HEMATO: blood and bone marrow cell images

HEMATO collects microscopy images from several public hematology datasets and puts them under one set of class names and one split scheme. There are four subsets, each for a different task. The current release is v4.0 (357,856 images).

This is a research and teaching dataset. It is not meant for clinical diagnosis.

Subset	Task	Classes	Images	Train / val / test
cls	white cell and bone marrow cell types	18	311,875	219,878 / 46,355 / 45,642
malaria	infected vs. uninfected red cells	2	32,450	24,651 / 3,995 / 3,804
all	acute lymphoblastic leukemia vs. normal	2	12,068	8,809 / 1,390 / 1,869
det	detection of RBC, WBC and platelets (YOLO)	3	1,463	1,020 / 243 / 200

Loading

from datasets import load_dataset
ds = load_dataset("KeremTT/hemato", "cls")   # or "malaria", "all"

Each row has the image, the label, a Turkish label name, the source dataset, the label used in the source, a patient or group id where one exists, and the source license. The detection subset is in det/ in YOLO format. zips/ has the same data as folder-per-class ZIP files.

cls

Single-cell images from bone marrow smears (Matek et al. 2021) and peripheral blood (Acevedo et al. 2020), extended in later versions with MLL23, KU-Optofil PBC, the multi-focus WBC set of Park et al. (2024), AML-Cytomorphology_LMU (Matek et al. 2019) and BCDD-SUST.

Class	Images	Class	Images
Segmented_neutrophil	72,258	Platelet	13,341
Lymphocyte	52,027	Monocyte	12,589
Erythroblast	31,047	Myelocyte	9,667
Blast	26,542	Plasma_cell	9,168
Artefact	19,979	Metamyelocyte	5,335
Band_neutrophil	15,798	Hairy_cell	3,669
Promyelocyte	15,454	Not_identifiable	3,538
Eosinophil	14,717	Basophil	3,018
Proerythroblast	2,740	Smudge_cell	988

Source labels were mapped to these names only when they meant the same thing. Reactive and large granular lymphocytes are counted as Lymphocyte, atypical promyelocytes as Promyelocyte, giant platelets and platelet clusters as Platelet. The original label is kept in original_label. Smudge_cell comes from one source only.

malaria

Segmented red cells from NIH/NLM (P. falciparum, 13,779 per class) and crops from BBBC041 (P. vivax, 2,446 per class). Both sources are balanced 1:1. The NIH part is split by patient, the BBBC041 part by field of view.

all

Segmented lymphoid cells from the C-NMC 2019 challenge (training set plus the preliminary test set), split by patient (100 patients).

det

BCCD plus TXL-PBC, with BCCD duplicates removed and class ids mapped to 0 Platelets, 1 RBC, 2 WBC.

How it was built

• Splits are 70/15/15. When a source has patient, smear or field ids, all images of one group stay in one split. MLL23, AML-LMU, TXL-PBC and the original bone marrow and PBC sets have no patient ids, so those are split by image.
• Each version keeps the earlier splits. The v1.0 test split has not changed.
• Images were checked for corrupt files, exact duplicates (SHA-256) and near duplicates (pHash, checked at pixel level), including duplicates across splits and the same image with two labels. Greyscale or unstained crops and crops with large black padding were removed. Thresholds were set after looking at flagged examples.
• The full decision log and per-version reports are in metadata/ (Turkish).

Things to keep in mind

• Class sizes are uneven (988 to 72,258 in cls).
• Sources differ in microscope, stain and crop size. Report results per source as well as overall, otherwise a model can look good by recognising the source.
• In cls, part of the data has no patient ids, so test scores there are likely a bit optimistic.
• 7,500 BCDD-SUST platelet crops in cls are blurry. Filter on source == "bcdd_sust" if you need sharp platelet images.

Licenses and citation

Each image keeps the license of its source (source_license column). Please cite the original papers.

Source	Used in	License
Matek et al. 2021, Blood (TCIA, doi:10.7937/TCIA.AXH3-T579)	cls	CC BY 4.0
Acevedo et al. 2020, Data in Brief (doi:10.17632/snkd93bnjr.1)	cls	CC BY 4.0
MLL23, Zenodo (doi:10.5281/zenodo.14277609)	cls	CC BY 4.0
KU-Optofil PBC, Zenodo (doi:10.5281/zenodo.17333317)	cls	CC BY 4.0
Park et al. 2024, Scientific Data (doi:10.6084/m9.figshare.26781052)	cls	CC0
Matek et al. 2019, Nature Machine Intelligence (TCIA, doi:10.7937/tcia.2019.36f5o9ld)	cls	CC BY 3.0
BCDD-SUST, Figshare (doi:10.6084/m9.figshare.33951064)	cls	CC BY 4.0
Gupta & Gupta 2019, C-NMC 2019 (TCIA, doi:10.7937/tcia.2019.dc64i46r)	all	CC BY 3.0
Rajaraman et al. 2018, PeerJ (NIH/NLM malaria datasheet)	malaria	public (U.S. NLM)
Hung et al. 2018, BBBC041 (Broad Bioimage Benchmark Collection)	malaria	CC BY-NC-SA 3.0
BCCD (Shenggan 2017), Roboflow 100 version	det	MIT / CC BY 4.0
Lu et al. 2024, TXL-PBC (Figshare 27073186)	det	CC BY 4.0

Because of BBBC041, the malaria subset can only be used non-commercially and shared under the same license.

Versions

Version	Changes
v1.0	4 subsets, 225,909 images
v2.0	8 more sources, 377,451 images
v3.0	more malaria cells and six new cls classes, 420,581 images
v4.0	malaria balanced per source; five of the six v3.0 subtype classes merged back into their parent classes; 357,856 images