Pale space in colorectal cancer
One fixed, training-free number from an H&E slide tracks survival in a public colorectal cancer cohort beyond stage. We tried to break it. This page records what survived, what it might be, which earlier claims it replaces, and what would settle it.
How much of the tumour is pale
H, the highlight mass, is the share of pixels in an image tile whose CIE lightness is at least 0.70. It is computed on the raw scanner image with no stain normalisation, averaged over tiles cut from inside the tumour outline drawn on each diagnostic slide. Nothing is learned: the formula is fixed, and it was not designed for pathology.
On expert-labelled tissue it behaves as a measure of open, pale space: highest in fat and mucus, intermediate in stroma and smooth muscle, lowest in dense tumour epithelium, debris and lymphocytes.
A signal, and four attempts to explain it away
In TCGA colon adenocarcinoma, patients with higher H had shorter disease-specific survival after adjusting for age, sex and stage, with tissue source site as strata: hazard ratio 1.57 per standard deviation (bootstrap 95% interval 1.25–2.02). The first three checks below were planned in writing before they were run; the fourth was a later, outcome-free look.
| Could it be… | Check | H's hazard ratio |
|---|---|---|
| The scanner or file route? | TCGA colon slides come through three production routes; 14 patients in this cohort have only a low-resolution derivative slide. Excluded them; stratified by site and route. (Model fitted jointly with two companion image measures on 175 patients, so the starting value differs from 1.57.) | 1.43 → 1.50 / 1.49 |
| Stroma or mucus as an AI classifier sees them? | A tissue classifier (frozen DINOv2 features with a logistic layer trained on nine expert tissue classes) estimated each patient's stroma and mucus. Added both to the model. | 1.57 → 1.51 (log HR −8.7%) |
| Where the tiles sit? | Tiles near the outline edge are paler. Measured H only on tiles at least two tiles inside the outline. | 1.56 → 1.54 |
| Simply a mucinous tumour type? | The 16 of 160 patients recorded as mucinous adenocarcinoma have higher H (AUC 0.70), but H varies as widely without them. | explains part, not most |
Limits: one cohort, which has been analysed many times; 34 deaths; one image measure among many examined in this programme. This is a lead, not a finding.
Four readings, one of them already graded by pathologists
Immature, myxoid stroma
Pathologists grade the desmoplastic reaction at the tumour front as mature, intermediate or immature; immature stroma is loose, pale and myxoid. It is an independent prognostic factor alongside T and N stage [1], with confirmations in prospective and recent cohorts [2, 3], though one study found it in stage II only [4]. A classifier trained on nine tissue classes cannot see it, since none is myxoid stroma; the stroma it does see runs against H (ρ −0.44). Caveat: the reaction is graded at the invasive front, while H is high throughout the tumour.
Mucin
Extracellular mucin is pale, and mucinous tumours have higher H. But recorded mucinous histology covers 10% of patients and explains only part of H's spread.
Retraction and handling
Clefts and open spaces from fixation and cutting are pale. If these drive H, part of it measures how the tissue was processed: the care of the artifact rather than the tumour.
Stain intensity
A tonal measure reads lightly stained or faded slides as higher. Comparing within each hospital reduces this; it does not remove it.
What this replaces in our earlier claims
Earlier versions of this work reported a hybrid structural composite with a disease-specific hazard ratio of 1.872 after Macenko stain normalisation, and withdrew H as a "staining proxy" because it weakened after normalisation. Audits in August and November 2026 found:
| Earlier claim | What we found |
|---|---|
| Composite DSS HR 1.872, OS HR 1.687 | Not reproducible from any surviving file, and retracted. Re-runs give 1.39–1.69 (DSS) depending on model settings; with a corrected normaliser the overall-survival estimate is 1.43. The direction holds; the site-controlled inference does not. |
| Results "after Macenko normalisation" | The normaliser used was degenerate: its two stain vectors were 0.17° apart, from an SVD sign ambiguity (a corrected estimate gives about 14°; textbook vectors are 39° apart). It did not separate haematoxylin from eosin. |
| H withdrawn as a staining proxy | The withdrawal rested on that normaliser. Measured raw, H holds within each hospital and production route. Whether it is staining is now candidate 4, to be tested directly. |
| Structural deviation Ω free of hospital effects | Retracted in August 2026: every structural axis carries a site signal. |
Along the way we found that 25 TCGA colorectal diagnostic slides are low-resolution derivatives whose stated pixel spacing in the NCI Imaging Data Commons is about half the true value: technical note.
An outside cohort and a pathologist's eye
An independent cohort. A one-page test was written and frozen before any outcome was seen: H on tumour tiles in an independent Scottish colorectal cohort (SurGen), direction fixed in advance, with a stated pass line. It closed in November 2026 without a survival estimate. The test required at least four confirmed tumour tiles in 135 of 168 patients; sampling reached 127, so the pre-agreed coverage check stopped it before outcomes were opened. H is untested there, neither supported nor contradicted (technical note). A fair test needs tumour regions outlined by a pathologist, as the TCGA tiles were.
A pathologist's read. A blinded set of high- and low-H cases is ready for a pathologist to describe and sort before learning what H is. If H tracks the desmoplastic-reaction grade, it would be a one-number, training-free proxy for a feature pathologists already value but grade by eye. If you are a pathologist and would look at 16 cases for twenty minutes, please get in touch.
- Ueno H, et al. Desmoplastic pattern at the tumor front defines poor-prognosis subtypes of colorectal cancer. Am J Surg Pathol 2017.
- Akimoto N, et al. Desmoplastic reaction, immune cell response, and prognosis in colorectal cancer. Front Immunol 2022. doi:10.3389/fimmu.2022.840198
- Impact of stromal maturity and proportion on prognosis and immune landscape in colorectal cancer. 2025 (PMC12777758).
- Desmoplastic reactions and epithelial-mesenchymal transition proteins in stages II and III colorectal cancer. Virchows Arch 2022;480:793–805. doi:10.1007/s00428-021-03266-4
- Kather JN, et al. 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo 2018; tiles of TCGA-CRC-DX slides, Zenodo 2020.
- TCGA colon adenocarcinoma diagnostic slides and clinical data: Genomic Data Commons; TCGA Pan-Cancer Clinical Data Resource.
Analysis by Russell Parrish with Claude (Anthropic). Measured facts are reported with their limits; the hypotheses above are hypotheses.