2019 年鲁棒医疗器械分割挑战赛
计算机视觉与模式识别
2020-05-20 v2
摘要
腹腔镜器械的术中追踪通常是计算机及机器人辅助介入的先决条件。尽管文献中已提出众多基于内窥镜视频图像检测、分割和追踪医疗器械的方法，关键局限仍有待解决：首先，鲁棒性，即最先进方法在具有挑战性图像（例如存在血液、烟雾或运动伪影）上运行时的可靠性能。其次，泛化性；针对某特定医院某特定介入训练的算法应泛化至其他介入或机构。为推动解决这些局限，我们组织了鲁棒医疗器械分割（ROBUST-MIS）挑战赛作为国际基准竞赛，特别关注算法的鲁棒性与泛化能力。在内窥镜图像处理领域，我们的挑战赛首次包含了二值分割任务，并同时处理了多实例检测与分割。该挑战赛基于一个手术数据集，包含来自三类不同手术的共 30 台手术采集的 10,040 张标注图像。针对三个任务（二值分割、多实例检测与多实例分割）的竞赛方法验证分三个阶段进行，训练与测试数据之间的域间隙逐步增大。结果证实了初始假设，即算法性能随域间隙增大而下降。尽管性能最优算法的平均检测与分割质量较高，未来研究应集中于小、交叉、运动及透明器械（部分）的检测与分割。
引用
@article{arxiv.2003.10299,
title = {Robust Medical Instrument Segmentation Challenge 2019},
author = {Tobias Ross and Annika Reinke and Peter M. Full and Martin Wagner and Hannes Kenngott and Martin Apitz and Hellena Hempe and Diana Mindroc Filimon and Patrick Scholz and Thuy Nuong Tran and Pierangela Bruno and Pablo Arbeláez and Gui-Bin Bian and Sebastian Bodenstedt and Jon Lindström Bolmgren and Laura Bravo-Sánchez and Hua-Bin Chen and Cristina González and Dong Guo and Pål Halvorsen and Pheng-Ann Heng and Enes Hosgor and Zeng-Guang Hou and Fabian Isensee and Debesh Jha and Tingting Jiang and Yueming Jin and Kadir Kirtac and Sabrina Kletz and Stefan Leger and Zhixuan Li and Klaus H. Maier-Hein and Zhen-Liang Ni and Michael A. Riegler and Klaus Schoeffmann and Ruohua Shi and Stefanie Speidel and Michael Stenzel and Isabell Twick and Gutai Wang and Jiacheng Wang and Liansheng Wang and Lu Wang and Yujie Zhang and Yan-Jie Zhou and Lei Zhu and Manuel Wiesenfarth and Annette Kopp-Schneider and Beat P. Müller-Stich and Lena Maier-Hein},
journal= {arXiv preprint arXiv:2003.10299},
year = {2020}
}
备注
A pre-print