CAMMA-public/SAGES_CVS_Challenge_2024

CVS Challenge Dataset

This document describes the directory and file structure accompanying this release of the CVS Challenge dataset. All details regarding dataset creation, annotation protocols, clinical definitions, and challenge design are described in the challenge manuscript:

https://arxiv.org/abs/2509.17100

Overall Directory Structure

cvs_challenge_dataset
│   README.md
│   LICENSE_CC_BY_NC_4.0.txt
│
└───train
│     ├── videos
│     │     video1.mp4
│     │     video2.mp4
│     │     ...
│     └── labels
│           │   metadata.csv
│           │
│           └── video1
│           │      frame.csv
│           │      video.csv
│           │
│           └── video2
│                  frame.csv
│                  video.csv
│
└───test
      ├── videos
      └── labels

Raw Videos

The videos/ folders contain 1 minute 30 second-long clips extracted from laparoscopic cholecystectomy procedures. Only selected frames within each clip are annotated with CVS criteria. Videos vary in resolution and acquisition characteristics.

Labels

Every video in this dataset has undergone a multi-rater annotation process for frame-level and video-level CVS criteria, in addition to associated metadata. Note: This repository contains the dataset and annotations used for the 2024 edition of the CVS Challenge. Labels and annotations from subsequent editions (e.g., 2025) are not included.

Metadata:

The metadata file, available at ./labels/metadata.csv, contains clinical and acquisition-related information described below.

Column Header	Value/Type	Description
video_name	string (unique)	Unique identifier linking videos with their corresponding annotations.
icg	integer (0 or 1)	0 ⇒ Indocyanine Green (ICG) was not used in this procedure. 1 ⇒ ICG was used.
ioc	integer (0 or 1)	0 ⇒ Intraoperative Cholangiography (IOC) was not used in this procedure. 1 ⇒ IOC was used.
robotic	integer (0 or 1)	0 ⇒ No robotic instruments were used. 1 ⇒ Robotic instruments were used.
country	integer (unique)	Anonymized identifier linking a video to a country.
device	integer (unique)	Anonymized identifier linking a video to a recording device.

A detailed description of all clinical metadata fields is available in the manuscript linked above.

Frame Labels:

Each 1:30 clip in ./videos/ is annotated every 5 seconds by 3 independent raters. These frame-level labels are found at:

./labels/<video_name>/frame.csv

The contents of each frame.csv file are described below:

Column Header	Value/Type	Description
frame_id	integer	Unique frame ID linking a row of annotations to a specific frame. Frame numbering is 0-indexed and based on the clip FPS (30).
c{x}_rater{y}	integer (0 or 1)	Whether rater y assessed Criterion x to be achieved. x ∈ {1,2,3}, y ∈ {1,2,3}.

Note that rater1, rater2, and rater3 do not correspond to specific individuals; different subsets of raters annotated different videos.

Video Labels:

Three independent raters assessed each clip for whether CVS criteria were achieved at the video level, as well as the difficulty of annotation. These are found in:

./labels/<video_name>/video.csv

The contents of each video.csv file are described below:

Column Header	Value/Type	Description
c{x}_rater{y}	integer (0 or 1)	Whether rater y assessed Criterion x to be achieved at the video level. x ∈ {1,2,3}, y ∈ {1,2,3}.
confidence_rater{y}	float (0–1)	0 ⇒ very difficult to annotate / not confident. 1 ⇒ very easy to annotate / very confident. y ∈ {1,2,3}.

Video-level ratings provide additional contextual information but are not explicitly evaluated in the challenge.

Code and Baselines

Reference code for loading the CVS Challenge dataset, computing evaluation metrics, and preparing challenge submissions is available here: https://github.com/SAIIL/CVS_challenge_code

The repository includes:

• datasets/ : example PyTorch dataloaders for the CVS Challenge data
• util/ : utilities for metric computation and result saving/loading
• notebooks/ : example Jupyter / Google Colab notebooks demonstrating dataset access, metric computation, and submission formatting

License

This dataset is released under the Creative Commons Attribution–NonCommercial 4.0 International License (CC BY-NC 4.0). You are free to share and adapt the material for non-commercial purposes, provided that appropriate attribution is given and no additional restrictions are applied.

A full copy of the license is included in this dataset as: LICENSE_CC_BY_NC_4.0.txt

Challenge Website

More information about the CVS Challenge, including challenge organization, rules, timelines, and details on previous and upcoming editions, is available at: https://www.cvschallenge.org/

Citation

If you use this dataset or any of its annotations in your work, please cite the following paper describing the challenge and dataset:

Alapatt D, Eckhoff J, Lyu Z, Ban Y, Mazellier J-P, Choksi S, Yang K, et al.
The SAGES Critical View of Safety Challenge: A Global Benchmark for AI-Assisted Surgical Quality Assessment.
arXiv preprint arXiv:2509.17100, 2025.
https://arxiv.org/abs/2509.17100

BibTeX

@article{alapatt2025sages,
  title={The SAGES Critical View of Safety Challenge: A Global Benchmark for AI-Assisted Surgical Quality Assessment},
  author={Alapatt, Deepak and Eckhoff, Jennifer and Lyu, Zhiliang and Ban, Yutong and Mazellier, Jean-Paul and Choksi, Sarah and Yang, Kunyi and Chiang, Po-Hsing and Zorzetti, Noemi and Cannas, Samuele and Neimark, Daniel and Bar, Omri and Yamlahi, Amine and Hennighausen, Jakob and Wang, Xiaohan and Li, Rui and Liang, Long and Wang, Yuxian and Koju, Saurabh and Bhattarai, Binod and Jaspers, Tim and Mao, Zhehua and Wijekoon, Anjana and Ma, Jun and Xu, Yinan and Weng, Zhilong and Okran, Ammar M. and Rashwan, Hatem A. and Shen, Boyang and Yang, Kaixiang and Zhang, Yutao and Wang, Hao and {CVS Challenge Consortium (2024)} and Li, Quanzheng and Filicori, Filippo and Li, Xiang and Mascagni, Pietro and Hashimoto, Daniel A. and Rosman, Guy and Meireles, Ozanan and Padoy, Nicolas},
  journal={arXiv preprint arXiv:2509.17100},
  year={2025}
}