README.md

MRGen: Diffusion-based Controllable Data Engine for MRI Segmentation towards Unannotated Modalities

This repository contains the curated MedGen-1M dataset proposed in MRGen: https://arxiv.org/abs/2412.04106/.

Some Information

Project Page $\cdot$ Paper $\cdot$ Dataset $\cdot$ Checkpoints

Dataset

Please check out MedGen-1M to download our curated dataset, including two parts: radiopaedia_data and conditional_dataset.

For the conditional dataset, we have directly provided our processed data, including the raw image, mask annotations, and text descriptions.

As described in our paper, considering the data privacy concerns of Radiopaedia, we only release the JSON files of this part here. For each case, the format is represented as ./radiopaedia/{patient_id}/{case_id}/{volume_id}/{slice_id}.jpeg, for example, ./radiopaedia/2564/1/MRI_4/1.jpeg. This format allows you to locate the corresponding original volume through the link provided in our json files. After obtaining official authorization from Radiopaedia, you may download the data corresponding to the JSON file on your own. Alternatively, you can send the authorization via email to us (haoningwu3639@gmail.com or Zhao_Ziheng@sjtu.edu.cn) to obtain the download link for the image data in our MedGen-1M.

Citation

If you use this dataset for your research or project, please cite:

@misc{wu2024mrgen,
        author  = {Wu, Haoning and Zhao, Ziheng and Zhang, Ya and Xie, Weidi and Wang, Yanfeng},
        title   = {MRGen: Diffusion-based Controllable Data Engine for MRI Segmentation towards Unannotated Modalities},
        journal = {arXiv preprint arXiv:2412.04106},
        year    = {2024},
  }

Contact

If you have any questions, please feel free to contact haoningwu3639@gmail.com or Zhao_Ziheng@sjtu.edu.cn.