SOURCES — public gastric H&E sources assessed against 4096x2048 @ 0.25 um/px

Work package: REAL-SOURCES. Date: 2026-09-09. Companion to INVENTORY.md.

0. The arithmetic that decides everything

A 4096 x 2048 px image at 0.25 um/px is a 1024 um x 512 um piece of tissue sampled at 8,388,608 pixels. A source can supply one only if it satisfies both:

• Sampling test: native pixel size <= ~0.25 um/px. Coarser means upscaling.
• Field test: the frame covers at least 1024 x 512 um. A smaller frame means the pixels exist but the tissue does not, and stitching is not an option for a single field.

Almost every published gastric dataset fails the field test, because almost every one of them is distributed as small patches (80-2776 px). Almost every published H. pylori H&E image fails the sampling test in the opposite direction — camera micrographs at 400x-1000x oversample the optics so heavily that a 12 MP frame covers only 100-700 um of tissue.

Only whole-slide images pass both tests. That is the single most important finding in this package.

1. The survey

Verified 2026-09-09 by direct API calls (figshare, Zenodo, GDC, Commons) and page fetches. Licence strings are quoted as published. "Organisms?" means: does the source assert or annotate the presence of H. pylori organisms (not merely a gastritis diagnosis)?

Searches that returned nothing usable

• Zenodo, "Helicobacter pylori" AND (histopatholog* OR histolog* OR "whole slide" OR biopsy), type=dataset: 8 records total. Seven are clinical/epidemiological tables with no images; the eighth is DeepHP (embargoed). There is no Zenodo record of H. pylori H&E whole slides.
• Kaggle: no gastritis or H. pylori H&E histology set. What exists is gastric cancer mirrors of GCHTID/TCGA, plus endoscopy sets (Kvasir, GastroVision) which are not histology.
• "HP-Neu": no dataset of this name exists. The NEU in the GitHub handle above is Near East University, Nicosia. The separate Girona/Barcelona HP work (arXiv:2412.13857) is immunohistochemistry, not H&E, and has no public release.
• Gastric CAMELYON/PANDA equivalent: none exists. PAIP has never run a gastric task.
• Paper supplementary figures: not viable. Published H&E figures showing organisms are typically 1000-2000 px on the long edge and JPEG-compressed. Upscaling 2-4x would destroy exactly the 2.5-5 um curved-rod morphology that makes the category what it is.

2. The H. pylori organism problem — the honest assessment

No public dataset provides H. pylori organism-level annotation on H&E at a resolution and field of view that can produce a 4096 x 2048 image at 0.25 um/px. That is the finding. It is not a gap in the search; it is the state of the public record on 9 September 2026.

The three ways this fails, one per candidate class:

(a) Sources that assert organisms but are too small and/or licence-blocked. The NEU dataset is the best of these: a pathologist graded HP density case by case, and the grades survive in the filenames (B129-19 HP (++), B198-19 HP (+++)). It is 1280 x 960 — a 3.2x linear upscale short — and it is CC BY-NC-ND 4.0, where the ND term forbids distributing the adapted material that any crop or resample would produce. Either failure alone is disqualifying. DeepHP would be the strongest source of all (394,926 labelled images, 111k HP-positive) and is embargoed until 31 December 2030.

(b) Sources that annotate organisms precisely, but not on H&E. The IEEE DataPort HP set has actual point coordinates for organisms — on immunohistochemistry. IHC is a different stain with a different appearance; it cannot be submitted as gastric H&E.

(c) The one source with the right geometry does not assert organisms. HiESD is 40x, 0.2458 um/px, CC BY 4.0, whole slides. Its ten annotation classes include Chronic Gastritis and Chronic atrophic gastritis but no H. pylori class. The dataset annotates lymphoid follicles, which its authors flag as a morphological indicator of HP infection — an indicator is not the organism.

We also tested this empirically rather than assuming it. Package B1-ORGANISMS (operations/research/sgh-program-20260908/B1_RESULT.md) scanned 753 mucosal-surface windows across 13 HiESD slides with a classical curved-rod detector and graded the top 60 candidates by eye:

• 0 of 60 graded "likely organisms"; 5 ambiguous, 55 not organisms.
• The top calls were collagen and muscle fibres, capillary walls, basement membrane, elongated endothelial/fibroblast/lymphocyte nuclei, surface fibrin and stain precipitate.
• The negative arm scored like the positive arm — normal-gland surface windows reached p90 20.8 against 24.0 for gastritis surfaces, and the single highest-scoring window in the first sweep was a negative.
• Tightening the search region from "dim pale" to genuinely bright luminal space — the compartment where the organism actually lives — collapsed the signal to zero on both arms.

That is an engineering result about a classical detector at 0.2458 um/px, not a pathologist's verification and not a claim that the HiESD slides are HP-negative. But it means we have no positive evidence of organisms on any HiESD field, and we must not label one as HP-positive.

Consequence for the submission. We can source, at native resolution and under a clean CC BY 4.0 licence, real gastric H&E fields for:

• intestinal metaplasia,
• chronic / chronic atrophic gastritis,
• co-occurring intestinal metaplasia and gastritis in one field,
• normal gastric mucosa with neither.

We cannot source a real field that is 4096 x 2048 at 0.25 um/px and in which a H. pylori organism is verified present. Any real image we label "H. pylori organism with gastritis" would be labelled on a diagnosis of gastritis, not on an observed organism.

3. The Wikimedia Commons route, measured rather than assumed

Commons holds the best publicly licensed gastric H&E images that visibly contain organisms — Nephron (Michael Bonert), CoRus13, Ed Uthman and others, at up to 6000 x 4000 px. Not one of them states a pixel size. So we measured it.

Method. Nuclear texture has a near-constant physical size, so the scale-normalised Laplacian-of-Gaussian response peaks at a sigma proportional to 1/(um per px). We calibrated the constant on HiESD, whose openslide.mpp-x is known to be 0.2458: K = sigma_peak * mpp = 1.61, stable to +/-1.5% across the native image and its 2x and 3x downsamples. Applying mpp = K / sigma_peak to each Commons file gives the table in candidates/wikimedia-commons/provenance.csv. The estimator is honest about its limits: where the peak pinned at the edge of the scale-space range, the row says "not measurable" rather than guessing.

Visual inspection confirms the numbers. The "very high" and "extremely high" HP frames are oil-immersion shots so heavily magnified that a 700 px crop contains a handful of cells and visible optical blur — empty magnification, not extra information. This is the trap in the whole Google-Images approach: the images that show the organisms best are precisely the ones whose field of view is smallest.

The structural reason. These are 12 MP DSLR frames (Canon EOS 450D / Rebel XSi / 1100D, 4272 x 2848, 5.196 um sensor pitch). A 12 MP frame at exactly 0.25 um/px would cover 1068 x 712 um — just enough for one 4096 x 2048 field, and only if the shot happened to be taken at ~0.25 um/px. Nephron's magnification tiers straddle that value rather than sitting on it, so the coincidence is rare. Exactly one of fourteen files landed in the usable band, and it is a normal mucosa image — the one category HiESD already supplies in quantity.

And the licence carries an obligation. Every Commons candidate is CC BY-SA (2.0 / 3.0 / 4.0). ShareAlike is not NoDerivatives — cropping is permitted — but the crop is adapted material that must itself be released under the same or a compatible licence, with attribution to the named author. If the challenge asserts any rights over submitted images, a CC BY-SA image creates a conflict that CC BY 4.0 material does not. Recommendation: do not put CC BY-SA material in the 70.

4. HiESD in detail — capacity, and one licence caveat

Licence caveat, to be resolved before submission. The figshare data record returns {"name": "CC BY", "url": "https://creativecommons.org/licenses/by/4.0/"} via the API. The Scientific Data article page renders a CC BY-NC-ND notice. These are different objects — the article text and the deposited data — and the data record governs the .svs files. Because a reviewer will ask, screenshot the figshare licence field on the day of download and keep it with the submission.

Capacity survey across the whole 104-slide collection

The ESD_40X_annotation_downsample64 masks are present locally for all 104 slides, including slides whose .svs we have never downloaded. That let us survey the entire collection's capacity without any further downloads. For every slide we slid a 4165 x 2083 level-0 window (which downsamples to exactly 4096 x 2048 at 0.25 um/px) over the mask and counted disjoint, non-overlapping windows meeting each category rule, with any carcinoma label excluding the window outright.

Script: hiesd_field_yield.py. Full per-slide table: hiesd_field_yield.csv.

Supply is not the constraint. Even before this package's downloads, the material on hand exceeded every requirement by more than an order of magnitude. What we bought with the six extra slides is specimen diversity — the IM and IM+gastritis pools roughly doubled, so a 14-image category can now be drawn from 18-20 distinct slides instead of 12-14.

A limitation to state plainly: HiESD is 104 slides from 44 patients, but the release publishes no slide-to-patient map. We can guarantee 70 images from distinct slides; we cannot guarantee 70 images from 70 distinct patients, and must not claim it.

The resampling is a downsample, and that matters

HiESD level 0 is 0.2458 um/px, and the target is 0.25 um/px. A 4096 x 2048 output at 0.25 um/px covers 1024 x 512 um, which is 4165 x 2083 level-0 pixels. So each field is read at 4165 x 2083 and downsampled by a factor of 1.017 to 4096 x 2048. This is a reduction in pixel count, never an increase. Nothing is interpolated into existence. Verification fields are in candidates/hiesd/ with the read coordinates, the source mpp and the resample factor recorded per file.

5. Sources still worth pursuing, and what blocks each

6. Recommended composition of the 70

Planned split 21 / 14 / 14 / 21. Every block below comes from HiESD — figshare DOI 10.6084/m9.figshare.28919840, CC BY 4.0, 40x scan, openslide.mpp-x 0.2458, cut at 4165 x 2083 level-0 pixels and downsampled 1.017x to 4096 x 2048 at exactly 0.25 um/px.

Selection discipline for whoever cuts the final 70: draw round-robin across slides so no single slide contributes more than ~4 images to a category; enforce a 300 um minimum centre-to-centre spacing so fields never overlap; exclude any window within 600 um of a carcinoma label; and apply the GrandQC quality mask (ESD_40X_Quality_Assessment_Mask, 0.9995 um/px) to keep out folds, pen marks, bubbles and out-of-focus tissue.

What NOT to include, and why

• Nothing from GasHisSDB. 160 px patches; 25.6x upscale.
• Nothing from the NEU HP dataset. CC BY-NC-ND forbids derivatives, and 1280 x 960 is 3.2x short.
• Nothing from Wikimedia Commons. ShareAlike obligation, no stated pixel size, and thirteen of fourteen candidates fail the geometry outright.
• Nothing from GAGL. Downsampled, and no file even reaches 4096 x 2048 px.
• No Google Images material. No provenance, no licence, no calibrated scale.
• No upscaled image of any kind, from any source.

Source statement to submit

The 70 real images are 4096 x 2048 px fields cut from HiESD — A Fully Annotated Pathology Slide Dataset for Early Gastric Cancer and Precancerous Lesions (Ge et al., Scientific Data, 2025; DOI 10.1038/s41597-025-05679-1), deposited on figshare under CC BY 4.0 at DOI 10.6084/m9.figshare.28919840. The slides are H&E-stained gastric endoscopic submucosal dissection specimens scanned at 40x objective. Each source slide reports openslide.mpp-x = 0.2458 um/px. Each field was read at 4165 x 2083 level-0 pixels — the exact extent of 1024 x 512 um — and resampled down by a factor of 1.017 to 4096 x 2048 px at 0.25 um/px. No image was upscaled. Category assignment follows the pathologist region annotations distributed with the dataset (Annotation_Category_Label_Description.txt): complete and incomplete intestinal metaplasia, chronic gastritis, chronic atrophic gastritis and normal glands. Fields containing any carcinoma annotation were excluded. HiESD does not record Helicobacter pylori organism status; images assigned to the H. pylori categories are so assigned on the basis of the annotated chronic gastritis pattern, not on observed organisms.

That last sentence is the one that must not be dropped.

Download public Markdown export