UniMed-5M

# UniMed-5M: Unified Medical Multimodal Dataset This repository provides the dataset resources used for training and evaluating **UniMedVL**, a unified medical multimodal model for medical image generation across 8 medical imaging modalities. ## Links - **Paper**: [arXiv:2510.15710](https://arxiv.org/abs/2510.15710) - **Project Page**: [UniMedVL Web](https://uni-medical.github.io/UniMedVL_Web/) - **Code**: [GitHub](https://github.com/uni-medical/UniMedVL) - **Model**: [UniMedVL on Hugging Face](https://huggingface.co/General-Medical-AI/UniMedVL) ## Dataset Organization The dataset is organized into **parts**, each containing Parquet metadata files for different medical imaging tasks: ### Structure ``` UniMed-5M/ ├── part1/ # Generation & synthesis tasks (Parquet files) │ ├── ACKNOWLEDGMENTS.md # Dataset citations and licensing info │ ├── LICENSES/ # License files for source datasets │ └── *.parquet # Metadata files (download images from original sources) └── part2/ # Coming soon ``` ### Part 1 - Generation & Synthesis Tasks #### Included Datasets **IXI Dataset** - MRI Super-Resolution (CC BY-SA 3.0) - 218,528 samples: 4x super-resolution for T1/T2 MRI - [Download](https://brain-development.org/ixi-dataset/) - Files: `ixi_t1_sr_4x_train.parquet`, `ixi_t2_v2_sr_4x_train.parquet` **SynthRAD2023** - CT-MR Synthesis (CC BY) - 107,936 samples: Bidirectional CT↔MR for brain/pelvis - [Download](https://zenodo.org/records/7260705) - Files: `synthrad_brain_ct_to_mr_train.parquet`, `synthrad_brain_mr_to_ct_train.parquet`, `synthrad_pelvis_ct_to_mr_train.parquet`, `synthrad_pelvis_mr_to_ct_train.parquet` **BraTS 2023** - MRI Modality Translation (CC BY 4.0) - 51,528 samples: Cross-modal synthesis (T1, T2, FLAIR, T1ce) - [Download](https://www.synapse.org/Synapse:syn51156910/wiki/621282) - Files: `brats23_train_modality_trans_v2.parquet` **DRIVE** - Retinal Vessel Segmentation - 40 samples: Fundus image vessel extraction - [Download](https://drive.grand-challenge.org/) - Files: `drive_all.parquet` **BCI (HE2IHC)** - Virtual Staining (Academic Only) - 3,896 samples: H&E to IHC for breast cancer pathology - [Download](https://bci.grand-challenge.org/) - Files: `he2ihc_train.parquet` ## Data Preparation ⚠️ **Important**: Our dataset provides **Parquet only**. Original images must be downloaded from official sources. ### What You Need To Do - 📥 Download original images from links below - 📋 Accept dataset licenses and terms - 🔗 Match image paths with our Parquet references ### Dataset Sources | Dataset | Samples | License | Download | Registration | |---------|---------|---------|----------|--------------| | **IXI** | 218,528 | CC BY-SA 3.0 | [Link](https://brain-development.org/ixi-dataset/) | No | | **SynthRAD2023** | 107,936 | CC BY | [Zenodo](https://zenodo.org/records/7260705) | No | | **BraTS 2023** | 51,528 | CC BY 4.0 | [Synapse](https://www.synapse.org/Synapse:syn51156910/wiki/621282) | Required | | **DRIVE** | 40 | See terms | [Challenge](https://drive.grand-challenge.org/) | Required | | **BCI (HE2IHC)** | 3,896 | Academic only | [Challenge](https://bci.grand-challenge.org/) | Approval needed | See [`part1/ACKNOWLEDGMENTS.md`](part1/ACKNOWLEDGMENTS.md) for citations and detailed information. ## Usage ```python from datasets import load_dataset # Load Part 1 datasets dataset = load_dataset("General-Medical-AI/UniMed-5M", data_dir="part1") # Download original images from sources listed in part1/ACKNOWLEDGMENTS.md # Match image paths in the Parquet files with your downloaded images ``` ## Licensing **UniMed-5M** is distributed under Apache License 2.0. Source datasets maintain their original licenses: - See `part1/LICENSES/` for individual dataset licenses - See `part1/ACKNOWLEDGMENTS.md` for attributions nd further details. **Copyright Clarification**: If you discover any copyright issues or need clarification regarding the dataset, please contact us directly at: **[e-mail]** ## Citation If you find this project useful in your research, please consider citing: ```bibtex @article{ning2025unimedvl, title={UniMedVL: Unifying Medical Multimodal Understanding And Generation Through Observation-Knowledge-Analysis}, author={Ning, Junzhi and Li, Wei and Tang, Cheng and Lin, Jiashi and Ma, Chenglong and Zhang, Chaoyang and Liu, Jiyao and Chen, Ying and Gao, Shujian and Liu, Lihao and others}, journal={arXiv preprint arXiv:2510.15710}, year={2025} } ```