When Pathologists Disagree, Do the Clones Disagree Too?

Pathologists who outline a tumor on the same tissue slide draw different boundaries. Across 120 board-certified pathologists, agreement on cell annotations reaches a Fleiss' kappa of only 0.43 (Variability Matters). The field has split on what to do about this. One camp treats disagreement as label noise and discards the least consistent annotators. The other treats it as signal about genuinely ambiguous tissue and trains models to reproduce the whole spread of expert opinion (MRNet, QUBIQ). Neither camp has an independent referee, because both judge annotations only against other annotations.

This proposal supplies a referee from molecular data. Spatial transcriptomics measures gene expression at thousands of tiny spots across a tissue section, and copy-number changes inferred from that expression can be used to rebuild the family tree of tumor clones in place. A clone here means a group of cells descended from a common ancestor and sharing its mutations. Doing this well is a dense-sampling problem: there are many spots and few distinguishing genetic events, so many different family trees explain the data equally well. Evolutionary biology has tools for exactly this situation. We will build an uncertainty-aware map of where the molecular tumor boundary lies. We will then test whether the places where pathologists disagree are the places where the clonal boundary is genuinely graded, over and above what image difficulty alone predicts. Finally, we will use the answer to decide, region by region, whether rater disagreement should be kept or discarded when training a segmentation model.

Core Claim

Rater disagreement on histology is a mixture of two things. The first is biological: tissue sitting on a real clonal transition front. The second is human error on tissue that is molecularly unambiguous. The two can be separated spatially if the molecular reference carries its own honest uncertainty. The key methodological point is that a single inferred clone tree, as used by Erickson et al., confuses two sources of apparent ambiguity. One is ambiguity created by picking one tree among many equally good ones. The other is true biological gradation. Only an ensemble over equally good trees can tell them apart.

Our claim is testable. A per-spot "molecular transition index" should predict per-spot rater disagreement beyond what image-based difficulty predicts, where the index is computed over an ensemble of clone trees weighted by how many spots share each genotype. If that holds, then filtering annotators by conformity will disproportionately remove raters whose boundaries track real clone fronts. A training rule that keeps disagreement only where the molecular ensemble supports it should give better calibrated models in transition zones without the noise cost that Variability Matters documents.

Where the Pieces Come From

Erickson et al. motivates the proposal. They inferred copy-number profiles from Visium spatial transcriptomics across more than 120,000 spots in prostate, skin, breast and brain tissue and built clone trees from them. They found that tumor-associated copy-number events such as MYC gains and PTEN losses extend into tissue that one pathologist marked benign. They compared against a single observer, however, used a single tree, and did not ask whether clone fronts explain disagreement among observers. We keep their measurement idea and add both missing pieces: multiple raters, and uncertainty over the tree.

Direction 1 supplies the method for that tree uncertainty. GCtree showed that when many samples share few mutations, the number of times each genotype is observed breaks ties among equally parsimonious trees. Abundant genotypes are better explained as persistent ancestors. GCtree formalizes this with a branching-process likelihood and reports trees measurably closer to the truth by Robinson-Foulds distance, a count of splits that differ between two trees. The history sDAG of Dumm et al. 2023 and Dumm et al. 2025 stores enormous sets of equally parsimonious histories in one compact graph by sharing common sub-histories. It expanded 800 seed trees into $10^{26}$ histories in seconds, and it supports dynamic-programming sums over the whole set. We reuse both, applied to discretized spot-level copy-number genotypes instead of B-cell sequences.

Two known limitations constrain the design. GCtree assumes every new mutation creates a new genotype, and recurrent copy-number events violate that. We therefore also score trees with a reconciliation-style penalty for repeated events, in the spirit of TreeFix-DTL, and report results with and without it. The second limitation is that both methods rely on parsimony seeds, so the true history may lie outside the ensemble. We will check this on simulations.

Research Program

The first component is data. We will obtain Visium sections with matched H&E images, either from the Erickson cohort if the matched images are available or from newly profiled prostate and breast sections. Five pathologists will each delineate tumor regions independently on the H&E, and each will re-annotate a subset weeks later. The repeat annotations give an intra-observer inconsistency floor, which is a direct empirical measure of pure human noise. All contours are rasterized to the 55 µm spot grid, so that each spot receives a rater disagreement score: the fraction of raters labeling it tumor, turned into binary entropy.

The second component is the molecular ensemble. We will infer copy-number profiles per spot as Erickson did and discretize them into event genotypes. We will seed parsimony trees and merge them into a history sDAG. Each tree is then weighted by the abundance likelihood, optionally times the recurrence penalty. Because Visium spots are mini-bulk averages of 5 to 10 cells, abundance is measured on spots rather than cells, and we will model the fraction of normal cells in each spot as a nuisance mixture. From the weighted ensemble, dynamic programming gives each spot a posterior probability of belonging to the tumor lineage and the variance of its position along that lineage. Together these form the molecular transition index.

The third component is the explanatory test. For each spot we will regress rater disagreement on image difficulty and then add the molecular transition index. Image difficulty is measured as the predictive entropy of an ensemble of H&E segmentation networks trained on majority-vote labels, together with simple stain and texture statistics. We will measure the increase in explained variance, with significance assessed by spatial block permutation, because neighboring spots are correlated.

The fourth component is training. We will train H&E-only segmentation models under several label regimes. The new regime, which we call biology-adjudicated labels, keeps each rater's soft vote in spots whose molecular transition index is high. In molecularly unambiguous spots it collapses the label to the molecular call, restricted to spots where the image model is not itself uncertain.

Baselines

The naive reference is a single model trained on pixel-wise majority vote over all five raters. It shows what ignoring disagreement costs.

The strongest competitor on the "disagreement is noise" side is conformity filtering from Variability Matters. That work scored each annotator's agreement with an anchor annotator and trained DeepLabv3+ only on the top 50% or 75%, which beat training on all annotations. We will run it with the same backbone, the same data, and the same hyperparameter search budget as every other regime. Because there is no designated anchor here, the anchor will be the leave-one-out consensus.

The strongest competitor on the "disagreement is signal" side is a deterministic ensemble trained on individual rater masks. The QUBIQ benchmark found this family among the top performers, with average Dice between 0.80 and 0.85 on its CT and MRI tasks. We will also run D-Persona, which learns a latent space bounded by the intersection and union of expert masks. On NPC-170 it reports a generalized energy distance of 0.2359 against 0.4465 for the Probabilistic U-Net. Generalized energy distance measures how well a set of sampled masks matches the set of expert masks; lower is better. These figures come from other modalities. No reported result was found on Lacuna for any multi-rater method on H&E sections paired with spatial clonal data, so all numbers on our data will be new.

Two ablations isolate the key ingredient. The first replaces the abundance-weighted sDAG ensemble with a single maximum-parsimony tree. The second replaces the tree with the raw smoothed copy-number score per spot. Beating the baselines means that biology-adjudicated labels match the soft-label ensemble on generalized energy distance against held-out raters, match or beat conformity filtering on Dice in molecularly unambiguous regions, and show clearly lower calibration error in molecularly defined transition zones.

Why not just…?

Why not just explain disagreement with image difficulty? Blurry staining and low-contrast borders surely drive much of the disagreement, and this is the main threat to the claim. That is why image-based predictive entropy enters the regression first, and only the added explanatory power of the molecular index counts. If that added power is negligible, the core claim fails and we will report it as a negative result. A negative result would still settle the noise-versus-signal dispute in favor of conformity filtering for this tissue type.

Why not just use the copy-number score map without any tree? A smoothed score may locate the tumor front just as well, and if it matches the ensemble index in the explanatory test, we will adopt the simpler signal. The tree should help in one specific case: when a boundary separates two tumor subclones rather than tumor from normal. A scalar score cannot see that boundary, but lineage position can. We will report the explanatory test separately on subclone-subclone borders and tumor-normal borders to check this.

Why not just use one parsimony tree, as Erickson et al. did? With many spots and few copy-number events, many trees tie, as GCtree documents in the analogous B-cell setting. A single arbitrary tree can place a front where the ensemble would call the region uncertain. The single-tree ablation measures directly whether this inflates or deflates the apparent agreement between molecular and rater uncertainty.

Why not just filter annotators and move on? Filtering improves average scores in Variability Matters. It may also systematically remove the raters who mark subtle infiltration. We will test this by computing, for each rater, how well their boundary agrees with the molecular front, and checking whether rater conformity correlates negatively with that agreement.

Decisive Checks

On simulations, we will generate copy-number clone trees with known spatial fronts, mix cells into spots, and synthesize rater noise with a known split between biological and random components. This tests whether the pipeline recovers that split and whether seed-restricted ensembles miss the true tree.

On real sections, three results decide the project. The first is whether the molecular transition index adds significant explanatory power over image difficulty under spatial permutation. The second is whether the disagreement it explains exceeds each rater's own intra-observer inconsistency floor. The third is whether conformity filtering removes the raters most aligned with the molecular front. If all three hold, biology-adjudicated labels resolve the noise-versus-signal tension region by region rather than globally. If none hold, the noise camp has the better argument for this tissue.