What does a spatial measurement detect inside a colorectal gland?

Paired-gland methods artifact · 10 November 2026

Across 82 matched normal–dysplastic gland pairs, the common component of the Parallax measurement (U) was higher in dysplastic tissue in 16 of 19 patients. Rearranging colours within bands at the same depth made the difference larger. A separate analysis showed that the change in colour from the gland outline inward accounted for much of the intact U difference. The opponent component (V) behaved differently.

Same slide, different glands

The GG-pT1 release contains 13 normal and 13 dysplastic gland crops from each of 20 patients. We matched glands within a slide on size, outline solidity, eccentricity, cell density, nuclear haematoxylin and cytoplasmic eosin. Matching without reuse yielded 90 pairs. Both partners in 82 pairs supported the full measurement and rearrangement comparison. We averaged contrasts within each patient before averaging across patients. Studer et al., original study

Normal gland

normal gland, IntactIntact
normal gland, RearrangedRearranged

Dysplastic gland

dysplastic gland, IntactIntact
dysplastic gland, RearrangedRearranged
Figure 1. First eligible pair in the archived pair manifest, shown intact and after the first deterministic shuffle. Whole RGB triples move only within the same 8-pixel band of distance to the gland outline or crop edge. Pixels outside the mask stay fixed. This pair illustrates the operation; the estimates below use all eligible pairs. Crops and masks from Studer et al., GG-pT1 release.

The rearrangement strengthened the contrast

The two measurements describe persistence of the standardized sum (U) and difference (V) of the original M-cone and L–M kernel fields. The separation was 16 native pixels. The crop files lacked verified physical calibration. For the control we shuffled first, applied the same kernels, and averaged three shuffles per gland.

Mean dysplastic minus normal contrast for U and V under intact, rearranged and colour-adjusted conditions. Values and intervals appear in the table.
Figure 2. Points are patient means; bars are separate 95% patient-bootstrap intervals. The third row is the residual from a model fitted to the intact images. The paired increase in U after shuffling was +0.155 [0.105, 0.208].
Patient-mean contrastU [95% interval]V [95% interval]
Intact+0.130 [0.070, 0.193]−0.051 [−0.099, 0.002]
Within-band shuffle+0.285 [0.220, 0.351]−0.142 [−0.199, −0.086]
Additional amount attributed to depth+0.108 [0.034, 0.176]−0.002 [−0.084, 0.083]
Remaining after colour adjustment−0.052 [−0.112, 0.052]−0.063 [−0.178, 0.079]

How much does the colour profile explain?

We divided each gland into four equal-area bands by distance from its outline or crop edge and measured mean RGB optical density in each band. We fitted slopes on the other patients, then subtracted the predicted contribution of colour differences in the held-out patient. Depth added a positive contribution to U beyond bulk colour in 18 of 19 patients. The model accounted for more than the mean intact U difference; the residual interval extended beyond the prespecified ±0.05 tolerance for complete accounting. V showed little average contribution from depth.

These analyses explain variation in the instrument’s response in selected glands. Whether U adds diagnostic information beyond the simpler colour profile remains an independent-cohort question.

Where the tissue interpretation ends

The depth bands include the gland contents. In a provisional AI review of one pair per patient, lumen could be outlined in both partners in only seven pairs. Zero pairs had two marked cellular-wall regions large enough for the frozen measurement. Epithelial localization therefore remains unresolved.

Reproduction record

The research archive includes the pinned inputs, original matching and measurement plans, control seeds, numerical outputs and provisional annotation decisions. An isolated macOS replay with the pinned Python environment reproduced all 14 reference CSVs and the summary figure byte for byte. This establishes computational repeatability on that environment. The research archive and Historical58 implementation are held privately.

Study design and data: Graph-based Classification of Intestinal Glands in Colorectal Cancer Tissue Images · Pinned public GG-pT1 dataset.