SegPath is released by its authors under CC BY-NC-SA 4.0 for non-commercial use. By requesting access you agree to those terms, to cite the SegPath paper, and to share any derivative under the same licence.

SegPath in the nuclei store format

A faithful conversion of SegPath (Komura and others, 2023) into one store format shared by several histology datasets: images and masks in sharded Zarr, one row per image in Parquet. All 158,687 images are here, at their original 984 px and 0.220818 um/px, in the train, val and test splits as distributed. Every pixel of every image and mask was checked equal to the original PNG files on 03-11-2026. The original data, its licence and its citation requirements are unchanged.

Source and licence

Source: SegPath, eight Zenodo records, one per antibody, listed below. Each H&E section was restained by immunofluorescence for one antibody, and the thresholded signal became the mask.
Licence: CC BY-NC-SA 4.0, as stated by each source record. This copy carries the same licence. Commercial use is not allowed. For commercial use, ask the authors, as the records say.
Cite, as the source asks: Komura, D., Onoyama, T., Shinbo, K., Odaka, H., Hayakawa, M., Ochi, M., Herdiantoputri, R. R., Endo, H., Katoh, H., Ikeda, T., Ushiku, T., Ishikawa, S. (2023). Restaining-based annotation for cancer histology segmentation to overcome annotation-related limitations among pathologists. Patterns, 4(2), 100688. doi:10.1016/j.patter.2023.100688

What changed in the conversion

Pixels: the H&E PNG files, read as RGB and stored unchanged. No crop and no resampling.
Masks: the source's binary mask (0 background, 1 target) becomes one semantic array, whose value is the cell type's code. The codes are shared by all eight stores, so masks can be stacked.
Each image keeps its split from the source's train_val_test column. The source's slide ID is kept in slide.
Each image's annotated_classes names the one cell type it was tested for. A mask says nothing about the other seven.

Layout

One store per antibody, each laid out as below:

native/segpath_erg/card.json        what the dataset is: source, licence, scale and its
                                    evidence, label origin, class names, arrays present
native/segpath_erg/data.zarr        Zarr v3, sharded, Zstd level 3, CRC32C checksums
    images    uint8 (10,647, 3, 984, 984)   RGB
    semantic  uint8 (10,647, 984, 984)      0 background, or the cell type's code
native/segpath_erg/images.parquet   one row per image
native/segpath_erg/objects.parquet  empty: SegPath has no per-nucleus labels

Images stay at the source's size and scale, and nothing here is cut for an experiment. Each image's split is in images.parquet. Slicing, resampling, merging classes and joining datasets happen when an experiment reads the store.

images.parquet columns: image (row number, the index into the arrays), key (the file name without its ending), height, width, um_per_px, split, patient (not reported), slide, organ (not reported), group (the slide), annotated_classes, label_origin, intensity_low and intensity_high (per-channel percentiles), and parent_image, parent_x, parent_y, parent_scale (empty here: they record where a slice came from when one is built from this store).

Store	Antibody	Cell type	Code	train	val	test	Total	Source record
`segpath_panck`	panCK	epithelium	1	21,912	2,259	2,338	26,509	https://zenodo.org/records/7412731
`segpath_asma`	alpha-SMA	smooth muscle	2	25,748	2,489	2,941	31,178	https://zenodo.org/records/7412732
`segpath_cd235a`	CD235a	red blood cell	3	21,595	2,224	2,090	25,909	https://zenodo.org/records/7412580
`segpath_cd3cd20`	CD3 and CD20	lymphocyte	4	10,453	1,082	738	12,273	https://zenodo.org/records/7412529
`segpath_cd45rb`	CD45RB	leukocyte	5	20,518	1,988	2,299	24,805	https://zenodo.org/records/7412739
`segpath_erg`	ERG	endothelium	6	9,497	613	537	10,647	https://zenodo.org/records/7412512
`segpath_mist1`	MIST1	plasma cell	7	11,320	947	964	13,231	https://zenodo.org/records/7412500
`segpath_mnda`	MNDA	myeloid cell	8	12,315	894	926	14,135	https://zenodo.org/records/7412690
All				133,358	12,496	12,833	158,687

Reading the store

With zarr (3.x) and pandas:

import json

import pandas as pd
import zarr

store = "native/segpath_erg"
card = json.load(open(f"{store}/card.json"))
codes = card["declaration"]["semantic_classes"]
root = zarr.open_group(f"{store}/data.zarr", mode="r")
images = pd.read_parquet(f"{store}/images.parquet")
train = images[images["split"] == "train"]
row = train.iloc[0]
image = root["images"][row.image]      # (3, 984, 984) uint8
mask = root["semantic"][row.image]     # (984, 984) uint8
cell_type = row.annotated_classes[0]   # the one type this image was tested for
target = mask == codes.index(cell_type)

Limits

A mask marks one cell type. A pixel of 0 means not that type, not background of every type.
The distributed splits share slides: patches of one slide appear in train, val and test. An evaluation that needs unseen slides must build its own split from slide.
slide holds the antibody, the cell type and a slide number, such as ERG_Endothelium_040. Inside one antibody, group by slide. Slide numbers repeat across antibodies, and the source records do not say whether one number is the same tissue in each. Until that is confirmed, group across antibodies by the trailing number, which can only make a split stricter.
Masks come from thresholded immunofluorescence, so they inherit its staining and registration errors. Nothing was corrected.
Models trained on this data inherit the non-commercial licence.

Downloads last month: -

Total file size:

386 GB