a speedup, and the key question is whether this spends high-resolution reads more efficiently than simply zooming into more regions.

Core Claim

Current coarse-to-fine pathology models attack speed, not misses. The Self-reform Multilayer Transformer (Hundredfold Accelerating for Pathological Images Diagnosis and Prognosis through Self-reform Critical Region Focusing) zooms through three levels, keeping the top $K$ patches at each level. It adds a backward "recheck" that penalizes the coarse selector when high-resolution evidence shows the chosen region was benign. That recheck fixes false positives only: a lesion never placed in the top $K$ is never revisited. MagNets (Magnifying Networks for Histopathological Images with Billions of Pixels) learns where to zoom with attention and has no recovery step. Its reported CAMELYON17 numbers show the cost. It reaches 84% AUROC on macro- and micro-metastases, but only 71% (three layers) to 75% (four layers) once isolated tumor cells are included. That gap is exactly the early-dismissal failure mode.

In label space, conformal prediction already controls this kind of cascade failure. CRSVP (Conformal Prediction in Hierarchical Classification with Constrained Representation Complexity) outputs sets of tree nodes that cover the true class at the nominal 90% rate. It also shows that allowing a few disjoint branches ($r=3$) shrinks PlantCLEF sets from about 123 to about 44 classes compared with a single ancestor node. We take that lesson, that coverage is cheap if the set may split across branches, and move it from class labels to image regions. The claim is that a calibrated, multi-branch set of regions to expand should give strictly better lesion recall than a fixed top-$K$ at the same average number of high-resolution pixels read. The gain should be largest on slides with small or multifocal tumors.

Mechanism

We tile the slide as a quadtree over magnification levels $l = 1, \dots, L$. Each node is a region, and its children are the four higher-magnification subregions. A light scorer $s_l$ at each level estimates whether a node's subtree contains tumor. A node is expanded if $s_l$ exceeds a threshold $\tau_l$. Suppose a lesion is missed at level $l$ with probability at most $\alpha_l$, meaning that among nodes whose subtree contains annotated tumor, the fraction pruned is at most $\alpha_l$. By a union bound over levels, the end-to-end miss rate is then at most

$$\alpha \le \sum_{l=1}^{L} \alpha_l ,$$

where $\alpha$ is the fraction of annotated lesions whose leaf patch is never read at full resolution.

We set each $\tau_l$ on a calibration split of slides, treating slides rather than patches as the exchangeable unit, because patches within one slide are correlated. The per-slide loss is the fraction of that slide's lesions pruned at level $l$. Choosing the threshold so that the average of this bounded loss stays below $\alpha_l$ with a finite-sample correction is standard risk-control conformal calibration. The design problem is how to split $\alpha$ across levels. We search over the split on a separate tuning split to minimize expected pixels read, because a coarse level that is nearly blind to micro-lesions should receive most of the miss budget only if its pruning saves a large share of the reads.

Two additions matter beyond calibration. The first is neighbor revival. Metastases are spatially contiguous, and CAMIL (CAMIL: Context-Aware Multiple Instance Learning for Cancer Detection and Subtyping in Whole Slide Images) showed that an 8-neighbor tissue prior sharply improves localization (Dice 0.515 versus 0.103 for TransMIL). So when a full-resolution patch is confirmed positive, its pruned spatial neighbors at every level go back into a priority queue ordered by coarse score. This is the recovery step that SMT's recheck lacks. It is included in the calibration, so the thresholds $\tau_l$ may become looser when revival is active.

The second is a bounded working set. The search runs best-first over this queue. It follows two ideas from memory-bounded search. Felner et al.'s finding (Compressing Pattern Databases) is that coarse guidance held in memory beats exact guidance paged from disk. The structured duplicate detection idea (Structured Duplicate Detection in External-Memory Graph Search) is that only the region adjacent to the active search frontier needs to reside in memory. Accordingly, pruned nodes are kept only as scalar coarse scores. Full-resolution embeddings are cached only while one of their spatial neighbors is still on the frontier, and are otherwise written to disk.

The memory component is a constraint on the design, not the scientific claim. We report peak memory to show that revival does not reintroduce the unbounded growth that beam-style expansion (BeamAggR: Beam Aggregation Reasoning over Multi-source Knowledge for Multi-hop Question Answering) suffers across hops.

Research Program

We build the quadtree search over pre-tiled slides at roughly 1.25x, 5x, 10x and 20x magnification (SMT uses three levels). Scorers are small heads on a frozen patch encoder shared across levels, so that differences between methods come from the search policy rather than the features. We then carry out four pieces of work.

First, we measure how much of the miss rate each level contributes in existing top-$K$ pipelines, as a function of lesion size. This tells us whether early dismissal of small lesions is the dominant failure, as the MagNets isolated-tumor-cell gap suggests.

Second, we calibrate and verify that the realized lesion miss rate on held-out slides tracks the target $\alpha$ across a range of values. We also measure how conservative the union bound is in practice.

Third, we study the split of $\alpha$ across levels and the effect of revival on pixels read.

Fourth, we test transfer across hospitals. Calibration assumes calibration and test slides are drawn from the same distribution, and CAMELYON17's five centers break that assumption. So we report both same-center coverage and leave-one-center-out coverage, and treat any loss of coverage as a finding rather than hiding it.

Baselines

The naive reference points are the two extremes. Uniform full-resolution processing reads every tile with the same encoder and an attention-based multiple-instance aggregator, where multiple-instance learning means classifying a slide from a bag of unlabeled patch embeddings. This version has zero pruning misses by construction and sets the upper bound on pixels and memory. The strongest full-resolution reference in the file is CAMIL, which reports 95.9% AUC on CAMELYON16 and 88.1% on CAMELYON17. We rerun CAMIL on our encoder features rather than relying on its reported numbers, and beating it is not the goal. The target is matching its slide-level AUC within about one point while reading a small fraction of the pixels. The other extreme is a thumbnail-only classifier.

The strongest competitor on the actual question is SMT-style fixed top-$K$ zooming with backward recheck. It is the best published method that already accounts for errors cascading between levels. The file gives only aggregate numbers: a 162.25x average speedup and accuracy within 2% of brute force across five datasets. There is no separate CAMELYON16 AUC for SMT, and no reported lesion-level recall. We therefore reimplement it on the same encoder, levels, training data and tuning budget. We sweep $K$ to trace a curve of recall against pixels read, and compare at matched average pixels per slide. Beating it means higher lesion-level recall, especially for micro-metastases and isolated tumor cells, at equal or fewer pixels, with no drop in slide-level AUC.

We also report MagNets' published CAMELYON17 figures (84% AUROC excluding isolated tumor cells, 71–75% including them) as an external reference. We compare at its stated budget of roughly 3–6 million pixels per slide.

Three ablations isolate the ingredients. The first uses per-level thresholds tuned on validation to hit a target recall without the finite-sample correction or joint allocation, which isolates the calibration. The second removes neighbor revival, which isolates recovery. The third uses a single global threshold at the coarsest level only, which isolates the multi-level split. All variants share encoder, data splits, a 24-hour-per-method tuning budget, and the same hardware memory cap.

Why not just…?

Why not just raise $K$ in top-$K$ zooming? This is the most serious threat. A large enough $K$ drives misses toward zero. The problem is that a fixed $K$ spends the same budget on a slide with one large tumor as on a slide with twelve scattered micro-foci. So matching recall on the hardest slides overpays on all the others. The label-space analogue in CRSVP, where allowing the set to adapt and split shrank it from about 123 to about 44 at equal coverage, suggests that adaptivity is where the savings lie. We do not assume this carries over to regions. The matched-pixel recall curve against the $K$ sweep settles it directly. If the curves coincide, calibration adds only a guarantee and not efficiency, and we will say so.

Why not just tune a threshold on validation data to hit a target recall? Honestly, this is close to what we do, and a well-tuned threshold may match average performance. The differences are that a naive single-level threshold ignores the fact that misses compound across levels, and that it gives no finite-sample statement about the miss rate on new slides. The first ablation above is exactly this alternative. If it matches our method's realized miss rate and pixel count at every target $\alpha$, the contribution reduces to the per-level allocation and revival. The CAMELYON17 cross-center test is where uncorrected thresholds are most likely to drift.

Why not just run flat full-resolution multiple-instance learning with more compute? That approach has no pruning misses at all and is the right choice when compute is free. The motivation for zooming is cost. SMT reports single-slide inference of 2.17–5.65 seconds and under 1% of the embedding storage of brute force. Our aim is to keep that saving while removing its hidden recall cost, so the full-resolution model is our accuracy ceiling rather than a rival.

Why not keep a beam of candidates, as BeamAggR does? A beam guards against committing to one wrong branch, but it grows multiplicatively with depth. Its width also has no link to a target miss rate. Neighbor revival is a narrower recovery step that only reopens regions adjacent to confirmed tumor. We report peak memory and nodes expanded against a width-matched beam to check that this narrowness does not cost recall.

Proposed Experiments

The primary benchmark is CAMELYON16, whose 399 slides carry lesion-level tumor annotations, making per-lesion miss rates measurable. CAMELYON17's 1,000 slides from five centers, with isolated tumor cell, micro- and macro-metastasis cases, test both small-lesion recall and calibration under a change of hospital.

The metrics are lesion-level recall stratified by lesion size, realized miss rate against target $\alpha$ for $\alpha \in \{0.01, 0.02, 0.05, 0.1\}$, slide-level AUC, high-resolution pixels read per slide, peak memory, and wall time. Every result is averaged over at least five random calibration splits.

Three checks decide the proposal. First, the realized miss rate must stay at or below $\alpha$ on same-center test slides across all targets. Failure there refutes the method. Second, at matched pixels, recall on micro-metastases and isolated tumor cells must exceed the best point on the top-$K$ curve. Failure there means calibration buys only a guarantee, not efficiency. Third, removing revival must measurably raise the pixels needed to reach a given $\alpha$. Otherwise the recovery step is unnecessary and should be dropped.