Surgical Visual Understanding (SurgVU) Dataset
收藏资源简介:
机器人辅助手术(RAS)日益普及,每年有超过一百万例手术,产生了大量适合机器学习应用的数据。Surgical Visual Understanding (SurgVU) 数据集提供了大量手术视频和标签的集合,这些视频和标签是在Intuitive Surgical, Inc.举办的培训课程中捕获的。外科医生在猪模型上执行标准任务,产生了超过840小时的视频,每秒60帧,产生了大约1800万张标记图像。该数据集支持年度机器学习挑战,如MICCAI会议上的内窥镜视觉(EndoVis)挑战,关注工具检测和步骤识别等问题。除了这些挑战,该数据集对更广泛的应用也有价值,包括视频分割和生成算法。SurgVU数据集降低了新机器学习爱好者的入门门槛,作为算法开发的基准,并加速了跨学科的研究。作为最大的公开手术视频数据集,它为手术数据科学领域提供了重要资源。
Robot-Assisted Surgery (RAS) has gained increasing popularity, with over one million procedures performed annually, generating a vast amount of data suitable for machine learning applications. The Surgical Visual Understanding (SurgVU) dataset provides a large collection of surgical videos and labels captured during training courses hosted by Intuitive Surgical, Inc. Surgeons performed standard tasks on porcine models, producing over 840 hours of video at 60 frames per second, resulting in approximately 18 million labeled images. This dataset supports annual machine learning challenges, such as the Endoscopic Vision (EndoVis) Challenge at the MICCAI conference, which focuses on topics including instrument detection and step recognition. Beyond these dedicated challenges, the dataset also offers value for a wider range of applications, including video segmentation and generation algorithms. The SurgVU dataset lowers the entry barrier for novice machine learning practitioners, serves as a benchmark for algorithm development, and accelerates interdisciplinary research. As the largest publicly available surgical video dataset, it provides a crucial resource for the field of surgical data science.
Surgical Visual Understanding (SurgVU) 数据集概述
数据集简介
- 目的:为外科数据科学领域的研究提供大规模手术视频及对应标签资源。
- 背景:机器人辅助手术(RAS)年超百万例,产生大量可用于机器学习的数据。
数据集内容
- 视频来源:Intuitive Surgical, Inc. 培训课程中外科医生在猪模型上执行标准任务。
- 数据规模:
- 视频时长:840小时(60帧/秒)
- 图像数量:约1800万帧(720p分辨率)
- 视频片段:280个(来自155个会话)
组成部分
-
工具标注:
- 工具类型:12种(如持针器、镊子等)
- 标注文件:
tools.csv - 特点:可能存在因可见性问题导致的标注噪声
-
任务标注:
- 任务类型:8种(如缝合、动脉操作等)
- 标注文件:
tasks.csv - 标注来源:临床专家标注
数据访问
- 手术视频:https://storage.googleapis.com/isi-surgvu/surgvu24_videos_only.zip
- 标签数据:https://storage.googleapis.com/isi-surgvu/surgvu24_labels_updated_v2.zip
- 工具检测验证集:https://storage.googleapis.com/isi-surgvu/cat1_test_set_public.zip
应用场景
- 年度机器学习挑战(如MICCAI的EndoVis挑战)
- 工具检测、步骤识别、视频分割、生成算法等研究
引用要求
bibtex @misc{zia2025surgicalvisualunderstandingsurgvu, title={Surgical Visual Understanding (SurgVU) Dataset}, author={Aneeq Zia and Max Berniker and Rogerio Nespolo and Conor Perreault and Ziheng Wang and Benjamin Mueller and Ryan Schmidt and Kiran Bhattacharyya and Xi Liu and Anthony Jarc}, year={2025}, eprint={2501.09209}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2501.09209}, }
许可协议
- 类型:CC BY-NC-SA 4.0国际许可
- 链接:http://creativecommons.org/licenses/by-nc-sa/4.0/




