EgoMed-IEMIS — Interactive Egocentric Medical Image Segmentation Dataset

First-person videos of a clinician reviewing medical images on a screen through smart glasses, with frame-level segmentation masks for the referred medical targets. This is the dataset used in “Understanding From Human Perspective: A Multi-agent System for Interactive Egocentric Medical Image Segmentation” (EgoMed-Agent).

The full IEMIS dataset is 523 videos / 173,657 frames, spanning 5 imaging modalities, 12 medical targets, and 5 everyday capture scenes.

⚠️ What this repository currently hosts

This repository hosts all five imaging-modality subsets: AMOS (CT), CAMUS (ultrasound), PolypGen (endoscopy), ACDC (MRI), and Montgomery County CXR (X-ray).

Code & method: EgoMed-Agent · Paper: arXiv:2607.17341

Sources (one public dataset per modality)

Modality	Source	#Videos	In this repo
CT	AMOS	207	✅
Ultrasound	CAMUS	80	✅
Endoscopy	PolypGen (PMC)	61	✅
MRI	ACDC	70	✅
X-ray	Montgomery County CXR (MCC)	105	✅
Total		523	523 hosted

How it was built

Each source image is displayed in a DICOM viewer (RadiAnt) on a screen and filmed from the first-person view with smart glasses (Xiaomi AI Glasses, Rokid Glasses) across five everyday scenes (workstation, living room, café, bedroom, classroom), under static viewing or head motion (left-right / forward-backward). The source pixel masks are transferred to every egocentric frame by SAM2 screen tracking + a per-frame homography projection, advancing a slice at each keyframe. Every video was reviewed for annotation quality (580 collected → 523 released, 57 discarded); only quality-passed videos are released. Full protocol is in the paper's supplementary material.

Repository layout (WebDataset-style tar shards)

data/<MODALITY>/<split>/<MODALITY>-<split>-NNNN.tar   # ~2 GB shards (AMOS_CT / CAMUS_US / PolypGen_Endo)
splits/<MODALITY>_splits.txt                          # split <tab> case_id
metadata.csv                                          # modality, split, case_id, num_frames

Within each shard, every frame is stored as a paired sample sharing one key:

AMOS_CT/train/76/76_0000.jpg     # egocentric frame
AMOS_CT/train/76/76_0000.png     # matching mask

Masks are label-preserving gray-value PNGs: pixel intensity encodes the medical target class. Targets per modality — AMOS: liver, left/right kidney, spleen, stomach; CAMUS: left atrium; PolypGen: polyp.

Splits are 5 : 2 : 3 (train / val / test) per modality.

Usage

from huggingface_hub import snapshot_download
snapshot_download(repo_id="daizywang/EgoMed-IEMIS", repo_type="dataset", local_dir="iemis")

Citation

If you use the full dataset, please cite our paper. When using one or more modality subsets, also cite the corresponding original source dataset paper(s) below.

@article{ge2026understanding,
  title   = {Understanding From Human Perspective: A Multi-agent System for Interactive Egocentric Medical Image Segmentation},
  author  = {Ge, Rongjun and Wang, Dongyang and Zhu, Heng and Li, Zhirui and Chen, Yang and He, Yuting},
  journal = {arXiv preprint arXiv:2607.17341},
  year    = {2026},
  url     = {https://arxiv.org/abs/2607.17341}
}

Original source datasets and acknowledgments

We thank the creators and maintainers of AMOS, CAMUS, PolypGen, ACDC, and Montgomery County CXR for making their datasets available. EgoMed-IEMIS contains first-person recordings of medical-image review with annotations projected from these source datasets. Please cite the relevant source paper(s) when using a corresponding subset.

BibTeX for source datasets

@article{ji2022amos,
  title={AMOS: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation},
  author={Ji, Yuanfeng and Bai, Haotian and Ge, Chongjian and Yang, Jie and Zhu, Ye and Zhang, Ruimao and Li, Zhen and Zhang, Lingyan and Ma, Wanling and Wan, Xiang and others},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  pages={36722--36732},
  year={2022}
}

@article{leclerc2019deep,
  title={Deep learning for segmentation using an open large-scale dataset in 2D echocardiography},
  author={Leclerc, Sarah and Smistad, Erik and Pedrosa, Joao and {\O}stvik, Andreas and Cervenansky, Frederic and Espinosa, Florian and Espeland, Torvald and Berg, Erik Andreas Rye and Jodoin, Pierre-Marc and Grenier, Thomas and others},
  journal={IEEE Transactions on Medical Imaging},
  volume={38},
  number={9},
  pages={2198--2210},
  year={2019}
}

@article{ali2023multi,
  title={A multi-centre polyp detection and segmentation dataset for generalisability assessment},
  author={Ali, Sharib and Jha, Debesh and Ghatwary, Noha and Realdon, Stefano and Cannizzaro, Renato and Salem, Osama E and Lamarque, Dominique and Daul, Christian and Riegler, Michael A and Anonsen, Kim V and others},
  journal={Scientific Data},
  volume={10},
  number={1},
  pages={75},
  year={2023}
}

@article{bernard2018deep,
  title={Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved?},
  author={Bernard, Olivier and Lalande, Alain and Zotti, Clement and Cervenansky, Frederick and Yang, Xin and Heng, Pheng-Ann and Cetin, Irem and Lekadir, Karim and Camara, Oscar and Ballester, Miguel Angel Gonzalez and others},
  journal={IEEE Transactions on Medical Imaging},
  volume={37},
  number={11},
  pages={2514--2525},
  year={2018}
}

@article{jaeger2014two,
  title={Two public chest X-ray datasets for computer-aided screening of pulmonary diseases},
  author={Jaeger, Stefan and Candemir, Sema and Antani, Sameer and W{\'a}ng, Y{\`i}-Xi{\'a}ng J and Lu, Pu-Xuan and Thoma, George},
  journal={Quantitative Imaging in Medicine and Surgery},
  volume={4},
  number={6},
  pages={475},
  year={2014}
}