NuCLS — Corrected Single-Rater Subset

Nucleus Classification, Localization and Segmentation in breast cancer (Amgad et al., GigaScience 2022). H&E fields-of-view (FOVs) cropped from TCGA-BRCA whole-slide images at 40× (~0.2 µm/px), with per-nucleus instance + semantic masks and a 13/7/4-class cell-type scheme.

Scope — please read. This repository contains only the corrected (pathologist-QC'd) single-rater subset: 1,744 FOVs / 59,485 nuclei from 124 TCGA-BRCA slides. It does not include the uncorrected single-rater subset (2,168 FOVs) nor the multi-rater evaluation/P-truth sets (the paper's gold unbiased-evaluation tier) — both remain at the official Google Drive. The corrected subset is the QC'd tier most downstream segmentation work uses, and the only one distributed as a single stable archive.

• Modality: Histopathology — H&E brightfield, variable-size FOVs at 40× (~0.2 µm/px)
• Target: nucleus segmentation + classification (breast)
• Ground truth: ready-made 3-channel instance+semantic masks (see encoding below)
• License: CC0 1.0 Universal (the dataset is public domain; the paper text is CC BY 4.0)
• Source: official portal · GitHub PathologyDataScience/NuCLS

⚠️ Hybrid bounding-box + segmentation dataset

Only 33.1% of nuclei (19,680 / 59,485) carry true polygon boundaries (type == 'polyline'). The other 66.7% (39,694) are bounding boxes (type == 'rectangle') rasterized as filled rectangles in the mask. The mask PNG does not distinguish them — the per-FOV type column (preserved in annotations_csv) is the only way to tell true contours from box approximations. Account for this before treating every mask object as a pixel-accurate nucleus.

Mask encoding (mask, 3-channel uint8 PNG)

Channel	Meaning
ch0 (R)	semantic GT_code
ch1 (G)	ROI/foreground flag (1 = inside annotated FOV ROI, 0 = outside; G==1 ⇔ R>0)
ch2 (B)	instance id (0 = outside ROI, 2 = ROI-filler region, 3..N = individual nuclei)

Semantic GT_code (ch0):

Code	Class		Code	Class
0	background (outside ROI)		8	myoepithelium
1	tumor		9	apoptotic_body
2	fibroblast		10	neutrophil
3	lymphocyte		11	ductal_epithelium
4	plasma_cell		12	eosinophil
5	macrophage		99	unlabeled-nucleus
6	mitotic_figure		253	ROI-region (non-nucleus tissue)
7	vascular_endothelium

import numpy as np
m = np.array(row["mask"])          # (H, W, 3) uint8 — read via numpy, do NOT convert("L")
nucleus = (m[..., 0] >= 1) & (m[..., 0] <= 99)   # binary nucleus mask (excludes bg & ROI-region)
inst    = m[..., 2]                # per-nucleus instance ids (ROI-filler id has R==253)
sem     = m[..., 0]               # semantic class codes (table above)

The 13-class raw codes group into the 7 main and 4 super classes used in the paper; the per-nucleus raw/main/super labels are in annotations_csv.

Columns

Column	Type	Notes
fov_id	string	FOV filename stem (encodes slide barcode + WSI crop coords)
image	Image (RGB)	H&E FOV, cropped to align with mask
mask	Image (RGB, 3-ch uint8)	instance+semantic mask, encoding above; cropped to match image
slide_name	string	TCGA slide barcode (e.g. TCGA-A1-A0SP-DX1) — cross-dataset dedup key
tcga_patient	string	TCGA patient barcode (e.g. TCGA-A1-A0SP)
hospital	string	TCGA TSS code (e.g. A1) — the slide-grouping used by the CV folds
fold_1..fold_5	string	train/test membership in each official slide-level CV fold
n_nuclei	int32	nuclei in the FOV (= CSV rows)
n_polyline	int32	nuclei with true polygon boundaries
n_rectangle	int32	nuclei stored as bounding boxes
height, width	int32	aligned image size
superclass_counts	string (JSON)	per-FOV super_classification histogram
annotations_csv	string	raw per-FOV CSV — every nucleus's raw/main/super_classification, type, bbox (xmin..ymax) and polygon (coords_x/coords_y). Lossless

Splits / folds

All 1,744 FOVs are in a single train split. NuCLS ships an official slide-level 5-fold cross-validation (grouping by slide prevents same-slide leakage); membership is preserved as fold_1..fold_5 (each train/test). For a single hold-out benchmark, fold_1 == 'test' (22 slides) is a reasonable default evaluation set. fold_999 (a 1-slide debug fold) is not included.

Provenance, naming and cross-dataset overlap

• Provenance: official author data (PathologyDataScience / Amgad–Cooper lab), CC0. The Dropbox NuCLS_dataset.zip is the exact single-archive re-host of the official "QC" Drive folder; FOV/nuclei counts (1,744 / 59,485 / 19,680 boundaries) match the paper's corrected single-rater numbers exactly.
• Faithful naming: corrected single-rater subset only (see scope note).
• Ground-truth tier: corrected = pathologist-QC'd. The paper's gold unbiased benchmark is the separate multi-rater P-truth set (not hosted here).
• Overlap (leakage hazard): FOVs are cropped from TCGA-BRCA WSIs, so this set shares source slides with other TCGA breast histopathology datasets — most directly BCSS (Breast Cancer Semantic Segmentation; same lab, same slides, NuCLS's 124 ⊂ BCSS's 151), and potentially PanNuke, MoNuSeg, MoNuSAC, Pan-Cancer-Nuclei-Seg. Deduplicate by the slide_name / tcga_patient TCGA barcode before any joint benchmark.

Citation

@article{amgad2022nucls,
  title   = {NuCLS: A scalable crowdsourcing approach and dataset for nucleus
             classification and segmentation in breast cancer},
  author  = {Amgad, Mohamed and Atteya, Lamees A. and Hussein, Hagar and
             Mohammed, Kareem Hosny and Hafiz, Ehab and Elsebaie, Maha A. T. and
             others},
  journal = {GigaScience},
  volume  = {11},
  pages   = {giac037},
  year    = {2022},
  doi     = {10.1093/gigascience/giac037}
}