Proposal Specification · Section 1

Datasets & Evidence Sources

Keep this list honest: HEST-1k does the load-bearing work (segmentation + transcriptomics); HumanST-1k widens molecular coverage; the rest provide supplementary H&E volume and reference annotations. If any end up unused, cut them rather than pad.

PRIMARY SPINE DATASET

HEST-1k (Histology & Spatial Transcriptomics)

Licence: CC BY-NC-SA 4.0 (Non-commercial, fine for FYP)

The spine of the project. 1,229 paired samples across ~26 organs: H&E whole-slide images with spatial transcriptomics and cell/nuclei segmentation. Provides cellvit_seg (CellViT nuclei masks, GeoJSON, five nuclear classes, ~76M nuclei across all slides) and, for Xenium samples, xenium_seg (nucleus and cell boundaries derived from DAPI and finely aligned to the H&E). Used for SAM2 fine-tuning, the transcriptomics-guided region grouping, and the per-organ failure study. Masks are model-derived (CellViT) except the Xenium DAPI masks — disclosed, not hidden.

PAIRED SAMPLES

1,229 WSIs

ORGAN COVERAGE

~26 Organs

SEGMENTED NUCLEI

~76 Million

PRIMARY ROLE

SAM2 Tuning + ST Grouping

Action Item: Confirm HEST's per-organ Xenium / cell-mask coverage before locking the four demo organs. Cell-resolution segmentation is only honest on the Xenium / Visium HD subset; on Visium-only organs it is spot-level (near-cell).

SECONDARY TRANSCRIPTOMICS SOURCE

HumanST-1k (Broader Molecular Organ Coverage)

Citation: STAMP (arXiv:2606.03644)

Secondary transcriptomics source, for broader organ coverage on the molecular side. ~1,004 human-only spatial-transcriptomics samples across ~30 organ/tissue types and 4 sequencing platforms, each a high-resolution H&E WSI co-registered with a spatial expression profile (~2.1M registered H&E patches paired with expression). Introduced by the STAMP paper (arXiv:2606.03644). Curated from public repositories — 10x Genomics, CNCB, Dryad, EMBL, HEST-1k itself, the Human Cell Atlas, HTAN, NCBI, Zenodo, and others.

HUMAN-ONLY SAMPLES

~1,004 WSIs

ORGAN/TISSUE TYPES

~30 Organs

SEQUENCING PLATFORMS

4 Platforms

REGISTERED PATCHES

~2.1 Million

Role and caveats: HumanST-1k widens organ coverage for the per-organ study (RQ2) and the transcriptomics grouping — it is not a SAM2 training source. It is built for molecular-alignment training (expression-paired patches) and does not ship ready-to-use nuclei/cell masks the way HEST does, so it does not replace HEST for segmentation. It also partly overlaps HEST (drawn from it), so it is a supplementary pool, not an independent dataset. Confirm its licence and per-organ cell-resolution coverage before relying on it.

SUPPLEMENTARY (ADDITIONAL H&E VOLUME AND REFERENCE)

Supplementary Datasets Catalog

Additional H&E volume and reference annotations for organ architectures and grading benchmarks.

DATASET	WHAT IT PROVIDES	ROLE IN GLASSBOX	LICENCE
TCGA (BRCA, LUAD/LUSC, COAD/READ, PRAD)	Public cancer H&E + clinical metadata	Extra H&E volume for diagnosis/report context; organ architecture reference	NIH open access
PANDA	Prostate biopsies, ISUP grade labels	Prostate grading reference / classification signal	Research use
CAMELYON16/17	Lymph-node metastasis WSIs	Metastasis morphology reference	CC0 / research
NCT-CRC-HE-100K	100k colorectal H&E tiles, 9 classes	Colorectal tissue-class reference, extra volume	CC BY 4.0
PatchCamelyon	327k lymph-node tiles, tumour/normal	Metastatic morphology reference	CC0 1.0
BreakHis	Breast tumour histology, 40x–400x	Breast grading appearance reference	Research use, on request
OpenSlide test data	Vendor-format WSIs	Viewer / deep-zoom tiling reference	Open, per file

HONEST DATA PRINCIPLE

Keep this list honest: HEST does the load-bearing work (segmentation + transcriptomics); HumanST-1k provides supplementary molecular breadth; the rest are supplementary H&E or references. If any end up unused, cut them rather than pad.

Explore Datasets in Multi-Agent Workbench →