KeenForgeAI/TXL-PBC-corrected

TXL-PBC-corrected — Peripheral Blood Cell Detection Dataset


A corrected version of the TXL-PBC dataset (Gan, Li & Wang, Scientific Data 12, 1694, 2025). TXL-PBC integrates four publicly available blood-cell datasets — BCCD, BCDD, PBC and Raabin-WBC — into a single YOLO-format detection benchmark for three cell types: WBC (white blood cell), RBC (red blood cell) and Platelets.

This release is identical to the published TXL-PBC, except that 4 duplicate images were removed (see Changes in this version).

中文说明见文末 中文版说明。

Dataset at a glance

Boxes per class

Directory structure

.
├── images/
│   ├── train/   # 878 png
│   ├── val/     # 252 png
│   └── test/    # 126 png
├── labels/
│   ├── train/   # 878 YOLO txt
│   ├── val/     # 252 YOLO txt
│   └── test/    # 126 YOLO txt
├── data.yaml                    # Ultralytics / YOLO dataset config
├── metadata_file.xlsx           # image filename → source dataset mapping (original 1,260 sample)
├── BCCD_selection.xlsx          # selected / excluded BCCD images
├── manual_annotation_protocol.pdf  # annotation guidelines
└── LICENSE

Label format

Standard YOLO, one line per object, coordinates normalized to [0, 1]:

class_id  x_center  y_center  width  height

class_id → 0 = WBC, 1 = RBC, 2 = Platelets (see data.yaml).

Data sources

TXL-PBC is built by integrating and re-annotating four public datasets. Image counts below refer to this release:

During integration the images were randomly shuffled and renamed (to opaque hashes) so that no source-specific naming leaks into the dataset. The original→new filename mapping is provided in metadata_file.xlsx.

Annotation

• —Initial manual annotation — 250 images were manually annotated to bootstrap the model (100 from PBC, 100 from Raabin-WBC, 30 from BCCD, 20 from BCDD).
• —Semi-automatic annotation — a YOLOv8n model trained on those 250 images was used to pre-label the remaining images.
• —Manual review — every image was then reviewed by a human annotator; a second annotator independently cross-checked the labels (cell type, box placement, missed cells).
• —Tool — X-AnyLabeling.
• —Only cells that could be confidently identified were annotated; ambiguous, severely overlapping or heavily blurred cells were intentionally left unlabeled.

Full guidelines: manual_annotation_protocol.pdf.

Usage

from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.train(data="data.yaml", epochs=100, imgsz=640)

metrics = model.val()   # evaluates on images/test/

data.yaml:

train: ./images/train/
val: ./images/val/
test: ./images/test/
nc: 3
names: ['WBC', 'RBC', 'Platelets']

Changes in this version

4 duplicate images (45 bounding boxes in total) were identified and removed from the original 1,260-image release. All four originate from the Raabin-WBC source:

Result: 1,260 → 1,256 images, 18,143 → 18,098 boxes. metadata_file.xlsx still documents the original 1,260-image sample (including these four rows).

Citation

If you use this dataset, please cite the TXL-PBC paper:

Gan, L., Li, X. & Wang, X. A Curated and Re-annotated Peripheral Blood Cell Dataset Integrating Four Public Resources. Sci Data 12, 1694 (2025). <https://doi.org/10.1038/s41597-025-05980-z>

@article{gan2025curated,
  title={A Curated and Re-annotated Peripheral Blood Cell Dataset Integrating Four Public Resources},
  author={Gan, Lu and Li, Xi and Wang, Xichun},
  journal={Scientific Data},
  volume={12},
  number={1},
  pages={1694},
  year={2025},
  publisher={Nature Publishing Group UK London}
}

If you use this corrected release, please cite it as well. It is not identical to the original 1,260-image sample: four duplicate images were removed and the labels and documentation were corrected, so a citation to the paper alone does not describe this version.

@misc{txlpbc_corrected,
  author    = {KeenForgeAI},
  title     = {TXL-PBC-corrected: a corrected release of the TXL-PBC peripheral blood cell dataset},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  doi       = {10.57967/hf/10527},
url       = {https://huggingface.co/datasets/KeenForgeAI/TXL-PBC-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from Gan et al. (2025),
               doi:10.1038/s41597-025-05980-z.}
}

The annotation tool (optional):

@software{keenforge,
  author    = {KeenForgeAI},
  title     = {KeenForge: a local-first, offline image annotation and model-training desktop tool},
  year      = {2026},
  publisher = {KeenForgeAI},
  url       = {https://github.com/KeenForgeAI/KeenForge},
  note      = {MIT licensed. Developed by Lu Gan and Sam Li.}
}

Source dataset references

@dataset{BCCD,
  author = {Shenggan and Nicolas Chen and cosmicad and akshaylamba},
  title  = {BCCD: Blood Cell Count and Detection},
  year   = {2018},
  url    = {https://github.com/Shenggan/BCCD_Dataset}
}

@article{acevedo2020dataset,
  title={A dataset of microscopic peripheral blood cell images for development of automatic recognition systems},
  author={Acevedo, Andrea and Merino, Anna and Alf{\'e}rez, Santiago and Molina, {\'A}ngel and Bold{\'u}, Laura and Rodellar, Jos{\'e}},
  journal={Data in Brief},
  volume={30},
  pages={105474},
  year={2020},
  doi={10.1016/j.dib.2020.105474}
}

@article{kouzehkanan2022large,
  title={A large dataset of white blood cells containing cell locations and types, along with segmented nuclei and cytoplasm},
  author={Mousavi Kouzehkanan, Zahra and Saghari, Sepehr and Tavakoli, Sajad and Rostami, Peyman and Abaszadeh, Mohammadjavad and Mirzadeh, Farzaneh and others},
  journal={Scientific Reports},
  volume={12},
  number={1},
  pages={1123},
  year={2022},
  doi={10.1038/s41598-021-04426-x}
}

@misc{BCDD,
  title  = {Blood Cell Detection Dataset (BCDD)},
  author = {draaslan},
  url    = {https://github.com/draaslan/blood-cell-detection-dataset}
}

License

Released under the MIT License (see LICENSE), following the original TXL-PBC release. Please also credit the four source datasets listed above.

Links

• —Paper: <https://doi.org/10.1038/s41597-025-05980-z>
• —Original dataset on GitHub: <https://github.com/lugan113/TXL-PBC_Dataset>
• —Original dataset on Figshare: <https://doi.org/10.6084/m9.figshare.27073186.v8>

中文版说明

这是 TXL-PBC 外周血细胞检测数据集的修正版（原论文：Gan, Li & Wang, Scientific Data 12, 1694, 2025）。 TXL-PBC 由四个公开血液细胞数据集整合再标注而成：BCCD、BCDD、PBC、Raabin-WBC，统一为 YOLO 格式， 共 3 类：WBC（白细胞）、RBC（红细胞）、Platelets（血小板）。

本版本与已发表的 TXL-PBC 一致，唯一区别是删除了 4 张重复图片。

各类别框数：WBC 1,294 ｜ RBC 16,261 ｜ Platelets 543

本版改动：删除 4 张重复图片（共 45 个框），均来自 Raabin-WBC 源 —— 20190527_114537_1.jpg、20190531_105858_0.jpg、95-5-5-1_108_1.jpg、95-8-10-1_980_1.jpg。 图片数 1,260 → 1,256，框数 18,143 → 18,098。

标注流程：先人工标注 250 张训练 YOLOv8n → 半自动预标注其余图片 → 人工逐张复核 → 第二位标注员交叉复核。工具：X-AnyLabeling。 详细规范见 manual_annotation_protocol.pdf。

数据来源：PBC（500 张）、Raabin-WBC（496 张）、BCCD（160 张）、BCDD（100 张）； 整合时随机打乱并重命名为哈希文件名，原始名与来源对照见 metadata_file.xlsx。

许可：MIT（沿用原 TXL-PBC 发布许可）；请同时注明上述四个源数据集。

引用：Gan, L., Li, X. & Wang, X. Sci Data 12, 1694 (2025). <https://doi.org/10.1038/s41597-025-05980-z>