ACNE04

1,457 facial acne photographs, rebuilt from the original ACNE04 release with clean image/label separation, verified 4-level Hayashi severity grades, and 18,983 lesion-level bounding box annotations.

   

---

Overview

This dataset packages the full ACNE04 corpus, 1,457 facial photographs graded for acne severity on the 4-level Hayashi scale, as a single, ready-to-load Hugging Face dataset. Every image is embedded directly in the parquet files alongside its full label set: severity grade (level/severity), the original lesion-count bucket label from the paper, and per-image lesion bounding boxes (18,983 total across the dataset). All 5 cross-validation folds used in the original paper are preserved as explicit train/test membership columns, plus a canonical split column (fold 0) for anyone who just wants a single train/test split.

The source images and labels come straight from the official release linked from the xpwu95/LDL GitHub repository (Classification.tar for images and grade labels, Detection.tar for the Pascal VOC lesion bounding boxes). Nothing was re-annotated or re-graded; the severity grades and lesion boxes are the original authors' annotations, joined row by row into a single table.

Statement of Need

ACNE04 is normally distributed as two separate tar archives hosted on Google Drive (Classification: images + five pairs of plain-text train/test label files for 5-fold CV; Detection: a parallel Pascal VOC XML tree of lesion bounding boxes), which have to be manually downloaded, matched by filename, and joined. Several ACNE04 mirrors circulating on the Hub are incomplete (as few as 394 of the 1,457 images), drop the lesion bounding box annotations entirely, or repurpose the label files for unrelated tasks (e.g. text-to-image prompt captions) with no severity grade preserved. This dataset fixes that: the complete, original 1,457-image corpus, severity grades and lesion boxes joined per image, and every one of the paper's 5 CV folds retained so results are reproducible against the published benchmark.

Intended Use

Intended for training and evaluating acne severity classification and lesion-detection models; benchmarking zero-shot or fine-tuned multimodal / VLM models on dermatological severity grading against the Hayashi ground truth; and research into scaling behavior, cost-efficiency, or bias in dermatology imaging AI. Suitable for academic research and model benchmarking. Not intended for clinical deployment or diagnostic use; this is a research dataset, not a validated clinical decision-support tool, and models trained on it should not be used to inform real patient care.

Limitations

Ground truth reflects the original annotators' Hayashi grading from the source study and was not independently re-adjudicated. lesion_count_label is the raw label from the paper's text files and combines lesion count with severity bucket in a way that isn't fully documented by the original authors; use level/severity as the primary severity label. The 5 test folds are close to but not a perfectly disjoint partition of the 1,457 images (a handful of images appear in more than one fold's test set, a property of the original release, not introduced here); the canonical split column uses fold 0 only, which is a clean single split. Images vary in resolution, lighting, and framing since they were collected as routine clinical photographs, not under standardized imaging conditions.

Dataset Structure

Splits

Split (canonical, fold 0)	Images
train	1,165
test	292

Severity distribution

Severity	Level	Count
mild	0	513
moderate	1	633
severe	2	182
very_severe	3	129

Features

Column	Type	Description
image	Image	The face photo.
file_name	string	Original filename.
level	int32	Primary label. Hayashi severity grade, 0-3.
severity	string	Human-readable grade name (mild, moderate, severe, very_severe).
lesion_count_label	int32	Original lesion-count bucket label from the paper's label files.
num_lesions_annotated	int32	Number of lesion bounding boxes annotated for this image.
lesion_boxes	struct	Per-lesion bounding boxes as parallel lists: xmin, ymin, xmax, ymax (pixel coordinates). Empty lists if no lesions were boxed.
split	string	train/test, canonical split (fold 0).
fold_0_split … fold_4_split	string	train/test membership for each of the paper's 5 CV folds.

Usage

from datasets import load_dataset

ds = load_dataset("Layered-Labs/ACNE04")
print(ds)

example = ds["train"][0]
example["image"].show()
print("severity:", example["severity"], "(level", example["level"], ")")
print("lesion boxes:", example["num_lesions_annotated"])

Example: canonical train/test split

train = ds["train"].filter(lambda r: r["split"] == "train")
test = ds["train"].filter(lambda r: r["split"] == "test")
print(len(train), len(test))

Example: full 5-fold cross-validation

for fold in range(5):
    col = f"fold_{fold}_split"
    fold_train = ds["train"].filter(lambda r: r[col] == "train")
    fold_test = ds["train"].filter(lambda r: r[col] == "test")
    print(fold, len(fold_train), len(fold_test))

Citation

If you use this dataset, please cite both this release and the original ACNE04 source:

@dataset{layeredlabs2026acne04,
  title     = {ACNE04},
  author    = {Ridwan, Abdullah and Hossain, Radhyyah and Noor, Saniya},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Layered-Labs/ACNE04},
  note      = {Images and labels rebuilt from the original ACNE04 release}
}

@inproceedings{wu2019joint,
  title     = {Joint acne image grading and counting via label distribution learning},
  author    = {Wu, Xiaoping and Wen, Ni and Liang, Jie and Lai, Yu-Kun and She, Dongyu and Cheng, Ming-Ming and Yang, Jufeng},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages     = {10642--10651},
  year      = {2019}
}

License

The underlying ACNE04 dataset is released by the original authors for academic/research use only. For any other use, contact the original author, Xiaoping Wu (xpwu95@163.com), per the terms stated in the official repository. This repackaging carries forward the same restriction: do not use this dataset for commercial purposes. Maintained by Layered Labs.