Which H&E measurements transfer across laboratories?

Centre signatures, colorectal tissue classification and same-section scanner repeatability

Parallax Metrology · Working paper · 10 November 2026

Abstract

Across 99,392 PathoROB tiles, adding 42 fixed histology measurements to six RGB summaries improved centre-identification balanced accuracy by 6.4 to 19.3 percentage points. In a separate colorectal experiment, 65 of 88 source-informative measurements retained tissue separation on SurGen, including 15 with a level shift. We then isolated the change of acquisition pipeline by comparing the same glass sections on two scanners. In 118 registered regions from ten veterinary H&E slides, eight of 42 Core42 measurements met prespecified agreement criteria with all three pointwise intervals inside their limits. The strongest agreement was in multiscale orientation diversity. Operational blob density shifted substantially on the same tissue. These findings identify candidate repeatable measurements for these two pipelines; human-tissue and third-scanner transfer remain open.

Introduction

Tissue composition, staining and scanning can each vary across medical centres. Identifying the centre from a tissue image measures their combined footprint. Isolating any one source requires additional controls. PathoROB provides a benchmark for studying how tissue class and medical centre shape representations [1]. A representation-level score, however, leaves open which observable properties of the image carry the centre signature. Channel means and standard deviations offer a simple colour reference. It discards local contrast, spatial arrangement and texture.

The first experiment tests centre identification. The second audits tissue contrast in a separate 218-input colorectal classifier. The third returns to the same Core42 instrument used in the first experiment and compares measurements of registered tissue on two scanners. The three endpoints distinguish centre information, class separation across collections and repeatability within a section.

Experiment 1. Detecting the medical centre

The PathoROB analysis used 60,690 TCGA tiles from 26 centres, 22,402 Camelyon tiles from five centres and 16,300 Tolkach oesophageal tiles from four centres. Images were 256 × 256 pixels. Parallax Core42 v1.0.0 produced 42 deterministic image measurements. Comparators were RGB channel means and standard deviations (six values) and 42 independent Gaussian random features.

Within each collection, a logistic regression predicted centre from each feature set. Five stratified group folds kept cases together for TCGA and slides together for Camelyon and Tolkach (seed 42). These splits evaluate new cases or slides from represented centres. Balanced accuracy is mean class recall, averaged over the five folds. Input scaling and mean imputation used the full collection before splitting. Eight raw feature entries across the three collections required imputation. The centre-prediction comparison was a follow-up to the planned benchmark run.

Table 1. Centre identification on held-out cases or slides

The combined inputs yielded higher observed balanced accuracy in all three collections. In Tolkach, RGB alone yielded higher balanced accuracy than Core42 alone, yet the combination improved on both. The number of centres and their prevalence differ across collections, so the within-collection gain is the relevant comparison. These are descriptive point estimates. No group-resampled uncertainty interval was available for the gains. The Gaussian baseline used 42 features; the analysis did not include an RGB-plus-42-random-features comparator to isolate the effect of adding dimensions.

Centre prediction is a diagnostic for accessible centre information. In these observational cohorts, a centre can also differ in case mix, section preparation and tissue sampling. The results alone cannot attribute the additional information to a scanner or stain. The destination experiment below asks a separate question about tissue calls made by an already fitted classifier.

Experiment 2. Auditing a frozen colorectal classifier

The nine-class colorectal classifier used 218 inputs and fitted readouts developed on NCT-CRC-HE-100K and a development subset of CRC-VAL-HE-7K [2]. The final version reached 94.3% accuracy on 5,380 held-back Kather tiles after five logged evaluations. This is a development-series estimate on collections sharing stain normalization. We applied the frozen classifier, including its two colour-triggered preprocessing rules, to 5,664 SurGen tiles from 177 patients [3]. Each 512-pixel tile supplied four 224-pixel crops. The largest mean class probability determined the tile call.

Two blinded AI readers had independently characterized a subset of 168 fields. They agreed on a predominantly glandular interpretation in 47 and a fibrous interpretation in 64. On those 111 fields the classifier returned tumour or normal mucosa in 7/47 glandular fields (15%) and stroma or muscle in 57/64 fibrous fields (89%). The difference describes agreement with readers. SurGen supplied no reference tissue-class diagnosis for these fields. Applying a prespecified Macenko normalization to the NCT target increased glandular agreement from 7 to 15 of 47 and reduced fibrous agreement from 57 to 41 of 64. Overall agreement fell from 64/111 to 56/111. The same fields were used in both comparisons.

Reader glandular · classifier stroma or muscle

Reader fibrous · classifier stroma or muscle

Measuring separation and level shift

The feature audit excluded three non-tissue inputs and used only crops that bypassed classifier recolouring. We averaged measurements over ungated crops within each field. Six glandular fields had no eligible crop, leaving 41 glandular and 64 fibrous fields. For each measurement we calculated source AUC for 14,317 tumour versus 23,982 stroma or muscle tiles, then destination AUC for glandular versus fibrous fields. The audit plan fixed source eligibility at an absolute distance of at least 0.15 from AUC 0.50. Retained separation required the same direction and a destination distance of at least 0.10. For each tissue group, the median change was divided by the corresponding source IQR/1.349. The larger absolute change determined whether the level shift reached one robust source standard deviation.

Table 2. Transport of individual measurements

Selected families are shown. Of 215 measurements examined, 127 fell below the source threshold. The eight retained stain-channel textures described haematoxylin. For example, coarse haematoxylin contrast yielded source and destination AUCs of 0.877 and 0.887. Coarse eosin contrast moved from 0.671 to 0.397, reversing direction. The family counts describe correlated measurements and do not constitute independent replications.

Experiment 3. Measuring the same tissue on two scanners

On BIFOCAL [4], five canine mammary tumour sections and five feline lymph node sections were each scanned on a Hamamatsu NanoZoomer S60v2 and an Olympus VS200. Each case supplied 12 matched field pairs. The native spacings were 0.22982 and 0.136908 µm/pixel respectively. Same-section pairing holds the specimen and stain preparation fixed.

CMT-HE-01_f1

Hamamatsu S60v2

Olympus VS200

FLN-HE-01_f1

Hamamatsu S60v2

Olympus VS200

Focus selection used the median estimated best focus among reliable patches at a common resolution, rounded to an acquired half-micron plane. Landmark-based similarity registration required at least 12 inliers, at least 25% inliers, coverage in both dimensions, a median inlier residual ≤0.5 µm and scale within 3% of the file calibration. All 120 fields registered. Two Olympus crops from one lymph node slide fell below the 20% brightness-based tissue threshold. All 42 measurements were available on the remaining 118 pairs. Median per-field inlier residual was 0.187 µm. One example field had been inspected for feasibility before the rules were frozen.

Agreement under the frozen rules

Each slide contributed 10% of the weight. Lin’s concordance correlation coefficient (CCC) measures closeness to the equality line. We required overall CCC ≥0.75, CCC after centering each scanner within slide ≥0.50, and an absolute mean scanner shift ≤0.50 Hamamatsu field SD. The sign of the shift below is Olympus minus Hamamatsu. Intervals resample slides within tissue group (3,000 draws). They are separate 95% intervals for each statistic; the decision does not provide joint or panel-wide 95% coverage.

Table 3. Eight measurements met all three limits throughout their intervals

Measurement	Overall CCC [95%]	Within-slide CCC [95%]	Shift, SD [95%]
gap_theta	0.950 [0.936, 0.958]	0.947 [0.934, 0.957]	-0.057 [-0.150, 0.037]
b_theta	0.937 [0.904, 0.952]	0.932 [0.904, 0.950]	-0.083 [-0.179, 0.009]
omega	0.888 [0.805, 0.931]	0.881 [0.804, 0.930]	0.065 [-0.027, 0.148]
g1_curv	0.873 [0.821, 0.949]	0.857 [0.794, 0.943]	-0.020 [-0.058, 0.036]
g4_s1	0.961 [0.953, 0.966]	0.964 [0.960, 0.968]	0.106 [0.043, 0.170]
g4_s8	0.989 [0.982, 0.994]	0.991 [0.981, 0.994]	-0.022 [-0.069, 0.035]
g4_s32	0.985 [0.978, 0.990]	0.986 [0.977, 0.991]	-0.050 [-0.067, -0.032]
g4_slope	0.983 [0.975, 0.988]	0.984 [0.975, 0.989]	-0.065 [-0.083, -0.049]

The four G4 measurements describe orientation diversity at three smoothing scales and the change across scales. Their overall CCCs ranged from 0.961 to 0.989. The other four were gap-field and baseline orientation organization (gap_theta and b_theta), curvature of the multiscale centroid path (g1_curv), and disagreement among mask centroids (omega, Ω). Ω retained within-slide concordance of 0.881, so its agreement also held among fields of the same slide.

Three additional measurements passed on point estimates only: k_xp, b_sdi and b_xp. The remaining 31 missed at least one point cutoff. Operational nuclear-blob density had overall CCC 0.085 and shifted +4.18 reference SD; contour tangent-change density (g3_infl) had CCC 0.042 and shifted −5.64 SD. The blob count is an image-derived proxy. A difference of this size across repeated scans cannot represent a change in the number of cells on the glass.

All other measurements and their intervals

Point criteria met; intervals cross a limit (3)

Measurement	Overall CCC [95%]	Within-slide CCC [95%]	Shift, SD [95%]
k_xp	0.760 [0.609, 0.815]	0.736 [0.602, 0.794]	0.029 [-0.091, 0.131]
b_sdi	0.819 [0.698, 0.860]	0.844 [0.719, 0.886]	-0.244 [-0.378, -0.187]
b_xp	0.819 [0.691, 0.873]	0.807 [0.697, 0.862]	-0.047 [-0.163, 0.044]

At least one point criterion missed (31)

Measurement	Overall CCC [95%]	Within-slide CCC [95%]	Shift, SD [95%]
gap_rv	0.649 [0.473, 0.782]	0.495 [0.302, 0.655]	0.226 [-0.086, 0.499]
gap_delta_vs_baseline	0.171 [0.051, 0.346]	0.443 [0.248, 0.584]	-2.219 [-3.410, -1.539]
gap_ds	-0.009 [-0.023, 0.010]	0.001 [-0.020, 0.088]	12.951 [6.153, 25.550]
blob_density	0.085 [-0.000, 0.097]	0.085 [-0.030, 0.149]	4.183 [2.183, 6.175]
blob_size_mean	0.120 [0.034, 0.249]	0.243 [0.006, 0.411]	1.539 [1.004, 2.401]
blob_size_cv	0.083 [-0.025, 0.251]	0.085 [-0.048, 0.249]	1.329 [0.676, 3.979]
k_rv	0.248 [0.051, 0.423]	0.271 [-0.005, 0.450]	-0.631 [-1.750, 0.194]
k_mu	0.218 [0.056, 0.418]	0.561 [0.331, 0.721]	-0.786 [-1.382, -0.078]
k_sdi	0.750 [0.463, 0.863]	0.778 [0.503, 0.885]	-0.065 [-0.220, 0.219]
k_rho	0.229 [0.028, 0.408]	0.225 [-0.011, 0.391]	0.675 [-0.136, 1.792]
k_theta	0.365 [0.166, 0.546]	0.479 [0.280, 0.638]	-0.664 [-1.012, -0.341]
k_ds	0.075 [0.029, 0.117]	0.012 [-0.172, 0.213]	4.220 [2.802, 6.177]
b_rv	0.438 [0.222, 0.652]	0.536 [0.357, 0.668]	1.064 [0.697, 1.545]
b_mu	0.532 [0.403, 0.613]	0.752 [0.696, 0.807]	0.959 [0.724, 1.252]
b_rho	0.429 [0.219, 0.642]	0.519 [0.342, 0.637]	-1.060 [-1.560, -0.688]
b_ds	0.133 [0.009, 0.251]	0.396 [0.233, 0.502]	2.449 [1.815, 3.247]
gamma	0.476 [0.329, 0.595]	0.564 [0.410, 0.742]	0.950 [0.828, 1.199]
delta_r	0.340 [0.167, 0.545]	0.380 [0.253, 0.479]	-1.345 [-2.103, -0.782]
beta	0.459 [0.247, 0.657]	0.520 [0.276, 0.716]	0.940 [0.370, 2.130]
g1_len	0.572 [0.536, 0.601]	0.744 [0.728, 0.761]	-0.631 [-0.701, -0.584]
g2_V	0.325 [0.159, 0.535]	0.794 [0.654, 0.875]	1.647 [1.075, 2.980]
g2_n	0.327 [0.121, 0.466]	0.350 [0.135, 0.495]	-0.278 [-0.356, -0.196]
g2_AR	0.218 [0.011, 0.448]	0.223 [0.015, 0.469]	-0.264 [-0.337, -0.181]
g2_cut	0.303 [0.151, 0.505]	0.651 [0.457, 0.830]	1.540 [1.050, 2.446]
g3_kvar	0.027 [-0.007, 0.070]	0.002 [-0.032, 0.045]	-5.477 [-13.100, -3.282]
g3_infl	0.042 [0.013, 0.081]	0.033 [0.001, 0.078]	-5.640 [-14.545, -3.381]
sci	0.430 [0.163, 0.590]	0.503 [0.208, 0.721]	0.936 [0.848, 1.083]
vcli_g	0.159 [0.032, 0.264]	0.247 [0.138, 0.384]	-1.600 [-2.355, -0.938]
rc_f	0.694 [0.559, 0.760]	0.685 [0.569, 0.763]	-0.114 [-0.304, 0.119]
rc_s	0.726 [0.618, 0.808]	0.704 [0.580, 0.806]	-0.216 [-0.407, -0.074]
drc	0.557 [0.316, 0.842]	0.569 [0.351, 0.869]	-0.177 [-0.412, 0.161]

Degenerate bootstrap resamples were excluded for blob_size_cv (2,760 valid draws) and g2_n (2,795); all other axes used all 3,000. Decisions use unrounded values.

Download the 42-axis scanner results (CSV)

Discussion

A level shift and a change in tissue ordering have different practical consequences. A measurement that retains its ordering across tissue classes while its level moves may support a recalibrated decision threshold. A threshold shift on a fixed increasing scale leaves the AUC unchanged. Cases near the audit cutoffs may change category under resampling. This distinction was visible in individual stain-channel measurements even though their classifier issued a single tissue call. The mixed normalization result also shows why a proposed fix should be checked in each tissue group: an aggregate improvement can conceal a loss of fibrous agreement.

Centre recognition in these PathoROB analyses suggests that a compact set of explicit measurements can complement routine colour summaries. The SurGen comparison illustrates how to describe a classifier’s operating range by tissue group and measurement family. The present data do not identify the contribution of any one field to an individual classification error. The per-measurement AUCs describe marginal associations. Several inputs change together, and source tumour versus stroma/muscle and destination glandular versus fibrous tissue are only approximate counterparts.

The paired experiment adds a control missing from the collection comparisons: a large measured shift occurred while the glass section stayed fixed. The agreement of orientation-related measures here and their retention of tissue separation in the colorectal audit motivate follow-up, but the feature sets, endpoints and tissues differ. These experiments do not identify which scanner effect, if any, caused the SurGen discordance.

Limits and next test

SurGen slides were unnormalized, whereas the Kather collections shared a stain-normalization procedure. The comparison combines that rendering difference with differences in acquisition, biological composition, sampling unit and labelling. Restriction to concordant reader fields and exclusion of recoloured crops narrow the population represented by the feature audit. Destination AUC intervals used 2,000 stratified field-level bootstrap draws. Fields from the same patient may be correlated, so these intervals can understate uncertainty. The audit reused reader fields previously examined for classifier agreement. The transport categories are descriptive thresholds, and estimates near a boundary can change category with modest sampling variation. PathoROB input scaling also used held-out covariate distributions. These experiments establish neither clinical utility nor cross-hospital correction.

The paired scans establish how the combined acquisition pipelines differed on these ten sections. Tissue type and species coincide in this dataset. Focus selection, resampling, compression and optical properties can each contribute to the observed differences. A decisive replication would use independently selected human colorectal specimens and a third scanner at matched physical scale. A separate controlled experiment is needed to attribute effects to staining or to any one acquisition step.

Data and analysis provenance

This paper uses image-only records from PathoROB, Kather, SurGen and BIFOCAL. The scanner analysis plan was frozen before Core42 extraction. In the second pass, all 236 retained crop hashes and pixel hashes were checked, and all 42 sets of point estimates and bootstrap interval endpoints were recalculated from paired measurements with direct equal-slide weighting. The public numerical table below records the complete scanner result.

Related project records: Colorectal classifier development and two prespecified transport tests.

References

[1] Towards robust foundation models for digital pathology. Nature Communications (2026). 10.1038/s41467-026-73923-2

[2] Kather JN, Halama N, Marx A. 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo (2018). 10.5281/zenodo.1214456

[3] SurGen: 1020 H&E-stained whole-slide images with survival and genetic markers. Dataset record, BioImage Archive. 10.6019/S-BIAD1285

[4] BIFOCAL: Bi-Scanner Focus Dataset, v1.0. Pinned data release and scanner documentation. Image data licensed under CC BY 4.0.