Russell Parrish · analysis carried out with AI assistants · Version 1, 10 November 2026
Abstract. A colorectal tissue classifier built from 218 fixed, named image measurements and a fitted readout reaches 94.3% on held-back tiles of the Kather benchmark [1], whose collections share one stain normalisation. We ran it, frozen and unchanged, on 22,656 crops from 177 patients of SurGen [2], a collection from another laboratory distributed without normalisation. SurGen has no tissue labels, so the comparison is with two blinded AI readers on the 111 fields they agreed were gland or fibrous tissue. The classifier matched the readers on 57 of 64 fibrous fields (89%) and 7 of 47 gland fields (15%), and called 0.6% of all SurGen tiles tumour. Because every input to the classifier is a named measurement, the boundary can be located field by field. Of 88 fields that separate tumour from stroma and muscle in the training collection, 50 held on SurGen, 15 kept their separation at a shifted level, and 23 broke. Among those, fibre-orientation fields held without exception (17 of 17), as did haematoxylin-channel textures (8 of 8). Fields that depend on how dark the stain is broke or shifted: optical density, shadow mass, nuclei counted per tissue area and eosin-channel texture. A prediction written before measurement, that Macenko normalisation to the training target would repair the tumour reading without harming the fibrous reading, was not supported (gland 15 of 47, fibrous 41 of 64), although it restored most individual field levels. Two normalisations each moved one tissue boundary and broke another. The per-field map is offered as a guide to which kinds of measurement can be expected to carry between laboratories without normalisation.
1. Question
A classifier that does well on a benchmark's held-back tiles may still depend on properties shared by the whole benchmark. The Kather colorectal collections [3] were prepared in one laboratory and distributed after one Macenko normalisation, so a good external score within them does not test a change of stain rendering. This note asks two questions of a classifier whose inputs are all named, fixed measurements:
How does it read slides from another laboratory, with nothing refitted?
Which of its 218 input measurements keep their meaning there, and which do not?
The second question is the reason to use this kind of instrument for the first. A learned embedding fails as a whole; a set of named fields can be checked one at a time.
2. Materials
Classifier. The version frozen at the end of the development record in [1]: 218 fixed measurements per crop (tonal, colour and stain, grey-level and stain-channel texture, nuclei, fibre orientation, gland and lumen geometry, layout, and the original structural axes), read by a fitted ensemble of logistic and gradient-boosted readouts trained on NCT-CRC-HE-100K. Two label-free colour gates sit in front of it: a violet gate and an orange gate, each of which recolours the crop before measurement when an unfamiliar rendering is detected. The measurements are deterministic; the readout and the gates' reference colours are fitted to the training collection.
External slides. SurGen [2]: colorectal cancer slides from a single-source Scottish cohort, distributed without stain normalisation. We used the 5,664 fields (512 × 512 pixels at about 0.50 µm per pixel, 32 per patient, 177 patients) already sampled across all tissue for a separate study [4]. Each field gives four non-overlapping 224-pixel crops, the scale of the Kather tiles, so 22,656 crops. The field call is the largest mean class probability over its four crops. No SurGen outcome, label or clinical file was read.
Reference labels. SurGen has no tissue-class annotation. Two AI readers, blinded to patient, measurements and each other, had labelled 168 of these fields by dominant appearance for the separate study. They agreed on gland or epithelium for 47 fields and on fibrous or fibromuscular tissue for 64. These 111 fields are the comparison set. Agreement with them is agreement between two instruments, not accuracy.
Plans. The classifier run, the per-field map and the normalisation prediction each had a plan written and hashed before computation.
3. The classifier on another laboratory's slides
Fields both readers agreed on
Fields
Classifier in the same group
Fibrous (stroma or muscle)
64
57 (89%): stroma 39, muscle 18
Gland or epithelium (tumour or normal)
47
7 (15%): normal 5, tumour 2. The rest: stroma 14, muscle 11, debris 11, mucus 4
Both (the planned statistic)
111
64 (58%)
Across all 5,664 fields the calls were muscle 36.8%, stroma 34.7%, debris 21.1%, mucus 4.0%, normal 1.9% and tumour 0.6%. One patient had four or more fields called tumour. The miss is in tumour, not in fibrous tissue (Figure 1).
Figure 1. Top: eight fields both readers called gland, which the classifier called stroma or muscle. Bottom: eight reader-agreed fibrous fields, read as fibrous. Images from SurGen [2].
The colour gates. The violet gate fired on 1,506 of 22,656 SurGen crops (6.6%), against 363 of 100,000 tiles in the training collection (0.36%), about 18 times the training rate. It was built to catch one single-hue rendering; here it responds to a general difference in rendering.
Is it colour? (exploratory, after the result). SurGen gland crops are bluer than training tumour tiles: median non-nuclear b* of −21.3 against −14.2, with similar a* (32.0 against 33.5) and somewhat higher lightness (67.4 against 61.8). Recolouring every crop of the 111 fields by brightness rank onto the training collection's colours raised gland agreement to 54% of crops and lowered fibrous agreement to 60%: stroma almost disappeared from the calls, because the recolour erases the stain balance that separates stroma from muscle.
4. The per-field map
4.1 Method
For each of the 215 tissue fields (three image-size constants excluded) two separations were computed:
Training collection. AUC for tumour (14,317 tiles) against stroma and muscle (23,982 tiles) in NCT-CRC-HE-100K.
SurGen. AUC for reader-agreed gland fields against reader-agreed fibrous fields, with a 95% interval from 2,000 stratified bootstrap draws. Only crops that neither colour gate had recoloured were used, which left 41 gland and 64 fibrous fields (410 crops).
A level shift was also computed: the SurGen median minus the training median for the matching tissue, in robust standard deviations of the training class. Categories were fixed in the plan:
Category
Rule
Fields
Not informative here
Training AUC within 0.35 to 0.65
127
Holds
Same direction, SurGen AUC at least 0.10 from 0.5, shift under 1 SD
50
Separates, level shifted
Same, with shift of 1 SD or more
15
Breaks
SurGen AUC within 0.10 of 0.5, or reversed
23
Of the 88 informative fields, 57% hold, 17% keep their separation at a shifted level and 26% break.
Figure 2. One dot per field. On the diagonal a field separates the two tissues as well on SurGen as in the training collection. Dot size is the level shift.
4.2 By family
Figure 3. Families sorted by the share of informative fields that hold.
Family
What happened
Fibre and orientation
17 of 17 informative fields hold: fibre coherence, anisotropy, orientation entropy and persistence, curvature, multiscale coherence.
Stain-channel texture
All 8 informative haematoxylin-channel textures hold. The 3 informative eosin-channel textures break.
Grey texture
8 hold, 3 shift, 4 break. Coarse contrast and entropy hold; fine-scale fields shift or break.
Nuclei
3 hold, 4 break. Orientation-type nuclear fields hold. Nuclei per tissue area loses its separation (AUC 0.83 to 0.54); three count and shape fields reverse.
Colour and tonal
Optical density (median and 90th percentile) and shadow mass (0.82 to 0.52) break. Haematoxylin median, a* and chroma hold.
Gland and lumen
22 of 23 are not informative even in the training collection (AUC 0.47 to 0.65); one breaks.
4.3 Where SurGen sits
Figure 4. SurGen median minus training median for the matching tissue, in robust SD of the training class. E: reader-agreed gland fields against training tumour. F: reader-agreed fibrous fields against training stroma and muscle.
Bluer and lighter. b* median −5.7 SD on gland fields and −2.9 SD on fibrous fields; L* +1.3 and +0.7 SD.
Less optical density, most at the dark end. 90th percentile −5.0 and −4.6 SD; median −2.6 and −1.7 SD; shadow mass −2.1 and −1.3 SD.
Haematoxylin median within 1 SD (+0.3, +0.7). The typical nuclear stain value is comparable; the darkest pixels are not.
Lumen geometry far lower on gland fields. 14 of 23 gland and lumen fields sit more than 1 SD below training tumour (eccentricity −6.1 SD). Whether this is fewer lumens, a detector threshold or field content is not separated here.
Orientation unchanged. Fibre orientation entropy and the structural axis g4_s1 sit within 0.4 SD.
4.4 What the readout leans on (exploratory)
For the logistic half of the ensemble, the share of absolute tumour-coefficient weight by category was: not informative here 53%, holds 28%, shifted 10%, breaks 9%. The readout leans mostly on fields that matter only in combination, so the tumour miss is not explained by a few broken fields carrying the weight. The gradient-boosted half has no comparable coefficient.
5. A frozen prediction: Macenko normalisation to the training target
The level shifts point to stain density first and hue second, which a stain-concentration normalisation addresses directly. The prediction was written and hashed after the map and before any normalised crop was measured: after Macenko normalisation [5] to a target fitted on 400 random training tiles, the unchanged classifier would agree with the readers on at least 19 of 47 gland fields and at least 52 of 64 fibrous fields. The normaliser is the sign-corrected one described in [6].
Reader-agreed fields
Raw
Hue recolour (exploratory, crops)
Macenko
Predicted
Gland (47)
7 (15%)
54%
15 (32%)
at least 19
Fibrous (64)
57 (89%)
60%
41 (64%)
at least 52
Figure 5. Agreement with the readers before and after normalisation. Dotted marks: the thresholds predicted for the normalised run.
Verdict: not supported. Neither threshold was met. No tile was refused by the normaliser and no crop tripped a colour gate afterwards. Fibrous fields were called stroma 37, debris 19, muscle 4 and other 4; muscle nearly disappeared and debris absorbed fields from both groups.
What normalisation did to the fields, by the same map rules:
Levels mostly restored. Optical density median and 90th percentile moved from −2.6 and −5.0 SD to within 1 SD; b* from −5.7 to −2.2 SD on gland fields.
Map improved. Breaks fell from 23 to 14 and holds rose from 50 to 53. Shadow mass regained its separation (AUC 0.52 to 0.84).
Haematoxylin overshot. Haematoxylin median moved from +0.3 to +1.9 SD on gland fields, and its separation weakened (0.80 to 0.63).
Lumen geometry did not move. Lumen eccentricity on gland fields stayed at −6.1 SD. That gap is not a stain effect.
Orientation stayed put. Fibre orientation entropy and g4_s1 remained within 0.5 SD and kept their separation.
6. Reading
Operating window of the classifier. It transfers within collections that share its training rendering. On raw-stained slides from another laboratory it reads fibrous tissue well and tumour badly.
Operating window per field. Structure that does not depend on stain intensity carried between these two laboratories: fibre orientation, haematoxylin-channel texture, coarse texture. Fields read from how dark the stain is did not: optical density, shadow mass, counts that rely on a darkness threshold, eosin-channel texture.
Normalisation repaired fields, not the fitted readout. Two different normalisations (hue rank and stain concentration) each moved one tissue boundary and broke another. A readout fitted over 218 fields did not follow a normalisation that only partly restores them.
The violet gate as a detector. A gate built for one rendering fired at 18 times its training rate on SurGen, so it may serve as a signal that a collection is outside the rendering the classifier knows. That use is suggested by one collection and is not tested here.
For readers who apply tools trained on the Kather collections to unnormalised slides from elsewhere, the map gives a concrete expectation: orientation and coarse-texture measures are the likeliest to carry over, and intensity-thresholded counts the least.
7. Limitations
One external collection. The 111 comparison fields were labelled by two AI readers of one model family, not by a pathologist. Agreement is consistency between instruments.
Reader fields are whole 512-pixel fields labelled by majority tissue, so gland fields contain stroma. This pulls SurGen separations towards 0.5 and makes "holds" conservative. Gland fields are not confirmed tumour, and the training contrast (tumour against stroma and muscle) is comparable to the SurGen contrast, not identical.
The 168 reader fields were used twice (classifier agreement, then the map). The map is descriptive, not a test.
Implementation correction: the map plan named the reader groups by intent but pointed at the wrong column of the comparison file. The first run selected fields by the classifier's own call; it was discarded and corrected before any field-level result was read. Its family totals were printed and seen, and are not used.
The hue recolour and the readout-weight analysis were run after results and are exploratory. The Macenko prediction was frozen, and it failed.
The lumen-geometry gap and the residual tumour miss are unexplained.
Data and code availability
Hashed plans, scripts, the per-field table (separations with intervals, shifts and category for all 215 fields), crop-level predictions and the lab books are held by the authors (folders SURGEN_CRC_CLASSIFIER_20261107 and SURGEN_FIELD_TRANSPORT_20261107; this note in FIELD_TRANSPORT_NOTE_20261110) and are available on request. Only public data were used.
References
Parrish R. Fixed, named H&E measurements on the external split of the Kather colorectal benchmark: a development record from 85.7% to 94.3% on held-back tiles. Technical note, 2026. https://www.parallaxmetrology.com/External_CRC_Benchmark_NOTE.html
Myles C, Um IH, Marshall C, Harris-Birtill D, Harrison DJ. SurGen: 1020 H&E-stained whole-slide images with survival and genetic markers. GigaScience 2025;14:giaf086. doi:10.1093/gigascience/giaf086. Images: EMBL-EBI BioImage Archive, S-BIAD1285.
Kather JN, Halama N, Marx A. 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo, 2018. doi:10.5281/zenodo.1214456
Parrish R. A frozen external survival test of two fixed H&E measurements on SurGen. Technical note, 2026. https://www.parallaxmetrology.com/SurGen_External_Test_NOTE.html
Macenko M, et al. A method for normalizing histology slides for quantitative analysis. IEEE International Symposium on Biomedical Imaging, 2009:1107–1110.
Parrish R. A missing sign check in a Macenko stain normaliser collapsed haematoxylin and eosin into one direction and created a site difference in TCGA colorectal measurements. Technical note, 2026. https://www.parallaxmetrology.com/Stain_Normaliser_NOTE.html
