MIS-VQA contains text only. The CholecT50 and EndoVis 2018 images are distributed by their authors under their own terms, and MIS-VQA may be used for non-commercial purposes only (CC BY-NC-SA 4.0).

MIS-VQA

849K question-answer pairs and scene descriptions for 32,545 frames from CholecT50 and EndoVis18.

Paper: MIS-VQA: An Annotation-Driven Surgical Vision–Language Dataset for Complex Surgical Understanding (MICAD 2026)
Code: github.com/M-Hamdy-M/MIS-VQA

MIS-VQA is generated from existing surgical annotations. For each frame, the phase, instruments, action triplets, instrument-tissue interactions and segmentation masks are written into a text prompt, and GPT-5-mini writes questions and answers from it, guided by examples written by surgeons. Every frame has five response types: a concise and a detailed description, single-fact questions (Conv-1), questions that combine several facts (Conv-2), and complex reasoning.

This repository contains text only. The images must be requested from their authors (see Images).

Configurations

	CholecT50 train	CholecT50 test	EndoVis18 train	EndoVis18 test
Frames	27,063	3,247	1,788	447
Concise	27,063	3,247	1,788	447
Detailed	27,063	3,247	1,788	447
Conv-1	254,819	30,386	18,080	4,854
Conv-2	234,573	28,117	16,251	4,149
Reasoning	157,384	19,069	12,632	3,233
Total	700,902	84,066	50,539	13,130

from datasets import load_dataset

cholec = load_dataset("M-Hamdy/MIS-VQA", "cholect50")      # train, test
endovis = load_dataset("M-Hamdy/MIS-VQA", "endovis18")     # train, test

Loading requires datasets 4.0 or later. Accept the conditions on this page and log in with hf auth login first.

Fields

Each row is one frame.

Field	Description
`image`	path of the frame in the source dataset (see Images)
`video`	CholecT50 video or EndoVis18 sequence number
`frame`	frame index in that video or sequence
`concise`	answer to "Describe the surgical scene concisely."
`detailed`	answer to "Describe the surgical scene in detail."
`conv1`	single-fact questions, as a list of `{question, answer}`
`conv2`	questions that combine several facts, as a list of `{question, answer}`
`reasoning`	complex reasoning questions, as a list of `{question, answer}`

The same data is provided as gzipped JSON under json/, in the format read by the training code on GitHub (conv1, conv2 and reasoning as lists of [question, answer]).

Images

CholecT50. Request the dataset from CAMMA. image is VIDxx/NNNNNN.png under CholecT50/videos/, and frame is the CholecT50 frame index (1 fps). Use the CholecT50 frames, not frames extracted from the Cholec80 videos.
EndoVis18. Download the training set of the EndoVis 2018 Robotic Scene Segmentation challenge. image is seq_X/left_frames/frameNNN.png.

from PIL import Image

row = cholec["test"][0]
image = Image.open(f"CholecT50/videos/{row['image']}")
print(row["conv1"][0]["question"], row["conv1"][0]["answer"])

Splits

CholecT50: test videos 5, 12, 26, 27 and 31, which are the Cholec80-VQA test videos available in CholecT50. The remaining 45 videos are used for training.
EndoVis18: test sequences 1, 5 and 16, as in EndoVis-18-VQA. The remaining sequences are used for training.

Citation

If you use MIS-VQA, please cite our paper:

@inproceedings{hamdy2026misvqa,
  title     = {MIS-VQA: An Annotation-Driven Surgical Vision--Language Dataset for Complex Surgical Understanding},
  author    = {Hamdy, Mohamed and Zouari, Iheb and Ahmed, Fatmaelzahraa and Abdel-Ghani, Muraam and Arsalan, Muhammed and Suganthan, Ponnuthurai and Al-Jalham, Khalid and Al-Ali, Abdulaziz and Balakrishnan, Shidin},
  booktitle = {Medical Imaging and Computer-Aided Diagnosis (MICAD)},
  year      = {2026}
}

Source datasets (please cite these in addition to MIS-VQA)

MIS-VQA is built on the CholecT50 frames and labels, the Cholec80 instrument labels, CholecSeg8k and FASL predictions, the EndoVis 2018 frames and scene labels, the ISINet instrument-type masks and the interaction labels of Islam et al. The test splits follow Surgical-VQA.

@article{nwoye2022rendezvous,
  title   = {Rendezvous: Attention Mechanisms for the Recognition of Surgical Action Triplets in Endoscopic Videos},
  author  = {Nwoye, Chinedu Innocent and Yu, Tong and Gonzalez, Cristians and Seeliger, Barbara and Mascagni, Pietro and Mutter, Didier and Marescaux, Jacques and Padoy, Nicolas},
  journal = {Medical Image Analysis},
  volume  = {78},
  pages   = {102433},
  year    = {2022}
}

@article{twinanda2017endonet,
  title   = {EndoNet: A Deep Architecture for Recognition Tasks on Laparoscopic Videos},
  author  = {Twinanda, Andru P. and Shehata, Sherif and Mutter, Didier and Marescaux, Jacques and de Mathelin, Michel and Padoy, Nicolas},
  journal = {IEEE Transactions on Medical Imaging},
  volume  = {36},
  number  = {1},
  pages   = {86--97},
  year    = {2017}
}

@article{hong2020cholecseg8k,
  title   = {CholecSeg8k: A Semantic Segmentation Dataset for Laparoscopic Cholecystectomy Based on Cholec80},
  author  = {Hong, W.-Y. and Kao, C.-L. and Kuo, Y.-H. and Wang, J.-R. and Chang, W.-L. and Shih, C.-S.},
  journal = {arXiv preprint arXiv:2012.12453},
  year    = {2020}
}

@inproceedings{abdelghani2025fasl,
  title     = {FASL-Seg: Anatomy and Tool Segmentation of Surgical Scenes},
  author    = {Abdel-Ghani, Muraam and Ali, Mahmoud and Ali, Mohamed and Ahmed, Fatmaelzahraa and Arsalan, Muhammad and Al-Ali, Abdulaziz and Balakrishnan, Shidin},
  booktitle = {European Conference on Artificial Intelligence (ECAI)},
  pages     = {1001--1008},
  year      = {2025}
}

@article{allan2020endovis18,
  title   = {2018 Robotic Scene Segmentation Challenge},
  author  = {Allan, Max and Kondo, Satoshi and Bodenstedt, Sebastian and Leger, Stefan and Kadkhodamohammadi, Rahim and Luengo, Imanol and Fuentes, Felix and Flouty, Evangello and Mohammed, Ahmed and Pedersen, Marius and others},
  journal = {arXiv preprint arXiv:2001.11190},
  year    = {2020}
}

@inproceedings{gonzalez2020isinet,
  title     = {ISINet: An Instance-Based Approach for Surgical Instrument Segmentation},
  author    = {Gonz{\'a}lez, Cristina and Bravo-S{\'a}nchez, Laura and Arbel{\'a}ez, Pablo},
  booktitle = {Medical Image Computing and Computer Assisted Intervention (MICCAI)},
  pages     = {595--605},
  year      = {2020}
}

@inproceedings{islam2020learning,
  title     = {Learning and Reasoning with the Graph Structure Representation in Robotic Surgery},
  author    = {Islam, Mobarakol and Seenivasan, Lalithkumar and Ming, Lim Chwee and Ren, Hongliang},
  booktitle = {Medical Image Computing and Computer Assisted Intervention (MICCAI)},
  pages     = {627--636},
  year      = {2020}
}

@inproceedings{seenivasan2022surgicalvqa,
  title     = {Surgical-VQA: Visual Question Answering in Surgical Scenes Using Transformer},
  author    = {Seenivasan, Lalithkumar and Islam, Mobarakol and Krishna, Adithya K. and Ren, Hongliang},
  booktitle = {Medical Image Computing and Computer Assisted Intervention (MICCAI)},
  pages     = {33--43},
  year      = {2022}
}

License

The dataset is released for non-commercial use only, under CC BY-NC-SA 4.0, consistent with CholecT50, Cholec80 and CholecSeg8k. The source images and labels are not redistributed and remain subject to their own terms. The text was generated with GPT-5-mini and is subject to the OpenAI Terms of Use.

Acknowledgements

This work was supported by the Qatar Research Development and Innovation Council (QRDI), grant ARG01-0522-230266.

Contact

Questions, issues and suggestions are welcome. Please open a discussion on this page or contact me.

Downloads last month: 30

Total file size:

138 MB

Papers for M-Hamdy/MIS-VQA

CholecSeg8k: A Semantic Segmentation Dataset for Laparoscopic Cholecystectomy Based on Cholec80

Paper • 2012.12453 • Published Dec 23, 2020 • 1

2018 Robotic Scene Segmentation Challenge

Paper • 2001.11190 • Published Jan 30, 2020