Parallax Pathology · Technical note
Russell Parrish · analysis carried out with Claude (Anthropic) · Version 2, 6 November 2026 (version 1, 5 November 2026, reported the first test and the first adjustment)
Within-dataset cross-validation with folds split by tile lets tiles from the same patient and slide fall on both sides of a split, so it measures how separable the classes are in one collection, not accuracy on new patients. Kather and colleagues released NCT-CRC-HE-100K as a training set and CRC-VAL-HE-7K as a separate test set from different patients [1, 2]. Fitting on the first and testing on the second is the comparison those datasets were built for, and the comparison published models report.
Data. NCT-CRC-HE-100K: 100,000 tiles of 224 × 224 pixels at 0.5 µm/px from 86 slides, nine classes. CRC-VAL-HE-7K: 7,180 tiles from patients with no overlap with the training set (50 patients per the dataset record [2]; the paper describes 25 patients and slides [1]). Both are distributed already Macenko-normalised by their authors. Both archives matched the published checksums.
Development tiles and sealed tiles. 1,800 CRC-VAL tiles (200 per class) had been used in our earlier work, including the error analysis that produced the extension measurements in round 2. After the first external test, the other 5,380 tiles were sealed: no labels, predictions, features or images were opened during development. Every choice was made on the 100,000 training tiles and the 1,800 development tiles. At the end of each round the version was frozen and hashed, then run once on the sealed tiles, and the result was entered in a look log. Looks 0 to 5 are all the evaluations made on those tiles.
What is fixed and what is fitted. Every measurement is a fixed, label-free calculation on one tile: a stain amount, a texture energy, a nucleus width, a gap fraction. The same calculation is applied to every collection, and nothing in it is learned. The readout that maps the measurements to nine classes is fitted on the training collection: a linear discriminant in looks 0 and 1, then gradient boosting (scikit-learn HistGradientBoostingClassifier), from look 3 averaged with logistic regression. Median imputation and standardisation are fitted on the training tiles.
Uncertainty. Wilson 95% intervals at tile level. Patient identifiers are not available for the tiles, so the intervals understate the true uncertainty. Changes between looks are compared tile by tile (paired sign test on the tiles that changed outcome).
The 81 fixed fields behind the 91.83% figure, with a linear discriminant fitted once on all 100,000 training tiles, scored 83.7% on all 7,180 test tiles (balanced 77.0%) and 85.7% on the 5,380 later-sealed tiles (balanced 76.7%). On the 1,800 development tiles it scored 77.8%, against 91.8% from within-dataset cross-validation on those same tiles. The exact 58-column instrument as published scored the same (83.8%). Fat, background, mucus and tumour transferred well (recall 0.90–0.97); stroma (0.38) and debris (0.45) did not.
| Look | Change (frozen before the look) | Accuracy, 5,380 sealed tiles | Balanced | Against previous look: newly right / newly wrong |
|---|---|---|---|---|
| 0 | 81 fixed fields, linear discriminant | 85.7% (84.7–86.6) | 76.7% | · |
| 1 | + violet protection | 87.0% (86.0–87.8) | 82.0% | 95 / 28 |
| 2 | + 73 extension fields; gradient boosting readout | 91.5% (90.7–92.2) | 89.4% | 442 / 198 |
| 3 | Readout only: orientation-free fields, boosting averaged with logistic regression | 91.6% (90.9–92.3) | 89.9% | 113 / 105 (p = 0.64) |
| 4 | + 70 new fields (colour, texture, nuclear form); six field views averaged | 93.8% (93.1–94.4) | 91.0% | 234 / 120 |
| 5 | + orange protection | 94.3% (93.6–94.9) | 92.1% | 29 / 0 |
All p-values for the paired comparisons are below 10−8 except look 3. Every figure in this table was recomputed from the saved per-tile predictions of each look.
No global stain difference between the collections was found. The debris failure had a specific cause: half of the development debris tiles are rendered in a single violet hue, against under 3% of training debris, and the instrument read them as tumour. Every tile of every collection is checked with one colour rule (mean blue minus mean red above 15). Flagged tiles keep their structure but are re-coloured, each pixel's brightness rank mapped to the average training tissue colour at that brightness, and are then re-measured. Debris recall rose from 0.47 to 0.96; debris called tumour fell from 58 to 1. Corrected Macenko re-normalisation, per-collection feature rescaling and per-channel colour mapping were tried on development tiles and rejected.
Two families of fixed measurements (33 and 40 fields), built in August 2026 from an error analysis of the 1,800 development tiles, were added unchanged: multiscale fascicle geometry, coupling between nuclei and fibres, gland-ring polarity, and mixture architecture. They are weak alone (66.7–75.4% on development tiles) and strong in combination with the original 81. With a gradient-boosting readout, stroma recall rose from 0.35 to 0.71 and every class reached 0.71 or above. With the linear discriminant, the same measurements scored 88.0%, so the readout contributed about 3.5 points of the 4.5-point gain.
Round 3 tuned the readout and dealt with eight fields that change when a tile is rotated (two pairs of direction offsets were replaced by their lengths; the other four were dropped). Averaging gradient boosting with logistic regression gained 1.9 points on development tiles, selected from about 18 combinations. On sealed tiles it gained nothing: 91.6% against 91.5%, 113 tiles newly right and 105 newly wrong. Errors moved between classes (normal called tumour fell from 61 to 48; stroma and muscle confusions rose from 94 to 118). On those 148 measurements the readout had reached its operating limit, and further gains needed new measurements.
A family of 70 new fixed fields measured the non-nuclear tissue directly: absolute stain amounts by fixed colour deconvolution (Ruifrok and Johnston [3]) with no per-tile rescaling, CIELab colour, rotation-invariant texture (local binary patterns, Gabor energy at 4, 8 and 16 px, eosin co-occurrence), nuclear width, length, bluntness and waviness, clear zones at the poles of elongated nuclei, and gaps. Sixty of the 70 fields separate stroma from muscle, or normal mucosa from tumour, in the same direction in both collections. The strongest is the haematoxylin-to-eosin ratio of non-nuclear tissue: smooth-muscle cytoplasm takes eosin strongly, while collagen and ground substance hold relatively more haematoxylin. It separates stroma from muscle in both collections (area under the curve 0.71 in training tiles, 0.86 in development tiles).
| Median H/E ratio, non-nuclear tissue | Muscle | Stroma | Debris | Tumour | Lymphocytes | Normal |
|---|---|---|---|---|---|---|
| Training (NCT) | 0.39 | 0.51 | 0.45 | 0.63 | 0.68 | 0.80 |
| Development (CRC-VAL) | 0.34 | 0.53 | 0.54 | 0.61 | 0.57 | 0.59 |
Adipose is omitted: with little tissue per tile, the ratio is unstable (5.4 and 1.08). The ratio also depends on a working stain separation, and reads orange-rendered tissue as nearly unstained (4.5).
Single readouts with the new fields moved errors rather than reducing them. The gain came from averaging the class probabilities of twelve readouts (gradient boosting and logistic regression on each of six field sets). Three of the six field sets were screened with development labels; a split-half check, screening on one half of the development tiles and scoring on the other, gave 92.1%. On sealed tiles the average scored 93.8%, against 91.6% for a single view on the same fields. Muscle called stroma fell from 70 to 3, normal recall rose from 0.86 to 0.91, and stroma recall fell from 0.72 to 0.64. A new error appeared: 38 sealed muscle tiles called mucus.
On development tiles the same error came from muscle rendered orange-tan, a second rendering of the test collection. A label-free gate catches it: the median CIELab b* of tissue pixels is above zero (yellow-shifted), which ordinary pink and purple H&E almost never is; by the corresponding non-nuclear measurement, only 2 of 100,000 training tiles (both background) pass. Re-colouring alone failed: re-coloured orange muscle read as stroma, because the orange rendering has erased the stain balance that separates the two, and no re-colouring can restore it without knowing the class. 
Gated tiles are therefore re-coloured, re-measured, and read by a separate readout restricted to the 84 fields that change by at most 0.25 standard deviations when training tiles are forcibly re-coloured: mostly texture, orientation and shape. Every other tile is read exactly as in round 4.
On sealed tiles the gate routed 47 tiles (38 muscle, 5 debris, 4 stroma). None had been correct at look 4; 29 became correct and none became wrong, and the other 5,333 tiles kept their look-4 predictions exactly. Muscle called mucus fell from 38 to 0; 19 muscle tiles are now called stroma, against 3 at look 4.
At look 5, 307 of 5,380 sealed tiles are misclassified, against 769 at look 0. The four largest confusions account for 161 of them.
| Dataset label → prediction | Look 0 | Look 5 |
|---|---|---|
| Adipose → muscle (of 1,138) | 7 | 52 |
| Stroma → muscle (of 221) | 54 | 51 |
| Normal → tumour (of 541) | 89 | 39 |
| Muscle → stroma (of 392) | 63 | 19 |
| Debris → tumour (of 139) | 58 | 0 |
| Lymphocytes → debris (of 434) | 43 | 13 |
Stroma called muscle has barely moved across the five rounds (54 at look 0, 51 at look 5; lowest 40, at look 2). Development stroma called muscle is dense, wavy and strongly aligned; stroma called correctly is loose and pale; the training collection's stroma is mostly loose, vacuolated or myxoid. Dense fibrous stroma therefore falls on the muscle side of what any readout fitted to the training collection learns. Whether those tiles are a stroma type absent from training, mixed or transitional crops, or labelling differences is a question for a pathologist's blinded read. Adipose called muscle rose from 7 at look 0 to 37 at look 2 and 52 at look 4, and was not investigated.
On tiles it was never tuned on, the instrument went from 85.7% to 94.3%, and from 76.7% to 92.1% balanced. Each gain came from something that can be named: a rendering the training collection lacks, flagged by a colour rule; a stain balance that separates two tissues; texture and nuclear form that survive re-colouring. Each step was followed by a frozen evaluation, and the steps can be read tile by tile. The round that changed only the readout produced no gain. The rounds that added measurements or protections did, though the look-4 gain needed both the new fields and averaging over field views: a single view on the same fields scored 91.6%.
The dataset authors' convolutional network reported 94.3% on all 7,180 test tiles [1]. This instrument reaches 94.3% on the 5,380 sealed tiles. The two figures are not on the same tile set: the 1,800 development tiles are excluded here, and the class mix differs. The comparison is approximate. It does place a fixed, named measurement set with a fitted readout at the level of the learned model on this benchmark.
The two rendering protections are part of the instrument for any collection: both are label-free colour rules, applied to every tile before measurement, and both flag almost nothing in the training collection. They say when a tile's colour should not be trusted, which is useful independently of the classifier.
The lab book for each round is published as written: round 1 (violet protection), round 2 (extension measurements), round 3 (readout only), round 4 (colour, texture and nuclear form), round 5 (orange protection). Frozen specifications, script hashes, per-tile measurements and predictions for every look, and the look log are held by the authors and are available on request. Only public data were used.