OpenKBP-Opt：对76个知识型计划流程的国际化可复现评估
医学物理
2022-09-28 v1 人工智能
计算机视觉与模式识别
摘要
我们建立了一个开放框架，用于开发放疗中知识型计划（KBP）的优化模型。我们的框架包含100例头颈癌患者的参考计划，以及 OpenKBP 大赛中不同研究组开发的19个 KBP 模型所给出的高质量剂量预测。这些剂量预测被输入4个优化模型，形成76条独特的 KBP 流程，共生成7600个计划。通过以下指标将预测与计划同参考计划比较：剂量评分（即模型在剂量上实现的逐体素平均绝对差异）；剂量体积直方图（DVH）标准的偏差；以及临床计划标准满足频率。我们还进行了理论探究以论证我们的剂量模仿模型。预测与其 KBP 流程间剂量评分的秩序相关范围为0.50至0.62，表明预测质量通常与计划质量正相关。此外，与输入预测相比，KBP 生成的计划在23项 DVH 标准中的18项上表现显著更优（P<0.05；单侧 Wilcoxon 检验）。类似地，每个优化模型生成的计划满足标准的百分比均高于参考计划。最后，我们的理论研究表明，剂量模仿模型生成的计划对于常规计划模型也是最优的。这是迄今评估 KBP 预测与优化模型组合的最大国际性工作。为可复现起见，我们的数据与代码可在 https://github.com/ababier/open-kbp-opt 免费获取。
引用
@article{arxiv.2202.08303,
title = {OpenKBP-Opt: An international and reproducible evaluation of 76 knowledge-based planning pipelines},
author = {Aaron Babier and Rafid Mahmood and Binghao Zhang and Victor G. L. Alves and Ana Maria Barragán-Montero and Joel Beaudry and Carlos E. Cardenas and Yankui Chang and Zijie Chen and Jaehee Chun and Kelly Diaz and Harold David Eraso and Erik Faustmann and Sibaji Gaj and Skylar Gay and Mary Gronberg and Bingqi Guo and Junjun He and Gerd Heilemann and Sanchit Hira and Yuliang Huang and Fuxin Ji and Dashan Jiang and Jean Carlo Jimenez Giraldo and Hoyeon Lee and Jun Lian and Shuolin Liu and Keng-Chi Liu and José Marrugo and Kentaro Miki and Kunio Nakamura and Tucker Netherton and Dan Nguyen and Hamidreza Nourzadeh and Alexander F. I. Osman and Zhao Peng and José Darío Quinto Muñoz and Christian Ramsl and Dong Joo Rhee and Juan David Rodriguez and Hongming Shan and Jeffrey V. Siebers and Mumtaz H. Soomro and Kay Sun and Andrés Usuga Hoyos and Carlos Valderrama and Rob Verbeek and Enpei Wang and Siri Willems and Qi Wu and Xuanang Xu and Sen Yang and Lulin Yuan and Simeng Zhu and Lukas Zimmermann and Kevin L. Moore and Thomas G. Purdie and Andrea L. McNiven and Timothy C. Y. Chan},
journal= {arXiv preprint arXiv:2202.08303},
year = {2022}
}
备注
19 pages, 7 tables, 6 figures