I understand that Derm1M-AgentAug is released for non-commercial research purposes only, under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) license, and that commercial use requires separate permission from the dataset creators. I acknowledge that the images and the original captions come from the Derm1M dataset (https://github.com/SiyuanYan1/Derm1M) and that I will comply with its terms; that part of the captions are agent-generated and were verified by retrieval rather than by clinicians; that the dataset is not a medical device; and that it must not be used for diagnosis, triage, or any other clinical decision-making. I further agree to use this dataset responsibly and ethically for advancing dermatological research and medical AI development.

Derm1M-AgentAug

Knowledge-enriched captions for 413,369 dermatological images, generated by MAGEN (Multi-Agent data GENeration) and used to pretrain O-MAKE.

MAGEN rewrites part of the corpus through a foundation-model-assisted captioning agent with a diagnostic tool, verifying each result by retrieval; captions it did not improve on keep the original Derm1M text, and the agent_generated column records which is which. Every caption is additionally decomposed into distinct knowledge aspects (subcaptions), an ontology caption naming the diagnosis, and a visual-concept caption listing the observed findings.

Source data. The images and the original captions come from Derm1M. This dataset is a derivative that adds MAGEN-generated captions and knowledge-aspect decompositions on top of it. Please cite Derm1M alongside this work and observe its terms of use.

📄 Paper: IEEE TMI (arXiv:2512.03445)
💻 Code: github.com/XiejiLi/MAGEN-O-MAKE
🗂️ Source dataset: github.com/SiyuanYan1/Derm1M

Splits

Split	Pairs	Agent-generated captions	Images
`train`	403,563	186,069 (46.1%)	46.3 GB
`validation`	9,806	0	1.1 GB

Path	Contents
`csv/MAGEN_train.csv`, `csv/MAGEN_valid.csv`	captions and metadata, no images
`data/train-.parquet`, `data/validation-.parquet`	the same rows with the images embedded

Content note. These are clinical dermatology photographs, including advanced disease, wounds, and other graphic presentations. The dataset viewer on this page is therefore pointed at the caption CSVs only — no image is ever rendered in the browser. The image shards are loaded explicitly, as shown below.

Usage

Captions and metadata only — this is what the preview on this page shows:

from datasets import load_dataset

ds = load_dataset('Xieji-Li/Derm1M-AgentAug', split='train')
print(ds[0]['filename'], ds[0]['truncated_caption'])

With the images:

ds = load_dataset(
    'Xieji-Li/Derm1M-AgentAug',
    data_files={'train': 'data/train-*.parquet',
                'validation': 'data/validation-*.parquet'},
    split='train',
)
ds[0]['image']        # a PIL image

script/pretrain.sh in the code repository reads images from disk, so materialise them once using each row's filename:

import os

for row in ds:
    path = row['filename']            # data/pretrain/images/<source>/<file>
    os.makedirs(os.path.dirname(path), exist_ok=True)
    row['image'].save(path)

Schema

Column	Description
`image`	the image itself
`filename`	`data/pretrain/images/<source>/<file>`, relative to the code repository root
`truncated_caption`	MAGEN caption, truncated to the text encoder's context length
`ontology_caption`	"This is a skin photo diagnosed as `<ontology path>`."
`visual_concept_caption`	"This skin photo shows `<concepts>`."
`subcaption_1` … `subcaption_8`	the caption split into knowledge aspects (blank where unused)
`sub_caption_mask`	8-element 0/1 mask marking which subcaptions are present
`knowledge_masks`	3-element 0/1 mask over (caption, ontology caption, visual-concept caption)
`ontology_label`	index into the Derm1M disease hierarchy, `-1` when unmapped
`agent_generated`	`True` if MAGEN rewrote this caption, `False` if it is the original Derm1M text
`source`	corpus of origin
`source_type`	coarse origin category

Composition by source (train)

Source	Images
youtube	184,344
IIYI_chinese	53,747
pubmed_english	46,597
public	35,947
pubmed_fail	30,955
textbook_english	24,481
textbook_fail	12,506
twitter_english	6,116

Limitations

Part of this dataset is agent-generated. MAGEN rewrote 186,069 of the 403,563 training captions (46.1%); the remaining 217,494, and all 9,806 validation captions, are the original Derm1M text. The agent_generated column marks which is which.

Citation

@article{li2025multi,
  title={Multi-Aspect Knowledge-Enhanced Medical Vision-Language Pretraining with Multi-Agent Data Generation},
  author={Li, Xieji and Yan, Siyuan and Liu, Yingsheng and Soyer, H Peter and Janda, Monika and Mar, Victoria and Ge, Zongyuan},
  journal={arXiv preprint arXiv:2512.03445},
  year={2025}
}

The MICCAI'25 conference version this extends:

@misc{yan2025makemultiaspectknowledgeenhancedvisionlanguage,
      title={MAKE: Multi-Aspect Knowledge-Enhanced Vision-Language Pretraining for Zero-shot Dermatological Assessment}, 
      author={Siyuan Yan and Xieji Li and Ming Hu and Yiwen Jiang and Zhen Yu and Zongyuan Ge},
      year={2025},
      eprint={2505.09372},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.09372}, 
}

Downloads last month: 11

Total file size:

47.5 GB

Papers for Xieji-Li/Derm1M-AgentAug

Multi-Aspect Knowledge-Enhanced Medical Vision-Language Pretraining with Multi-Agent Data Generation

Paper • 2512.03445 • Published Dec 3, 2025

MAKE: Multi-Aspect Knowledge-Enhanced Vision-Language Pretraining for Zero-shot Dermatological Assessment

Paper • 2505.09372 • Published May 14, 2025 • 1