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

MoNuSAC in the nuclei store format

A faithful conversion of the MoNuSAC 2020 dataset (Verma and others, 2021) into one store format shared by several nuclei datasets: images and masks in sharded Zarr, one row per image and one row per nucleus in Parquet. Every image's pixels were checked equal to the original .svs file and 46,885 of the 46,899 stored nuclei were matched to their original XML outline by an independent rasteriser on 31-10-2026, with equal classes for all but 1 (a nucleus overlapping a neighbour of another class). Of the 14 unmatched, 13 are outlines that cross themselves in the XML: the store repairs them into their full area, where a plain rasteriser leaves holes; one is a three-vertex outline a few pixels wide. The original data, its licence and its citation requirements are unchanged.

Source and licence

Source: MoNuSAC 2020 challenge, https://monusac-2020.grand-challenge.org, training and test data as distributed: 209 training and 101 test regions of varied size from TCGA whole-slide images of four organs (lung, prostate, kidney and breast), every nucleus outlined and classed by hand in Aperio ImageScope as epithelial, lymphocyte, macrophage or neutrophil, with ambiguous areas marked separately.
Licence: CC BY-NC-SA 4.0, as stated on the source page. This copy carries the same licence. Commercial use is not allowed.
Cite, as the source asks: Verma, R., Kumar, N., Patil, A., Kurian, N. C., Rane, S., Graham, S., Vu, Q. D., Zwager, M., Raza, S. E. A., Rajpoot, N., and others (2021). MoNuSAC2020: a multi-organ nuclei segmentation and classification challenge. IEEE Transactions on Medical Imaging, 40(12), 3413 to 3423.

What changed in the conversion

Pixels come from the .svs file of every region, read as RGB and stored unchanged. The distribution also holds .tif copies of 294 regions whose red and blue channels are swapped; they are not used.
Labels: each XML outline of the four nucleus classes is drawn into one instance map per image, larger outlines first so a small nucleus drawn inside a larger one keeps its pixels; a region with no pixel of its own or no centre inside the image is dropped. 10 of the 46,909 XML nuclei were dropped this way, leaving 46,899. The outline itself is kept as each nucleus's geometry, and the class is the XML layer's name in lower case.
Ambiguous areas are not nuclei: they become valid 0, 1.78% of all pixels, so a model is neither trained nor scored there.
Pixel size: each XML states MicronsPerPixel and it is kept per image. Patient and slide come from the TCGA folder names; organ is not recorded per image by the source files.

Layout

native/monusac/card.json         what the dataset is: source, licence, scale and its
                                 evidence, label origin, class names, arrays present
native/monusac/data.zarr         Zarr v3, sharded, Zstd level 3, CRC32C checksums
    images     uint8  (310, 3, H, W)   RGB, padded to the largest region;
                                                  true sizes are in images.parquet
    instances  uint32 (310, H, W)      0 background, one id per nucleus
    valid      uint8  (310, H, W)      0 over ambiguous areas and padding
native/monusac/images.parquet    one row per image
native/monusac/objects.parquet   one row per nucleus (GeoParquet, outline polygon)
patches/monusac/                 the same, cut into 2887 patches of 256 px
                                 at 0.25 um/px; a region smaller than a patch is padded
                                 and the padding is not valid; a nucleus crossing a patch
                                 edge appears in both patches
views/monusac_splits.parquet     the source's own split as roles: train and test; a
                                 validation split is the user's choice

images.parquet columns: image (row number, the index into the arrays), key (train/<region> or test/<region>), height, width (the true size inside the padded array), um_per_px (per image), split, patient (TCGA participant), slide, organ (not reported), group, annotated_classes, label_origin, intensity_low and intensity_high (per-channel percentiles), and parent_image, parent_x, parent_y, parent_scale (filled in the patches tier).

objects.parquet columns: image, object (the nucleus's id in instances), native_class, label_origin, x_min, y_min, x_max, y_max (bounding box in pixels) and geometry (the outline, WKB polygon).

Class names are MoNuSAC's own, never regrouped. A class code is the position in card.json's native_classes, counted from 1; 0 means no class:

code	native_class	nuclei
1	epithelial	21,746
2	lymphocyte	23,457
3	macrophage	894
4	neutrophil	802

split	images	nuclei
train	209	31,407
test	101	15,492

Region sizes run from 33 to 2500 rows and 35 to 2162 columns. Nuclei per image: minimum 2, median 59, maximum 861.

Reading the store

With zarr (3.x), pandas and shapely:

import pandas as pd
import shapely.wkb
import zarr

root = zarr.open_group("native/monusac/data.zarr", mode="r")
images = pd.read_parquet("native/monusac/images.parquet")
row = images.iloc[0]
image = root["images"][0, :, : row.height, : row.width]      # (3, h, w) uint8
instances = root["instances"][0, : row.height, : row.width]  # (h, w) uint32
valid = root["valid"][0, : row.height, : row.width]          # (h, w) uint8, 0 = ignore
objects = pd.read_parquet("native/monusac/objects.parquet")
outline = shapely.wkb.loads(bytes(objects["geometry"][0]))

With pathomix[nuclei-datasets], the package this format belongs to, a store opens with Store.open, a view loads with load_role, and a PyTorch loader comes from pathomix.nuclei.datasets.loading; both read the patches tier and carry valid.

Limits

Outlines and classes are hand-made and inherit the source's judgement calls; nothing was corrected. Macrophage outlines are markedly larger than the other classes and may trace whole cells rather than nuclei; treat that class with care.
Training on this store needs a loss that honours valid; ambiguous areas and padding are not background.
Organ is not recorded per image by the source files.
Models trained on this data inherit the non-commercial licence.

Downloads last month: 21

Total file size:

868 MB