CholecTriplet2021：一个用于手术动作三元组识别的基准挑战赛
计算机视觉与模式识别
2023-04-11 v2
摘要
手术室中的情境感知决策支持可通过利用来自手术工作流分析的实时反馈来促进手术的安全性与效率。大多数现有工作在粗粒度层面识别手术活动，如阶段、步骤或事件，遗漏了关于手术活动的细粒度交互细节；然而这些细节对于手术室中更有帮助的AI辅助是必需的。将手术动作识别为<器械, 动词, 目标>组合的三元组，可提供关于手术视频中发生活动的全面细节。本文提出CholecTriplet2021：一个在MICCAI 2021上组织的、用于腹腔镜视频中手术动作三元组识别的内镜视觉挑战赛。该挑战赛授予了对大规模CholecT50数据集的私有访问权限，该数据集标注了动作三元组信息。在本文中，我们介绍了挑战赛的设置以及对参与者在挑战期间提出的最先进深度学习方法的评估。共呈现了来自挑战赛组织者的4种基线方法和参赛团队的19种新深度学习算法，用于直接从手术视频中识别手术动作三元组，取得的平均精度均值(mAP)介于4.2%至38.1%之间。本研究还分析了所呈现方法所得结果的重要性，对它们进行了详尽的方法学比较、深入的结果分析，并提出了一种用于增强识别的新型集成方法。我们的分析表明，手术工作流分析尚未解决，并且突出了未来关于细粒度手术活动识别研究的有趣方向，这对手术中AI的发展至关重要。
引用
@article{arxiv.2204.04746,
title = {CholecTriplet2021: A benchmark challenge for surgical action triplet recognition},
author = {Chinedu Innocent Nwoye and Deepak Alapatt and Tong Yu and Armine Vardazaryan and Fangfang Xia and Zixuan Zhao and Tong Xia and Fucang Jia and Yuxuan Yang and Hao Wang and Derong Yu and Guoyan Zheng and Xiaotian Duan and Neil Getty and Ricardo Sanchez-Matilla and Maria Robu and Li Zhang and Huabin Chen and Jiacheng Wang and Liansheng Wang and Bokai Zhang and Beerend Gerats and Sista Raviteja and Rachana Sathish and Rong Tao and Satoshi Kondo and Winnie Pang and Hongliang Ren and Julian Ronald Abbing and Mohammad Hasan Sarhan and Sebastian Bodenstedt and Nithya Bhasker and Bruno Oliveira and Helena R. Torres and Li Ling and Finn Gaida and Tobias Czempiel and João L. Vilaça and Pedro Morais and Jaime Fonseca and Ruby Mae Egging and Inge Nicole Wijma and Chen Qian and Guibin Bian and Zhen Li and Velmurugan Balasubramanian and Debdoot Sheet and Imanol Luengo and Yuanbo Zhu and Shuai Ding and Jakob-Anton Aschenbrenner and Nicolas Elini van der Kar and Mengya Xu and Mobarakol Islam and Lalithkumar Seenivasan and Alexander Jenke and Danail Stoyanov and Didier Mutter and Pietro Mascagni and Barbara Seeliger and Cristians Gonzalez and Nicolas Padoy},
journal= {arXiv preprint arXiv:2204.04746},
year = {2023}
}
备注
CholecTriplet2021 challenge report. Paper accepted at Elsevier journal of Medical Image Analysis. 22 pages, 8 figures, 11 tables. Challenge website: https://cholectriplet2021.grand-challenge.org