HEPROBench CRC-CODEX reviewer demo

This is a small real-data software-verification subset for HEPROBench. It contains paired, registered H&E and CRC-CODEX patches derived from Schürch et al., Coordinated cellular neighborhoods orchestrate antitumoral immunity at the colorectal cancer invasive front, created by Christian Schürch, Mendeley Data 10.17632/mpjzbtfgfr.1, licensed under CC BY 4.0.

The bundle has 2 patches from each of 2 anonymized FOVs per split (4 train, 4 validation, and 4 test patches). Cell counts in the included masks are train=721, valid=751, and test=578 after aggregation by anonymous FOV and cell ID.

Contents

• images/: 256x256 RGB H&E JPEGs, re-encoded without EXIF metadata.
• targets/: aligned 256x256x4 uint8 arrays (DAPI, CD3, CD20, PanCK).
• masks/: 256x256 non-negative integer Mesmer cell-ID masks; zero is background.
• metadata.csv and splits/*.csv: the HEPROBench patch contract.
• cell_annotations.csv: continuous cell means and GMM-derived positive labels.
• channel_stats.json, gmm_gates.json, provenance.json, and SHA256SUMS: normalization, label, provenance, and integrity records.

The four reviewer channels are an exact subset of the normalized 58-channel arrays: DRAQ5 is exposed as DAPI, CD3 as CD3, CD20 as CD20, and Cytokeratin as PanCK. Values already lie in the 8-bit range, so selecting the four channels and storing them as uint8 does not rescale or otherwise transform intensity.

The source landing page and its CC BY 4.0 license link were publicly accessible on 2026-09-05. The DataCite record also contained a legacy embargoedAccess entry; current public access and the explicit license were therefore checked directly before this release.

Scope and privacy

This subset is selected for marker/class diversity and is intended only to verify train -> validation/test inference -> image/per-slide/cell evaluation. It is not representative of the full cohort and must not be used to reproduce the paper's scientific performance numbers.

No patient identifiers, original TMA/core names, clinical outcomes, source paths, or source metadata tables are included. FOV and file names are newly assigned per split. JPEG metadata are stripped during export. The images are derived from human biospecimens and should still be handled according to the source dataset's terms and the CC BY 4.0 attribution requirement.