This dataset contains de-identified clinical white-blood-cell images used in the WBCBench 2026 ISBI challenge. Access requires accepting the data-use agreement below.

By requesting access, you agree to the following:

You will use this dataset for non-commercial research and teaching only, under the WBCBench 2026 Data Use Licence (see LICENSE).
You will cite the WBCBench 2026 paper (Tian et al., ISBI 2026) in any publication or presentation that uses this dataset.
You will not attempt to re-identify any patient in the data, nor link it to external clinical records.
You will not redistribute the dataset or share your access with third parties.
You will not use this dataset to train commercial products or services (model training intended for commercial deployment is prohibited).
You will report any data leak or unauthorized access to the WBCBench 2026 team.
You will keep the data on secure systems that only you can access, and delete all copies if your access ends.

WBCBench 2026 - Robust White Blood Cell Classification Under Class Imbalance

Single-cell images of white blood cells from peripheral blood films, labelled into 13 cell types. The dataset was used in the WBCBench 2026 ISBI challenge, which tests classifiers under severe class imbalance and realistic image degradation.


Images	101,716: 55,002 original ("pristine") images and 46,714 degraded copies
Cases	493 blood films (one laboratory accession each)
Classes	13 white-cell types (see below)
Image format	RGB JPEG, 368 × 370 or 368 × 368 pixels (width × height)
Source	King Chulalongkorn Memorial Hospital, Chulalongkorn University (slides analysed 2023–2024)
Version	1.0 (October 2026)
Licence	WBCBench 2026 Data Use Licence v1.0 (`LICENSE`): non-commercial research and teaching only, no redistribution. Scripts in `scripts/` are MIT.
Paper	arXiv:2604.10797
Challenge	Website · Kaggle
Zenodo	Same data as zip files, access on request: doi:10.5281/zenodo.23544488
Contact	Xin Tian (xin.tian@well.ox.ac.uk) · Nantheera Anantrasirichai (N.Anantrasirichai@bristol.ac.uk)

Access

Click Request access at the top of this page, fill in the form and accept the data-use agreement. Access is granted automatically.

You may publish results, aggregate statistics and analysis code, show up to 10 example images in a publication or presentation (with citation), and share model weights for non-commercial use. The data themselves must not be shared; everyone who needs them must request access. The corresponding author mentioned in the licence is Xin Tian (xin.tian@well.ox.ac.uk).

Splits

Split	Cases	Pristine images	Degraded images	Use
`phase1_train`	74	8,288	–	Training
`phase2_train`	222	24,897	24,897	Training
`phase2_eval`	49	5,350	5,350	Validation (may also be used for training)
`phase2_test`	148	16,467	16,467	Test (official leaderboard)

Cases were assigned to splits at random (about 15%, 45%, 10% and 30% of cases), with the five cases that contain prolymphocytes spread over all four splits.

Each degraded image is a copy of a pristine image from the same split; original_image_id links the two. Labels are included for all splits, including phase2_test.

Challenge setting. Phase 1 of the challenge released the pristine phase1_train images. Phase 2 released only the degraded phase2_* images, and the leaderboard was scored on the degraded phase2_test images. The pristine copies of the phase2_* splits were not part of the challenge.

Cases, not patients. Each case (patient_hash) is one blood film, and every case appears in only one split. Some people had more than one film, so the splits are not fully separated by person. Where films could be linked to hospital records (340 of 493 cases), they come from 277 people: 36 people have 2–7 cases, and 24 of them appear in more than one split (about 14% of phase2_test images come from these 24 people). Their cases are different films taken on different days, not duplicate images. The other 153 cases could not be linked, so these figures are minimums. Please describe patient_hash as a case, not a patient. A person-level grouping key may be added in a future version if approved by the data provider.

Use patient_hash to group cells from the same case for cross-validation.

Changes since the challenge

Some test labels were re-checked. After the challenge, the 341 phase2_test images that all of the top 10 teams misclassified were re-checked by an expert: 257 labels were corrected, 10 images were removed (9 showed more than one cell, 1 was too difficult to classify) and 74 labels were confirmed (16,477 → 16,467 test images). No other images were re-checked. metadata/label_changes.csv lists these 267 images with their pristine and degraded IDs and their challenge-era and current labels. Scores computed with these labels are not directly comparable with the Kaggle leaderboard or the challenge papers.
Differences from the challenge paper. The paper describes 55,012 images of 368 × 368 pixels, split by patient with group-stratified sampling. This release has 55,002 pristine images (after the review), two image sizes (368 × 370 and 368 × 368) and random case-level splits (see Splits and Cases, not patients).

Cell types

Label	Cell type	Pristine images
SNE	Segmented neutrophil	25,532
LY	Lymphocyte	15,100
MO	Monocyte	5,194
EO	Eosinophil	1,640
BA	Basophil	857
VLY	Variant (atypical) lymphocyte	711
BNE	Band-form neutrophil	648
MMY	Metamyelocyte	650
MY	Myelocyte	792
PMY	Promyelocyte	192
BL	Blast cell	3,568
PC	Plasma cell	102
PLY	Prolymphocyte	16

The same list is in metadata/class_legend.csv.

Columns

Column	Description
`image`	The cell image (JPEG)
`image_id`	Image ID, e.g. `00173214`
`label`	Cell type (one of the 13 labels above)
`wbcbench_split`	`phase1_train`, `phase2_train`, `phase2_eval` or `phase2_test`
`severity`	`pristine`, `mild`, `moderate` or `extreme` (always `pristine` in the pristine config)
`original_image_id`	For degraded images: the ID of the pristine source image (empty otherwise)
`patient_hash`	Case ID: a salted hash of the laboratory accession number (one blood film, not one person)

All rows are also listed in metadata/labels.csv, which has one extra column, folder (pristine or degraded).

Quick start

pip install -U datasets huggingface_hub pillow
hf auth login   # use the account that requested access (browser login, or paste a "Read" token from https://huggingface.co/settings/tokens)

In a notebook, run from huggingface_hub import login; login() instead of hf auth login.

from datasets import load_dataset

ds = load_dataset("Xin-Tian/wbcbench2026", "pristine", split="phase2_train")
row = ds[0]
print(row["image_id"], row["label"], row["image"].size)   # 00004004 MY (368, 370)

Configs: pristine (all four splits) and degraded (the three phase2_* splits).
Loading any split downloads its whole config (about 1.1 GB for pristine, 2.8 GB for degraded) and keeps a converted copy of the same size, so allow about twice that in disk space.
With streaming=True, read rows with next(iter(ds)) instead of ds[0]. Each data file (up to about 525 MB) is still read in full before its first row is returned.

Class IDs. label is a string. To get the same class IDs in every split, cast it with the full list of 13 labels:

from datasets import ClassLabel

LABELS = ["SNE", "LY", "MO", "EO", "BA", "VLY", "BNE", "MMY", "MY", "PMY", "BL", "PC", "PLY"]
ds = ds.cast_column("label", ClassLabel(names=LABELS))

Source of a degraded image. original_image_id is the image_id of the pristine image in the same split:

deg = load_dataset("Xin-Tian/wbcbench2026", "degraded", split="phase2_train")
index = {iid: i for i, iid in enumerate(ds["image_id"])}
source = ds[index[deg[0]["original_image_id"]]]

Metadata files. load_dataset does not load the files in metadata/; download them directly, and read them with dtype=str so image IDs keep their leading zeros:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("Xin-Tian/wbcbench2026", "metadata/labels.csv", repo_type="dataset")
labels = pd.read_csv(path, dtype=str)

Saving images exactly as published. You can train on the decoded images directly. To write the files to disk, don't use image.save(), which re-encodes the JPEG; write the original bytes instead:

import os
from datasets import Image

os.makedirs("out", exist_ok=True)
raw = ds.cast_column("image", Image(decode=False))
for row in raw:
    with open(f"out/{row['image_id']}.jpg", "wb") as f:
        f.write(row["image"]["bytes"])

Degradations

The degraded images simulate imperfect imaging. Each degraded image has one severity level:

pristine: no degradation applied. The image was re-saved as JPEG, so it is not byte-identical to its source (mean difference about 1 grey level).
mild: 1–3 operations drawn, with weak settings
moderate: 1–4 operations drawn, with stronger settings
extreme: 3–5 operations drawn, with the strongest settings

Operations are drawn at random from crop-and-resize, Gaussian blur, motion blur, Gaussian noise, Poisson noise, saturation, gamma, brightness, vignetting and color jitter. Motion blur and color jitter are then applied only with a set probability, so some images received fewer operations than were drawn (about 5% of mild and 2% of moderate images received none). The mix of severities differs by split: pristine-level images make up 7% of degraded phase2_train, 40% of phase2_eval and 35% of phase2_test. Prolymphocyte (PLY) images were given only pristine or mild severity.

The settings used for each image are in metadata/degradation_params.csv, one row per degraded image. Rows for pristine-level images have no settings. In the other rows, an empty cell means that operation was not drawn (for color jitter, color_jitter_enable is False instead); motion blur and color jitter settings are listed whenever they were drawn, even if they were not applied; the rotate_* columns are 0 because rotation was disabled; and rng_seed is the image's random seed. Join it on new_name, which is the degraded image_id plus .jpg (orig_newname is the source image). The code is in scripts/.

How the data were collected

Specimens: peripheral blood films, prepared on automated Sysmex instruments and stained in-house with Wright-Giemsa using a standardised automated protocol.
Imaging: CellaVision DI-60 analysers with a 100× oil-immersion objective. The analyser finds and crops each white cell automatically. The physical pixel size is not recorded.
Two analysers: 368 × 370 images come from one analyser (218 cases) and 368 × 368 images from another in the same laboratory (275 cases). All images of a case come from one analyser. The 368 × 368 analyser was used throughout (January 2023 to July 2024), but the 368 × 370 one only from September 2023 to June 2024, so image size is partly linked to time.
Cells per case: the white cells from the analyser's verified differential for that slide, minus the exclusions below (median 112 per case). Cells were not chosen by type, apart from the exclusions below. Non-white-cell objects (NRBCs, platelets, smudge cells, artifacts) are not included.
Labels: each image was labelled by an experienced haematopathologist; ambiguous cases were discussed by five haematopathologists in consensus sessions. Some test labels were corrected after the challenge (see Changes since the challenge).
Exclusions: 1,019 of 56,021 exported images (1.8%) were removed: crops with more than one cell, objects outside the 13 classes, images without a label record, and the 10 images removed in the review. Most cases lost two images or fewer; the largest loss was 38% of one case.
Class balance: the classes are very imbalanced and do not reflect a general population (for example, blast cells appear in 215 of the 493 cases).

Evaluation

The official challenge metric is macro-averaged F1 over the 13 classes, so every class counts equally regardless of its size. Ties were broken by balanced accuracy, then macro precision, then macro specificity. phase2_test contains only 2 prolymphocyte (PLY) and 12 plasma cell (PC) images, so per-class scores for these two classes are unstable.

Participant code

Code released by WBCBench 2026 participants. It is provided by the authors and has not been tested by the WBCBench 2026 team.

Paper	Code
Nguyen et al., Synergizing Deep Learning and Biological Heuristics for Extreme Long-Tail White Blood Cell Classification (arXiv:2603.16249)	https://github.com/trongduc-nguyen/WBCBench2026
Le et al., Robust White Blood Cell Classification with Stain-Normalized Decoupled Learning and Ensembling (arXiv:2603.01976)	https://github.com/luulecmg/WBCBench2026
Gitau et al., Multi-Stage Fine-Tuning of Pathology Foundation Models with Head-Diverse Ensembling for White Blood Cell Classification (arXiv:2603.20383)	https://github.com/Antony-gitau/msfit
Srivastava et al., Ensemble of Small Classifiers For Imbalanced White Blood Cell Classification (arXiv:2603.20856)	https://gitlab.com/siddharthsrivastava/wbc-bench-2026 (branch `isbi-final`)

To have your WBCBench 2026 code listed here, please contact the WBCBench 2026 team.

Privacy and ethics

Direct identifiers (laboratory accession numbers, hospital numbers, source paths, scan dates, analyser serial numbers) have been removed. patient_hash is a truncated SHA-256 of a private salt plus the accession number, so cases can be grouped without revealing the accession number. The images contain no EXIF metadata. OpenCV may warn about "extraneous bytes" when reading the pristine images; this is harmless padding left by the original JPEG encoder.

Ethics. The data were commercially obtained from Chulalongkorn University; additional ethical approval was not required (see the challenge paper).

Citation

@inproceedings{wbcbench2026,
  title={{WBCBench} 2026: A challenge for robust white blood cell classification under class imbalance},
  author={Tian, Xin and Ma, Xudong and Yang, Tianqi and Achim, Alin and Papie{\.z}, Bart{\l}omiej W and Watanaboonyongcharoen, Phandee and Anantrasirichai, Nantheera},
  booktitle={2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI)},
  pages={1--4},
  year={2026},
  organization={IEEE},
  doi={10.1109/ISBI61048.2026.11515899}
}

Downloads last month: 186

Total file size:

3.94 GB

Papers for Xin-Tian/wbcbench2026

WBCBench 2026: A Challenge for Robust White Blood Cell Classification Under Class Imbalance

Paper • 2604.10797 • Published Apr 12

Synergizing Deep Learning and Biological Heuristics for Extreme Long-Tail White Blood Cell Classification

Paper • 2603.16249 • Published Mar 28

Ensemble of Small Classifiers For Imbalanced White Blood Cell Classification

Paper • 2603.20856 • Published Mar 21

Multi-Stage Fine-Tuning of Pathology Foundation Models with Head-Diverse Ensembling for White Blood Cell Classification

Paper • 2603.20383 • Published Mar 20

Robust White Blood Cell Classification with Stain-Normalized Decoupled Learning and Ensembling

Paper • 2603.01976 • Published Mar 2