Where Pre-Annotation Misleads: Spatial Entropy and Anchoring in Mitosis Labeling

Pathology datasets are now often built by having a detector propose candidate cells and asking experts to accept or reject them. This is faster than labeling from scratch, but it quietly changes the task. The expert stops searching the tissue and starts checking the machine's list. Missed cells the machine never proposed tend to stay missed, and plausible false alarms tend to be accepted. We propose to test a specific, falsifiable hypothesis about where this damage happens. It concentrates in tissue regions whose cellular layout is spatially mixed and disordered, which is exactly where searching by eye is hardest. The degree of mixing can be measured cheaply with spatial entropy, a Shannon-entropy score computed over the kinds of cells that neighbor each other in a region. If the hypothesis holds, an entropy map lets an annotation tool hide machine suggestions only in the disordered regions. Experts would then label those regions from scratch, keep most of the speedup elsewhere, and stop feeding the model's blind spots back into its own benchmark.

The third ingredient is model specialization. Pathology foundation models are pretrained on huge collections of tissue images, and they now make strong mitosis detectors. Their mistakes are also likely to be more convincing than those of older networks, because they fire on true look-alike cells. We therefore also ask whether a specialized pre-annotator anchors experts more deeply than a generic one at similar accuracy. In practice, it may be precisely the high-entropy regions where it does so.

Core Claim

We make one claim with two parts. First, after controlling for the detector's own confidence and for cell density, the local spatial entropy of a region predicts two kinds of anchored error: accepted false-positive suggestions and missed mitoses the detector failed to propose. Second, gating suggestions by entropy recovers from-scratch label quality at a lower hiding budget than gating by detector confidence or hiding at random.

The distinction from confidence gating is the point of the proposal. Detector confidence measures how hard a region is for the model. Entropy measures how hard the scene is for a human searching it. Anchoring is a human failure, so the second quantity may carry information the first does not. If it does not, the experiments below will show that, and the correct recommendation becomes confidence gating.

Relation to Prior Work

Three studies establish that the problem exists in pathology but do not locate it spatially. Stuck on Suggestions measured automation bias in 28 pathologists estimating tumor cell percentage. It found a 7% rate of experts overturning their own correct judgments (38 of 560) and an AI weight-of-advice coefficient of 0.44 against 0.55 for the experts' own judgment. It also found that confidence during assisted work predicts more over-reliance. That study concerns a single aggregate proportion per patch, not the localization of individual cells.

Are pathologist-defined labels reproducible? re-labeled the TUPAC16 mitosis set with RetinaNet suggestions. It raised the count from 1,552 to 1,999 mitoses, added 10,483 hard negatives, and lifted detector F1 from 0.549 to 0.735, while acknowledging confirmation bias in the process. The COALA study (COALA) names correction-based ground truth as a source of inflated agreement in CT segmentation. All three motivate our proposal. None measures how bias varies across the image.

The evaluation-inflation mechanism we test comes from NLP. Anchoring and Agreement in Syntactic Annotations showed that a parser evaluated on data corrected from its own output looks 34–49% better in error rate than on from-scratch data. We port that "home-field advantage" measurement to mitosis detection. We also take Berzak et al.'s remedy, which shows suggestions only where different models agree, as a baseline rather than an assumption. They themselves found that different models share blind spots.

Does AI Assistance Preserve or Collapse Disagreement? shows that pre-annotation collapses legitimate disagreement between annotators. We therefore track inter-reader agreement as a quantity that can be inflated, not just as a quality score.

Spatial entropy supplies the method. Density and entropy of immune cells computes Shannon entropy of immune-cell layouts on whole slides and finds that it carries prognostic signal that density misses. Spatial entropy as an inductive bias for vision transformers computes entropy over thresholded connected components of attention maps and uses it as a training loss. Both apply entropy to the tissue or to the model. Neither applies it to the human annotation process.

The research file notes two known weaknesses of spatial entropy. It is blind to spatial arrangement, because plain Shannon entropy ignores which bins neighbor which. It is also sensitive to the choice of bin size. Both constrain our definition below.

Specialization enters through Beyond Classification. It pairs frozen pathology foundation models with light detection heads. Frozen H-optimus-0 with RetinaNet reached in-domain F1 0.7718 against 0.7917 for a fully fine-tuned ResNet50 on MIDOG++, and out-of-domain 0.7349 against 0.7193. Two detectors with near-equal accuracy but different pretraining are exactly the pair needed to ask whether specialization changes how persuasive the errors are. We do not rely on the specialization-beats-scale argument of Lawma, which concerns text. We use only the observation that specialized models are what annotation pipelines will increasingly deploy.

Research Program

Entropy map. We first build the entropy map. A lightweight nucleus classifier labels each nucleus in a high-power field with one of a few classes: tumor, stroma, immune, mitosis candidate, or look-alike. This classifier is separate from the mitosis pre-annotator, so the map is not a restatement of that model's confidence.

To respect spatial arrangement, we compute entropy over pairs of neighboring nuclei rather than over single nuclei. For a window $w$, let $p_{ab}(w)$ be the fraction of nucleus pairs in $w$ that are neighbors in a Delaunay triangulation (a mesh that connects each nucleus to its nearest surrounding nuclei) and have classes $a$ and $b$. The window's co-occurrence entropy is then

$$H(w) = -\sum_{a \le b} p_{ab}(w)\,\log p_{ab}(w).$$

This score is low when like cells cluster together and high when classes are finely intermixed. We compute it at three window sizes and report every downstream result at all three, which addresses the bin-size problem directly rather than hiding it.

Reader study. The second component is a controlled reader study in which experts mark mitoses in fields drawn from MIDOG++ and the TUPAC16 auxiliary set, stratified into entropy terciles. Each reader sees each field under exactly one condition, balanced across readers. The conditions are:

  • no suggestions (from scratch);
  • all suggestions from the H-optimus-0 detector;
  • all suggestions from the ResNet50 detector;
  • suggestions hidden in the top-entropy regions;
  • suggestions hidden in the lowest-confidence regions;
  • suggestions hidden in randomly chosen regions.

The three hiding conditions are matched so that each hides the same fraction of area.

On top of each detector's natural errors, a small, known set of false-positive and false-negative suggestions is injected. This follows the incorrect-advice design of Rosbach et al., and it makes automation-bias rates identifiable per region.

Home-field test. The third component asks whether all of this matters for model evaluation. Detectors are trained on the label sets produced under each condition and scored twice: on a held-out set labeled from scratch by a separate panel, and on held-out sets labeled with each detector's assistance. The gap between the two scores is the home-field inflation.

Baselines

Unassisted annotation. The naive reference is annotation from scratch. It is the quality ceiling we are trying to recover and the throughput floor we are trying to beat. It is run with the same readers, fields, and interface, with suggestions simply absent.

Full pre-annotation. The second reference is full pre-annotation from the specialized H-optimus-0 + RetinaNet detector. It is the strongest existing detector-backbone configuration in the file out-of-domain (F1 0.7349; Beyond Classification) and the setup a lab would realistically deploy. It is the throughput ceiling. The ResNet50 detector, with in-domain F1 0.7917 in the same paper, is run alongside it under identical conditions to isolate the effect of specialization on anchoring.

Confidence gating. The strongest competitor to our gate hides suggestions wherever detector confidence falls in the lowest band, or where a Monte Carlo dropout estimate of uncertainty is highest. It uses the same hiding budget and the same detector. It is the strongest competitor because it is cheap, standard, and plausibly targets the same regions. Beating it means lower missed-mitosis and false-acceptance rates than confidence gating at an equal hidden-area fraction, with non-overlapping bootstrap intervals over readers.

Ensemble-agreement gating. Following Anchoring and Agreement, this baseline shows only suggestions on which H-optimus-0, UNI, and ResNet50 detectors agree. It is run on the same fields. Compute cost is reported alongside label quality.

Random gating. Hiding randomly at the same budget is the ablation that isolates the entropy signal. If random hiding does as well, the benefit comes from hiding per se, not from where the hiding happens. A second ablation replaces co-occurrence entropy with plain single-class entropy to test whether accounting for spatial arrangement matters.

There are no reported results on Lacuna for gated or region-selective pre-annotation in histopathology. The closest reported bias magnitudes are the 7% automation-bias rate and 0.44 weight of advice of Rosbach et al., and the 34–49% error inflation of Berzak et al. in syntax.

Why not just…?

Why not just gate by detector confidence? This is the real threat. Rosbach et al.'s confidence paradox concerns human confidence, so it does not by itself rule out gating on the model's confidence. The proposal stands or falls on a mixed-effects logistic model of anchored error per suggestion, with reader and field as random effects and entropy, detector confidence, and nucleus density as fixed effects. If the entropy coefficient vanishes once confidence is included, and confidence gating matches entropy gating in the reader study, we will report that entropy adds nothing and recommend confidence gating. One reason to expect a gap is that the anchoring most damaging to evaluation comes from errors the model makes confidently, which confidence gating by construction never hides.

Why not just require agreement among several models? Berzak et al. report that distinct parsers share blind spots that annotators then overlook. Running several billion-parameter pathology encoders per slide is also costly. Both points are tested rather than assumed: the ensemble-agreement condition is scored on the same missed-mitosis metric, with its compute cost reported alongside.

Why not just always label from scratch, or add a second reader? Scratch labeling is the quality reference. The question is how much of it can be kept at what fraction of its time cost, and the study measures time per field under every condition. A second unassisted reader also costs a full second read and does not remove the shared anchor when both readers see the same suggestions, since the video-disagreement study shows that shared suggestions collapse the variance between readers.

Why not just use a better detector? Bertram et al.'s labels were produced with a RetinaNet, and the specialized models in the file are only a few F1 points apart. Improving accuracy shrinks the number of errors but may make the remaining errors more persuasive. The comparison between the H-optimus-0 and ResNet50 detectors in this study tests exactly that possibility.

Decisive Checks

Three results settle the proposal.

First, the entropy coefficient in the anchored-error model must remain positive and significant after controlling for confidence and density, at all three window sizes.

Second, entropy gating must beat both confidence gating and random gating on the missed-mitosis rate at a matched hidden-area budget, while retaining at least half of full pre-annotation's time savings.

Third, detectors trained and evaluated on assisted labels should show a measurable home-field inflation relative to scratch-labeled test data, analogous to Berzak et al.'s finding. That inflation should be concentrated in high-entropy fields and reduced by entropy gating. Separately, if the specialized detector's natural false positives are accepted at a higher rate than the ResNet50 detector's at similar precision, that is direct evidence that specialization deepens anchoring.

A null result on the first check falsifies the core claim. A null result on the third means pre-annotation bias in mitosis labeling is mostly a throughput-versus-recall trade-off, not an evaluation problem.