MoNuSAC2020

Multi-organ Nuclei Segmentation And Classification Challenge 2020 (ISBI 2020 satellite challenge). H&E-stained histopathology sub-images cropped from TCGA whole-slide images (scanned at 40x), with per-nucleus instance segmentation and 4-way cell-type classification across four organs.

Composition

• 310 RGB tiles (variable size, ~33x35 .. 2500x1987 px) = 209 train + 101 test
• 71 patients (46 train + 25 test; no patient overlap between splits)
• 4 organs (tissue): Breast, Kidney, Lung, Prostate
• 46,909 scored nuclei (31,411 train + 15,498 test) across 4 classes, plus 2,403 test-only Ambiguous don't-care regions

Split	Images	Patients	Scored nuclei	Ambiguous	Organs (Breast/Kidney/Lung/Prostate images)
train	209	46	31,411	0	58 / 55 / 39 / 57
test	101	25	15,498	2,403	22 / 31 / 22 / 26

Schema

Field	Type	Description
patient	str	12-char TCGA patient barcode (e.g. TCGA-73-4668)
tissue	ClassLabel(4)	Organ: 0 Breast, 1 Kidney, 2 Lung, 3 Prostate
image	PIL RGB	H&E tile, uint8
type_map	PIL grayscale uint8 (L)	Semantic class per pixel (see table below)
inst_map	PIL grayscale uint16 (I;16)	Unique nucleus id per pixel (0 = background)
num_nuclei	int32	Count of the 4 scored classes (excludes Ambiguous)
num_ambiguous	int32	Count of Ambiguous / don't-care nuclei (test only)

Semantic class encoding (type_map)

Value	Class	Scored?
0	Background	-
1	Epithelial	yes
2	Lymphocyte	yes
3	Macrophage	yes
4	Neutrophil	yes
5	Ambiguous	no — don't-care (faint / uncertain / out-of-challenge nuclei; test set only; excluded from challenge scoring)

The four scored classes are the challenge ground truth. Class 5 (Ambiguous) should be excluded from evaluation (treated as a don't-care / ignore region). Ambiguous instances were painted before the scored classes, so a scored nucleus always wins where the two rarely overlap.

Instance map (inst_map)

Each connected nucleus gets a unique 16-bit id per tile (from 1; max ~870 per tile). The class of instance k is recoverable as the type_map value at any pixel where inst_map == k.

Provenance & derivation

• Canonical source: official MoNuSAC2020 Grand Challenge release (organizer Google Drive: .svs/.tif tiles + Aperio ImageScope polygon XML), https://monusac-2020.grand-challenge.org/ .
• Immediate source: the RationAI/MoNuSAC parquet mirror (Masaryk University), which rasterizes each XML Region polygon to a binary mask at full image resolution. We verified its per-class instance counts reconcile exactly to the paper (31,411 train; 15,498 test scored, i.e. 17,901 test instances minus 2,403 Ambiguous).
• Our processing: merged the per-instance masks + categories into the dense type_map / inst_map pair above. No resampling or intensity change to the RGB tiles.

Cross-dataset overlap (leakage caveat)

All tiles come from TCGA, and patient is the TCGA barcode. The same TCGA cases may appear in other nuclei datasets — notably MoNuSeg (same lab, same 4 organs; binary segmentation only), PanNuke, NuCLS (TCGA-BRCA), and CryoNuSeg. Deduplicate by TCGA barcode before any joint benchmark.

License

CC BY-NC-SA 4.0 (non-commercial, share-alike, attribution). Underlying WSIs are public TCGA (NIH) data.

Citation

@article{9446924,
  author={Verma, Ruchika and Kumar, Neeraj and Patil, Abhijeet and Kurian, Nikhil Cherian and Rane, Swapnil and Graham, Simon and Vu, Quoc Dang and Zwager, Mieke and Raza, Shan E. Ahmed and Rajpoot, Nasir and others},
  journal={IEEE Transactions on Medical Imaging},
  title={MoNuSAC2020: A Multi-Organ Nuclei Segmentation and Classification Challenge},
  year={2021},
  volume={40},
  number={12},
  pages={3413-3423},
  doi={10.1109/TMI.2021.3085712}
}