UnumLocalia

Matched H&E, spatial RNA and protein images from four human tissue cores: hepatocellular carcinoma, matched non-tumour liver, tonsil and hepatocellular adenoma. Xenium provides cell and nucleus boundaries alongside a 5,001-target RNA panel. COMET adds 15 protein markers and DAPI. The measurements and H&E come from the same tissue section.

The data are CC BY 4.0. Credit Duchini and colleagues and Felix Marsh-Wakefield, cite the study and original release, and describe any changes when sharing derived data. This version crops images and converts supplied boundaries into pixel masks.

Images and masks

There are 513 H&E crops, each 512 × 512 pixels at about 0.203125 µm per pixel. H&E pixels are unchanged. Each crop has separate nucleus and cell instance masks. Background is 0. Original Xenium numeric IDs are preserved; pairs links nucleus IDs to cell IDs. The full source has 46,076 cells and 45,891 nucleus instances. Some cells have multiple nuclei and 2,121 cells have no supplied nucleus boundary. Crops can cut through instances; the full links and boundaries are retained in objects.parquet and source/.

image_if holds all 16 COMET channels as a native uint16 CYX TIFF crop. The IF crop has its own size and pixel grid at 0.28 µm per pixel. Its values are unchanged. if_x, if_y and he_to_if give the coordinates needed to compare it with H&E. if_valid marks H&E pixels covered by the IF image. Do not place the two arrays directly on top of each other without this mapping. Some H&E crops extend beyond the available COMET image. has_if marks whether a native IF crop is available; otherwise image_if is empty and if_valid is 0.

The channel list includes DAPI, CK8, VEGF, CD45RA, vimentin, E-Cadherin, CD1c, IL-10, podoplanin, FXIIIa, CD66, CD20, alpha-SMA, CD45, HepPar1 and CD206. Original channel strings are in each row. The image metadata and measurement tables spell alpha-SMA differently; the original spellings are retained.

The COMET/InstanSeg segmentation is provided separately in inst_comet_cells and inst_comet_nuclei. It is an alternative segmentation, with its own IDs. Its numeric mask IDs map to original feature UUIDs in objects.parquet. Do not join these IDs to Xenium IDs.

COMET masks are unavailable for all 123 non-tumour liver crops because the source predictions overlap extensively. has_comet_masks is false, both COMET mask fields are empty, and skip_training_comet is true for those crops. Their Xenium masks remain available. Original COMET geometry is retained for provenance and marked unavailable for use as a mask in objects.parquet. The remaining 390 crops retain their COMET alternatives; six exact crop pairs were visually approved. Other COMET crops remain unreviewed. Native and COMET quality flags are independent.

Cell types and measurements

No ready-made cell-type classes are supplied. type and type_cell use 255 for unknown instances and 0 for background. Tissue diagnosis describes the core, not every cell. No cell types were inferred for this version.

measurements/ contains the original per-cell RNA and protein tables, compressed without changing their CSV contents. Join them to the Xenium objects by core and string cell_id. Their numeric label column is an export ID and differs from boundary IDs. There are 73 boundary cells without a measurement row.

instance-counts.csv records each crop's counts for both segmentation alternatives. class-distribution.csv gives source and exported unique instance counts. Crops near image edges can overlap, so adding per-crop counts repeats some cells.

Annotation	Source instances	Unique instances represented in crops
Xenium nuclei	45,891	45,890
Xenium cells	46,076	46,075
COMET nuclei	42,417	33,950
COMET cells	51,533	39,438

Splits and quality

This version has one training split. There are three patient groups: the tumour and non-tumour liver cores share one patient; tonsil and adenoma belong to two other donors. The group names are assigned here from the study description, not original patient identifiers. Keep each patient together when making splits. There is no independent validation or test set in this version.

Eight crops were visually reviewed, with nuclei and cells checked separately. All 16 decisions were usable. The other 505 crops remain unreviewed. The COMET alternatives were reviewed separately on the same eight crops. Six crop pairs were usable; the two non-tumour liver pairs were excluded. nuclei_quality_status and cells_quality_status preserve separate decisions; human_training_approved is true only when both native masks for the exact crop were approved. skip_training flags excluded or uncertain crops and overlaps with rejected areas. No Xenium crops were excluded in the native-mask review.

A usable review does not certify every cell, cell type or complete annotation. These are generated segmentations and can miss, split or merge cells. The supplied alignment has local errors, particularly near some tissue boundaries. The masks are not certified as complete benchmark labels.

Some original polygons cross themselves. Their original boundaries are retained; for pixel masks, valid components are recovered with Shapely make_valid and flagged. Where different instances claim the same pixel, that pixel is set to 0 and marked invalid. Use nuclei_valid with inst, and cells_valid with inst_cell, to ignore repaired or conflicting pixels in segmentation losses. The COMET masks have their own corresponding validity masks. Zero at an invalid pixel does not mean verified background. The quality flags record the current review decisions.

Load an example

from io import BytesIO

import matplotlib.pyplot as plt
import numpy as np
import polars as pl
import tifffile
from PIL import Image

sample = (
    pl.scan_parquet("train.parquet")
    .filter(pl.col("has_if"))
    .select("image", "inst", "inst_cell", "image_if")
    .limit(1)
    .collect()
    .row(0, named=True)
)
image = np.array(Image.open(BytesIO(sample["image"])))
nuclei = tifffile.imread(BytesIO(sample["inst"]))
cells = tifffile.imread(BytesIO(sample["inst_cell"]))
markers = tifffile.imread(BytesIO(sample["image_if"]))

fig, axes = plt.subplots(1, 4, figsize=(16, 4))
for ax, values, title in zip(
    axes,
    [image, nuclei, cells, markers[0]],
    ["H&E", "Nucleus IDs", "Cell IDs", "Native IF: DAPI"],
):
    ax.imshow(values, cmap=None if values.ndim == 3 else "nipy_spectral")
    ax.set_title(title)
    ax.axis("off")
plt.tight_layout()

Reference

Duchini E, Tsao C, Madore J, Ashhurst T, De Almeida Silva J, Shin J, Gupta R, McCaughan G, Palendira U, Liu K, Ferguson A, Marsh-Wakefield F. An open multimodal spatial resource integrating same-tissue transcriptomics, proteomics, and histology. bioRxiv, 2026. Study. Data.