Parallax Pathology · Technical note
Russell Parrish · analysis carried out with AI assistants · Version 1, 10 November 2026
Parallax Pathology measures H&E images with fixed, named formulas that are not trained on pathology data. In TCGA colon adenocarcinoma (TCGA-COAD), two such measurements were associated with survival after adjustment for stage. That cohort has been analysed many times within this programme, so those associations are leads, not findings. This note records what happened when both were taken to an independent cohort.
U, common persistence. Two fixed colour-opponent response fields (labelled M-cone and L−M in the original instrument) are pooled on a grid and standardised within each tile. U is the mean, over six fixed separations (1, 2, 4, 8, 12 and 16 grid cells), of the normalised lagged product of their sum. Its companion V uses their difference. A patient's value is the mean over eligible tiles. In TCGA the lower-U patients had shorter survival, so the fixed risk score is −U.
H, highlight mass. The share of pixels in a tile with CIE lightness of at least 0.70, measured on the raw image with no stain normalisation. In TCGA-COAD, higher H accompanied shorter disease-specific survival: hazard ratio 1.57 per standard deviation after adjustment for age, sex and stage (bootstrap 95% interval 1.25 to 2.02; 161 patients, 34 deaths) [2].
The question for each was the same: with the measurement, its direction and the analysis fixed in advance, does the association appear in patients the programme has not analysed?
Cohort. SurGen SR386 [1]: 427 colorectal cancer slides from a single-source Scottish cohort, with the authors' official split of 255 training, 84 validation and 84 test patients. Label and split files were pinned to a fixed commit of the authors' repository.
Endpoint. All-cause death within five years of diagnosis (1,826.25 days), following the cohort paper's convention: patients released as alive are censored at five years, and deaths after five years are censored at five years. Three cases with redacted survival fields have no endpoint. This is not the TCGA endpoint (disease-specific survival), so the test is of transport to another cohort and endpoint, not an exact replication.
Sampling. Each slide was read at about 0.50 µm per pixel in 512 × 512-pixel fields. Up to 32 fields per slide were chosen by a fixed rule that spreads them across all tissue found by a simple optical tissue screen. The rule does not look for tumour. In TCGA, by contrast, the archived tiles came from inside tumour outlines drawn on each slide.
Blinding. Image selection and measurement were frozen before any outcome join. The training partition had been modelled earlier in the programme with other measurements; the validation and test partitions had not been analysed against outcomes. Their label files were first opened in this work, to audit endpoint definitions before the protocol was frozen.
Protocol. Concordance (Harrell C) of the continuous score −U in the official test partition. Pairs count when one patient has a strictly earlier death; tied scores get half credit. The interval is the 2.5th to 97.5th percentile of 2,000 patient-bootstrap draws. No model is fitted and no threshold or sign is chosen on SurGen.
Result. All 84 planned patients were included, with 29 deaths and 2,001 comparable pairs. C = 0.460 (0.348 to 0.565). Of 2,688 sampled fields, 2,677 were eligible, with 29 to 32 per patient. An independent implementation reproduced the score and all 2,000 bootstrap values.
These checks were specified in writing after the primary result and before each was computed. They are exploratory, and none changes the primary.
| Possible explanation | Check | Finding |
|---|---|---|
| The endpoint changed | The same fixed −U on the 175 TCGA patients, with a five-year all-cause endpoint in place of disease-specific survival | C 0.632 (0.540 to 0.724), against 0.623 (0.513 to 0.731) for disease-specific survival. The TCGA association does not depend on the endpoint. |
| Fewer tiles per patient | 32 of 50 TCGA tiles drawn at random, 1,000 times, 173 patients | Median C 0.606 for disease-specific survival; no draw at or below 0.50. |
| Image rendering | 336 SurGen fields re-rendered three ways (bicubic; via 256 pixels; via 256 pixels with the calibrated operator) | Patient rank agreement for U of 0.998 or higher. |
| Hospitals ranked against each other in TCGA | Pairs restricted to patients from the same TCGA site | C 0.624 (0.509 to 0.735), but one site (AA) supplies 88% of the pairs. |
Two descriptive differences between the cohorts were recorded. U was lower and less spread in SurGen (mean 0.126, SD 0.034) than in TCGA (0.195, 0.053). Patient rankings were also less repeatable across disjoint halves of each patient's fields: median rank correlation 0.72 in SurGen against 0.87 in TCGA.
Two AI readers, blinded to patient, measurement and outcome, labelled 168 SurGen fields (two per test patient) by dominant appearance; they agreed on 158 (94%). Among 15 patients with both an agreed gland-rich and an agreed fibrous-rich field, U was lower in the fibrous field: mean paired difference −0.051 (descriptive 95% interval −0.095 to −0.014). The V interval included zero. Restricting to fields with no artefact flagged by either reader left 9 patients and a U difference of −0.076 (−0.129 to −0.032).
A denser study followed: 30 further fields from each of 32 patients, 960 fields, labelled by three AI readers in pairs (agreement 77% to 81%). Its primary analysis needed ten patients with at least four agreed gland fields and four agreed fibrous-matrix fields. Nine qualified, so the primary inference was not run; in all nine, U was lower in matrix (mean difference −0.038). At two fields per class, 25 patients qualified: U was lower in matrix in 22, mean difference −0.046 (nominal 95% interval −0.067 to −0.024). For V the mean difference was −0.00005 (−0.028 to 0.030).
The labels are provisional AI morphology, not pathologist annotation. They support a tissue-composition contribution to U. They do not establish why the survival association failed: a measure that reads lower in fibrous tissue will move with how much fibrous tissue the sampling collects, and SurGen's whole-tissue sampling collected more fibrous than gland fields (63 against 47 in the first study), but no analysis here ties that to the concordance result.
H was specified as a one-page rider, hashed on 4 November 2026 before any SurGen H value was joined to an outcome. Because H had been found on tumour-outline tiles, the rider added a tumour gate, which is a fitted layer and is named as such: each field is split into four quadrants, each quadrant is read by a frozen DINOv2 ViT-S/14 encoder [3] with a logistic layer trained on nine expert tissue classes from one laboratory's unnormalised tiles [4], and the field is kept when the mean tumour-epithelium probability is at least 0.50. On TCGA, with no outcome involved, this gate kept 61% of tiles, left 160 of 167 patients with at least four tiles, and reproduced the archived patient H at Spearman 0.951.
Three checks had to pass before any outcome join:
The reading rule was also fixed: C of at least 0.55 with a lower bound above 0.50 would count as transport; C of 0.50 or below would end H as a survival marker; anything else was inconclusive and would trigger a separately reported run on the training partition. Stage (I to IV, fixed order) was frozen as a positive control to be scored beside H on exactly the same patients.
Whole-tissue sampling. On the 32 fields per slide already acquired, 76 of 168 patients had four gated fields, and the median gated share was 9.4%. Both machine checks failed. The visual check was still carried out and recorded: all twelve top-ranked fields showed tumour epithelium (reviewed by the author with an AI assistant, not by a pathologist). Outcome joins stayed closed.
Tumour-targeted sampling. The decision amendment of 6 November allowed one remedy: denser tumour sampling under the same gate, frozen before any join. A map of likely tumour regions was built from archived low-resolution thumbnails by a ridge logistic classifier over simple colour statistics, fitted in five patient folds on the other folds' gate decisions, so that no patient's own gate scores shaped its own map. Up to 16 fresh fields per patient were taken from mapped regions in a fixed farthest-first order, 2,426 fields in all. Four patients had no mapped region and counted as zero. Result: 113 of 168 patients, median gated share 37.5%. The share check passed; coverage did not.
One top-up. An amendment recorded during acquisition, before any gate result, allowed a single top-up of up to eight further fields for patients still short, with no further rounds. It added 264 fields and 14 patients. Final: 127 of 168, eight short of 135; median gated share 37.5%.
Disposition. H closed on SurGen as untested because sampling did not cover enough confirmed tumour. No survival estimate exists for H or for the stage control, and the secondary models were not released.
For the 18 patients with 24 targeted fields each and fewer than four gate passes, all 432 fields were laid out with scores hidden and described by an AI reader, then compared with the frozen gate scores. Six fields were also inspected at full size (Figure 5).
Both explanations apply, sometimes within one slide. The review cannot estimate the detector's sensitivity, because the fields were selected for low yield, the reader was an AI and not a pathologist, and glandular epithelium can be benign. A second, independent tissue classifier trained on the same laboratory's tiles also read tumour poorly on SurGen while reading fibrous tissue well; that record is reported separately [5].
A fair external test of either measure would use tumour regions outlined by a pathologist, with sampling inside them fixed before outcomes are seen, in a cohort with enough deaths for a useful interval.
Frozen protocols, amendments with their SHA-256 hashes, patient-level coverage tables, the 432-field review and the independent audit records are held by the authors (folder KERNEL_CANDIDATE_EXTERNAL_KF010_20261103 and its lab books; this note in SURGEN_SURVIVAL_NOTE_20261110) and are available on request. SurGen images and labels are public [1]. Only public data were used.