External CRC benchmark, round 4: lab book

Parallax Pathology · Russell Parrish, analysis carried out with Claude (Anthropic) · 5–6 November 2026 · lab book for round 4 of the external CRC benchmark · published November 2026

Published as written at the time, apart from this header. File and folder names refer to the project record, available on request. Benchmark note · Claim ledger

Summary. Round 4 added 70 new fixed measurements (the r4 family: colour of the non-nuclear tissue, texture, nuclear form, perinuclear halo, gaps) and read them with an average of twelve readouts over six field views. Frozen and run once on the 5,380 sealed tiles (look 4), it reached 93.8% accuracy (balanced 91.0%), up from 91.6% (89.9%) at look 3; 234 tiles newly right against 120 newly wrong. The strongest new measure is the ratio of haematoxylin to eosin in the non-nuclear tissue, which separates stroma from muscle in both collections. Muscle called stroma fell from 70 to 3, but stroma itself fell from 0.72 to 0.64, and a new error appeared: 38 sealed muscle tiles called mucus. On development tiles the same error comes from muscle rendered orange-tan, a second rendering variant of the test collection.

Entry 1 · Bookkeeping

Before round 4 the sealed tiles had four logged looks (0 to 3). Same rule: develop on NCT (all) and the 1,800 development tiles only; one logged look at a frozen version (../EXTERNAL_CRC_ROUND2_20261105/LOOK_LOG.md).

Entry 2 · The r4 family

scripts/r4_axes.py: 70 fields, fixed and label-free, each a named quantity of the tile. The existing v1/v2 families separate the stains but rescale each one within the tile, which removes absolute stain amounts; r4 keeps them.

Colour of the non-nuclear tissue: fixed Ruifrok–Johnston haematoxylin/eosin/DAB deconvolution (scikit-image rgb2hed, no rescaling) and CIELab, measured on tissue with nuclei (Otsu on haematoxylin, dilated 2 px) removed: percentiles, haematoxylin-to-eosin ratio, hue, chroma, eosin smoothness at two scales.
Texture: rotation-invariant local binary patterns; Gabor energy at wavelengths 4, 8 and 16 px (mean, and maximum over mean across 8 orientations); eosin co-occurrence averaged over 4 angles at distances 1, 3 and 7.
Nuclear form: width, length, end bluntness (width near the ends over maximum width), waviness (skeleton length over axis length), chromatin darkness and texture.
Perinuclear halo: optical density 3–6 px beyond the poles of elongated nuclei minus beside their flanks (the clear vacuoles of smooth muscle).
Gaps: clefts inside tissue: fraction, count, slit-ness, share of thin gaps.

The fields are built to be rotation-invariant. On 6 test tiles rotated by 90°, one field moved by more than 5% (nuclear waviness, through the pixel grid). Speed about 0.08 s per tile. Measured with round-1 violet protection on all 100,000 NCT tiles (from the external drive, checkpointed), the 1,800 development tiles and, after the freeze, the 5,380 sealed tiles: no failures; the gate flagged 363, 103 and 75 tiles, as in every round.

Entry 3 · Do the new fields transfer?

Area under the curve of each field, NCT and development tiles (outputs/02_transfer.csv); 60 of 70 fields point the same way in both collections with |AUC − 0.5| > 0.05 for at least one of the two pairs.

Stroma vs muscle	NCT	Development	Normal vs tumour	NCT	Development
H/E ratio, non-nuclear tissue	0.714	0.863	Gabor energy, 16 px	0.801	0.923
Haematoxylin, 10th percentile	0.602	0.761	Lightness (L*), median	0.828	0.908
Chromatin texture	0.406	0.246	Haematoxylin, 90th percentile	0.242	0.103
Red-green (a*), median	0.285	0.254	DAB channel, median	0.179	0.106
DAB channel, median	0.416	0.263	Chroma, median	0.154	0.119

AUC above 0.5 means higher in stroma (left) or normal (right). The best earlier stroma/muscle field reached about 0.79 on development tiles.

Entry 4 · The haematoxylin-to-eosin ratio

Median haematoxylin over median eosin in the non-nuclear tissue. Smooth-muscle cytoplasm takes eosin strongly; collagen and ground substance hold relatively more haematoxylin.

Median ratio	Muscle	Stroma	Debris	Tumour	Lymphocytes	Normal	Adipose
NCT	0.39	0.51	0.45	0.63	0.68	0.80	5.4
Development	0.34	0.53	0.54	0.61	0.57	0.59	1.08

Limits: it helps the stroma/muscle boundary, not the whole picture (debris, tumour and lymphocytes sit nearby); it is unstable on tiles with little tissue (adipose); it depends on a working stain separation; and it reads orange-rendered tissue as nearly unstained (Entry 9).

Entry 5 · Development readouts

Development tiles	Gradient boosting (500)	Logistic	Average of both	Stroma (avg)
Round-3 fields (148)	89.1%	86.4%	91.0%	0.74
+ r4, all (218)	90.6%	86.0%	91.5%	0.68
+ r4, transfer-screened (208)	90.0%	85.3%	90.3%	0.66
Shift-screened, max class shift ≤ 0.5 SD (59)	89.2%	82.0%	88.9%	0.61
Shift-screened, ≤ 0.75 SD (110)	90.7%	85.9%	91.4%	0.61
Shift-screened, ≤ 1.0 SD (152)	90.6%	83.9%	90.8%	0.64

Shift screen: for each field and class, (development median − NCT median) / NCT standard deviation; fields kept if the largest absolute shift over the nine classes is within the bound. Every version gains muscle and loses stroma. Development stroma is denser and more aligned than NCT stroma (Gabor anisotropy is about 0.5 SD higher in both development stroma and development muscle, which moves stroma towards muscle), and development adipose has far less tissue per tile (tissue fraction shifted −4.3 SD), which destabilises its colour fields.

Entry 6 · Averaging over views

The six views make different mistakes. Averaging the class probabilities of all twelve readouts (six views × two readouts) gives 92.4% on development tiles. About 240 subsets of eight readouts from four of the views were also scored (best 92.6%); the rule "every view, both readouts" was taken instead of the best subset. Three views are chosen with development labels (transfer and shift screens), so a split-half check was run: screens chosen on one stratified half, ensemble scored on the other, then swapped (scripts/04_split_half.py): 92.1%.

Entry 7 · Freeze (02:35 UTC, 6 November)

FROZEN_ROUND4.md and hashes. Primary: the twelve-readout average, all fitted once on the 100,000 NCT tiles; screens recomputed at look time from NCT and the 1,800 development tiles, identical to development. Secondary, same look: round-3 fields + r4 all, gradient boosting and logistic only (no screening on development labels).

Entry 8 · Look 4 (sealed, run once)

5,380 sealed tiles	Accuracy (95% tile-level)	Balanced	Stroma	Muscle	Normal
Look 2: + v1/v2 extensions, gradient boosting	91.5% (90.7–92.2)	89.4%	0.706	0.852	0.819
Look 3: direction-free, boosting + logistic	91.6% (90.9–92.3)	89.9%	0.719	0.814	0.860
Look 4: + r4, six views averaged	93.8% (93.1–94.4)	91.0%	0.643	0.890	0.915
Look 4: + r4, single view (secondary)	91.6% (90.9–92.3)	89.5%	0.611	0.885	0.884

Primary recall by class: adipose 0.931, background 1.000, debris 0.914, lymphocytes 0.963, mucus 0.980, muscle 0.890, normal 0.915, stroma 0.643, tumour 0.956. 5,044 of 5,380 correct (336 errors). Tile by tile against look 3: 234 newly right, 120 newly wrong (paired sign test p ≈ 1 × 10⁻⁹).

Sealed confusion	Look 3	Look 4
Muscle → stroma (of 392)	70	3
Stroma → muscle (of 221)	48	50
Stroma → tumour / debris	5 / 0	15 / 8
Normal → tumour (of 541)	48	39
Muscle → mucus (of 392)	0	38
Adipose → muscle (of 1,138)	40	52

Entry 9 · Orange-rendered muscle (development tiles)

Development muscle called mucus, muscle called correctly, and mucus — Rows: development muscle called mucus by the round-4 average (the first 8 of 17); development muscle called correctly; development mucus.

On development tiles the twelve-readout average calls 17 muscle tiles mucus (round-3 fields alone: 1). They are rendered orange-tan. The fixed stain vectors read orange as almost unstained: H/E ratio 0.06 against 0.37 for correctly called muscle, eosin median 0.011 against 0.034 (mucus 0.010). Label-free, mean red minus mean blue above 30 flags 28 development muscle tiles, 4 debris and 2 stroma, and no other class. The sealed muscle-to-mucus error was not opened; on development tiles it has this cause.

Entry 10 · Reading

Look 4 is a measured gain on tiles not used to choose anything: 93.8% against 91.6%, balanced 91.0% against 89.9%. The development estimate (92.1% split-half) was conservative.
The single-view secondary equals look 3 (91.6%): on sealed tiles the r4 fields moved errors rather than reducing them, and the gain appears when views built on different field sets are averaged. Every view but one contains r4 fields, so the two cannot be fully separated.
The gain is in muscle (0.814 to 0.890), normal (0.860 to 0.915), mucus (0.927 to 0.980) and tumour (0.936 to 0.956); stroma (0.719 to 0.643) and debris (0.942 to 0.914) lost. The sealed set has more muscle (392) than stroma (221), which favours this trade in overall accuracy; balanced accuracy also rose, so the mix does not explain all of it.
The dataset authors' convolutional network reached above 94% on all 7,180 test tiles; on the 5,380-tile sealed subset the gap is now under one point (an approximate comparison).
Open, for a next round: an orange-rendering gate with recolouring, by analogy with the violet gate (development evidence only so far); stroma, now 0.64, where the test set's dense aligned stroma lies outside the training collection's stroma.
The sealed tiles have now had five logged looks (0 to 4). Further rounds on this test set are development; independent confirmation needs a new collection.

Entry 11 · Note for other workstreams (H/E ratio)

Ratio of haematoxylin to eosin in non-nuclear tissue: fixed Ruifrok–Johnston deconvolution (scikit-image rgb2hed, no per-tile rescaling); nuclei masked by Otsu on haematoxylin within tissue, dilated 2 px; median H ÷ median E over the remaining tissue. Stroma vs muscle AUC 0.714 on NCT-CRC-HE-100K and 0.863 on CRC-VAL development tiles, same direction. In look 4 it is one of 70 new fields inside a six-view average that reached 93.8% on sealed tiles; its separate contribution on sealed tiles has not been isolated. Conditions: a minimum amount of non-nuclear tissue; a working stain separation (the earlier degenerate Macenko normaliser, with stain vectors 0.17° apart, would flatten it); and a rendering check, since orange-rendered tissue reads as nearly unstained. Code: EXTERNAL_CRC_ROUND4_20261105/scripts/r4_axes.py.