| --- |
| pretty_name: MedIGen Medical Illustration Dataset |
| task_categories: |
| - image-to-text |
| - text-to-image |
| tags: |
| - medical |
| - biomedical |
| - medical-illustration |
| - image-captioning |
| - anatomy |
| - scientific-illustration |
| configs: |
| - config_name: examples |
| default: true |
| data_dir: examples |
| drop_labels: true |
| --- |
| |
| <div align="center"> |
| <h1>MedIGen Medical Illustration Dataset</h1> |
| </div> |
|
|
| ## Introduction |
|
|
| This dataset accompanies **MedIGen: Reliable Medical Illustration Generation via Interleaved Introspective Reasoning**. MedIGen formulates medical illustration generation as a high-constraint visual reasoning problem: an image must be visually plausible while also satisfying biomedical semantics, anatomical structure, spatial relationships, and procedural logic. |
|
|
| The dataset supports domain-specific visual–semantic pretraining and research on medical image understanding, image captioning, retrieval, and text-to-image generation. |
|
|
| ## Data |
|
|
| | Subset | Content | Approximate size | |
| | --- | --- | ---: | |
| | `original/` | Medical illustrations with paired text files. Images and captions are packed together in ZIP shards. | 1,216,161 image–TXT pairs | |
| | `high_quality/` | High-quality image collection with paired metadata. Images are distributed as ZIP shards and captions are stored in `metadata.csv`. | 139,639 images | |
|
|
| Each original sample consists of: |
|
|
| ```text |
| image_file.jpg |
| image_file.txt |
| ``` |
|
|
| The TXT file contains the associated caption, figure description, or source alt-text. The high-quality metadata file uses the following fields: |
|
|
| ```text |
| image,prompt |
| ``` |
|
|
| ## Repository layout |
|
|
| ```text |
| Med-Illustration-Data/ |
| ├── README.md |
| ├── examples/ |
| │ ├── images/ |
| │ ├── metadata.jsonl |
| │ └── README.md |
| ├── high_quality/ |
| │ ├── metadata.csv |
| │ └── data/ |
| │ └── images_*.zip |
| └── original/ |
| └── images_*.zip |
| ``` |
|
|
| The ZIP archives are data shards, not independent subsets. `images_1.zip`, `images_2.zip`, and so on should be treated as one combined dataset. |
|
|
| ## Loading the example split |
|
|
| Each JSONL record points to an image relative to `metadata.jsonl` and provides its caption. |
|
|
| ```python |
| from datasets import load_dataset |
| |
| dataset = load_dataset("imagefolder", data_dir="examples", split="train") |
| print(dataset) |
| print(dataset[0]["image"]) |
| print(dataset[0]["caption"]) |
| ``` |
|
|
| ## Limitations and responsible use |
|
|
| This dataset is intended for research and development. Captions may contain extraction noise, incomplete figure descriptions, outdated terminology, or errors inherited from source material. The presence of an image or caption does not imply clinical validity, diagnostic suitability, or endorsement of the source. |
|
|
| Models trained on this data should not be used for diagnosis, treatment decisions, surgical guidance, or patient-facing medical advice without appropriate expert validation and safeguards. |
|
|
| ## Citation |
|
|
| If you use this dataset or the associated MedIGen resources, please cite: |
|
|
| ```bibtex |
| |
| ``` |
|
|
