ORION-CRC — 19-channel crops

Overview

ORION-CRC contains 3,337 paired fluorescence and H&E image crops of human colorectal cancer, with masks outlining individual nuclei and whole cells. The crops come from 41 tissue samples representing 40 cases. Each crop is 512 × 512 pixels at 0.325 micrometres per pixel.

The fluorescence images retain all 19 original channels. These show nuclei, immune cells, epithelial cells and other tissue features, and include autofluorescence and blank channels. The H&E images show the same tissue regions.

License: CC BY 4.0. You may use, modify and share the data, including commercially, with credit to the creators, a license link and a description of changes.

Files and splits

Split	File	Crops	Nuclei	Whole cells
train	train.parquet	2,946	550,303	537,862
val	val.parquet	272	41,738	42,118
test	test.parquet	119	19,223	19,063

The original VitaminP train, validation and test assignments are kept. Every case stays within one split. CRC33_01 and CRC33_02 belong to case CRC33 and stay together. Some crops within a split overlap, so object counts can include the same cells more than once.

These are the same regions and mask labels as VitaminP ORION-CRC, with the full original fluorescence panel. Keep matching regions in the same split when combining the two releases. vitaminp_id identifies the matching crop.

Fields

Each row contains one H&E image, its 19-channel fluorescence image and both instance masks.

Field	Contents
id	Unique crop name
image	H&E PNG bytes; RGB uint8, shape (512, 512, 3)
image_if	Fluorescence TIFF bytes; uint16, shape (19, 512, 512), axes channel, height, width
channel_names, ome_channel_names	Channel names in image order, with the original OME names also kept
inst, inst_cell	Nucleus and whole-cell masks; uint32 TIFF bytes, shape (512, 512)
type	Nucleus foreground; uint8 PNG bytes, 0 background and 1 nucleus
pairs	Inferred links with nucleus_id and cell_id
mask_format, mask_dtype	tiff, uint32
image_if_format, image_if_axes, image_if_dtype	tiff, CYX, uint16
case, sample	Source case and tissue-sample names
source_index, source_split, source_path, vitaminp_id	Original VitaminP crop number, split, mask folder and crop name
source_wsi_url, source_wsi_etag	Original fluorescence image location and source version
roi_index	Source selected-region number
x_start, y_start, x_end, y_end	Crop coordinates in the original image; x is column and y is row
mpp_x, mpp_y	Micrometres per pixel: 0.325 in each direction

Original image values and mask numbers are kept. Zero is background in both instance masks. Object numbers can repeat across crops showing the same region.

Fluorescence channel order:

Index	Channel
0	Hoechst
1	AF1
2	CD31
3	CD45
4	CD68
5	Argo550
6	CD4
7	FOXP3
8	CD8a
9	CD45RO
10	CD20
11	PD-L1
12	CD3e
13	CD163
14	E-cadherin
15	PD-1
16	Ki67
17	Pan-CK
18	SMA

AF1 is the autofluorescence channel. Argo550 is named 07_Blank_Argo550 in the original image metadata.

Labels

The nucleus and whole-cell masks come from VitaminP. Segmentation models produced the outlines, and pathologists selected regions where they looked reliable.

Nucleus and cell numbers are independent. pairs links a nucleus to a cell only when every nucleus pixel lies inside that cell. Some cells have several linked nuclei. All unmatched objects and partial overlaps remain in the separate masks. Missing links can result from small differences between the two outlines.

type marks nucleus foreground. The masks do not assign biological cell types.

Download and read a sample

from io import BytesIO

import numpy as np
import polars as pl
import tifffile
from huggingface_hub import hf_hub_download
from PIL import Image

path = hf_hub_download(
    repo_id="csmp-hub/orion-crc", filename="train.parquet", repo_type="dataset"
)
row = pl.scan_parquet(path).head(1).collect().to_dicts()[0]


def read_png(data):
    with Image.open(BytesIO(data)) as image:
        return np.array(image)


image = read_png(row["image"])  # HWC RGB uint8, (512, 512, 3)
image_if = tifffile.imread(BytesIO(row["image_if"]))  # CYX uint16, (19, 512, 512)
nuclei = tifffile.imread(BytesIO(row["inst"]))  # HW uint32, original nucleus numbers
cells = tifffile.imread(BytesIO(row["inst_cell"]))  # HW uint32, original cell numbers
nucleus_classes = read_png(row["type"])  # HW uint8, 0 background / 1 nucleus
pairs = row["pairs"]  # inferred links between nucleus and cell numbers

channel_names = row["channel_names"]
pixel_size_um = row["mpp_x"]  # 0.325 micrometres per pixel

Plot examples

Two training crops are shown below. Fluorescence contrast is adjusted only for display.

import matplotlib.pyplot as plt
from matplotlib.patches import Patch
from skimage.segmentation import find_boundaries

sample_ids = ["orion_crc_CRC01_00000", "orion_crc_CRC02_00000"]
samples = (
    pl.scan_parquet(path)
    .filter(pl.col("id").is_in(sample_ids))
    .sort("id")
    .collect()
    .to_dicts()
)


def display_channel(channel):
    low, high = np.percentile(channel, [1, 99])
    return np.clip((channel.astype(np.float32) - low) / max(high - low, 1), 0, 1)


fig, axes = plt.subplots(2, 5, figsize=(14, 6))
for row_axes, sample in zip(axes, samples):
    rgb = read_png(sample["image"])
    channels = tifffile.imread(BytesIO(sample["image_if"]))
    nuclei = tifffile.imread(BytesIO(sample["inst"]))
    cells = tifffile.imread(BytesIO(sample["inst_cell"]))
    row_axes[0].imshow(rgb)
    for ax, index in zip(row_axes[1:4], [0, 14, 3]):
        ax.imshow(display_channel(channels[index]), cmap="gray", vmin=0, vmax=1)
    overlay = rgb.astype(np.float32) / 255
    overlay[find_boundaries(cells, mode="inner")] = [0, 1, 0.2]
    overlay[find_boundaries(nuclei, mode="inner")] = [0.2, 0.6, 1]
    row_axes[4].imshow(overlay)
    row_axes[0].text(0, 1.02, sample["id"], transform=row_axes[0].transAxes, fontsize=8)
    for ax in row_axes:
        ax.axis("off")
for ax, title in zip(
    axes[0], ["H&E", "Hoechst", "E-cadherin", "CD45", "H&E with outlines"]
):
    ax.set_title(title, pad=25)
fig.legend(
    handles=[
        Patch(color=(0.2, 0.6, 1), label="Nuclei"),
        Patch(color=(0, 1, 0.2), label="Whole cells"),
    ],
    loc="lower center",
    ncol=2,
)
fig.subplots_adjust(
    left=0.01, right=0.99, bottom=0.08, top=0.88, hspace=0.14, wspace=0.03
)
plt.show()

Masks without the images:

from skimage.color import label2rgb

mask_fig, mask_axes = plt.subplots(2, 2, figsize=(7, 7))
instance_colors = [(1, 0.3, 0.3), (0.2, 1, 0.5), (0.3, 0.7, 1), (1, 0.8, 0.2)]
for row_axes, sample in zip(mask_axes, samples):
    for ax, field in zip(row_axes, ["inst", "inst_cell"]):
        labels = tifffile.imread(BytesIO(sample[field]))
        ax.imshow(
            label2rgb(labels, colors=instance_colors, bg_label=0, bg_color=(0, 0, 0)),
            interpolation="nearest",
        )
        ax.axis("off")
    row_axes[0].text(0, 1.02, sample["id"], transform=row_axes[0].transAxes, fontsize=7)
for ax, title in zip(mask_axes[0], ["Nucleus instances", "Whole-cell instances"]):
    ax.set_title(title, pad=25)
mask_fig.subplots_adjust(
    left=0.01, right=0.99, bottom=0.01, top=0.91, hspace=0.14, wspace=0.03
)
plt.show()

Colours distinguish objects within each mask. Matching colours across the two masks do not identify pairs.

Preparation and validation

The fluorescence crops were read from the original ORION-CRC images at the coordinates released by VitaminP. The H&E crops, masks and split assignments were kept from VitaminP.

Every fluorescence crop was compared with its source values, and its Hoechst channel matched VitaminP exactly. Every saved H&E image and both masks were checked against the source. Images were saved without resizing, clipping or normalising their values. Source details, checksums and split checks are in manifest.json, validation.json and split-audit.json.

Limitations

• Model-generated outlines can be inaccurate, and nucleus and cell outlines do not always fit together exactly.
• Some crops within a split overlap. The VitaminP and full-panel releases share regions and labels.
• Marker intensity and image quality vary across samples. Autofluorescence and blank channels are retained as supplied.
• Case grouping and image-similarity checks do not rule out every possible shared region. See the split audit for the scope of the checks.

License and citation

The source dataset and this preparation use CC BY 4.0. Credit both the VitaminP researchers and the original ORION researchers, link to their sources and the license, retain supplied notices, and describe your changes. Full terms are in LICENSE; preparation changes are described in NOTICE.md.

Shokrollahi, Y. (2026). VitaminP: Processed Zarr Datasets for Cross-Modal Whole-Cell Segmentation in Histopathology, version v1.0.0. Zenodo. doi:10.5281/zenodo.21224234.

Shokrollahi, Y.; Pinao Gonzales, K. B.; Barrientos Toro, E. N.; Acosta, P.; Patient Mosaic Team; Chen, P.; Yuan, Y.; Pan, X. (2026). VitaminP: cross-modal learning enables whole-cell segmentation from routine histology. Paper.

Lin, J.-R.; Chen, Y.-A.; Campton, D.; et al. (2023). High-plex immunofluorescence imaging and traditional histology of the same tissue section for discovering image-based biomarkers. Nature Cancer 4, 1036–1052. doi:10.1038/s43018-023-00576-1. Original ORION data.

The HTAN imaging-data policy specifies CC BY 4.0 for the original imaging data. The VitaminP labels also use CC BY 4.0. This preparation combines the original 19-channel crops with VitaminP's paired images, masks and split assignments.

The original ORION data were generated by Lin, Chen, Campton and colleagues at the Harvard tissue-atlas team and RareCyte. Their study was supported by NCI grants U54-CA225088 and U2C-CA233262, NCI SBIR grant R41-CA224503, and RareCyte. Image-processing and data-science work also received support from the Bill and Melinda Gates Foundation (INV-027106), Gray Foundation, David Liposarcoma Research Initiative, Emerson Collective and Ludwig Cancer Research; see the original study for full acknowledgments.