头颈放疗中AI辅助GTV勾画的前瞻性随机评估
该单中心前瞻性随机对照试验将85例头颈癌根治性放疗患者随机分为手动勾画组与AI辅助勾画组,评估AI辅助GTV勾画的编辑程度。主要终点盲法多学科靶区会议中归一化新增路径长度(APL)的中位差异在GTV-T为-0.02[-0.32;0.07],GTV-N为0.06[-0.01;0.14];会议时长、勾画时间及感知复杂度组间相似,无不良事件,未显示效率获益。
要点
- AI辅助GTV-T勾画在盲法会议中的编辑程度与手动勾画无相关差异。
- AI辅助GTV-N勾画在同行评审中APL趋向更多编辑,但其他补充指标未显示。
- 会议时长、自报勾画时间及感知复杂度组间相似,未显示效率获益。
核心结论
- GTV-T的APL中位差异与无相关效应一致,GTV-N的APL趋向更多编辑但补充指标未显示,未证明效率获益,总体不提示担忧但需进一步研究。
研究设计与样本
- 前瞻性单中心随机对照评估,2023年9月11日至2025年1月20日入组85例,83例纳入分析(对照组41例,AI辅助组42例)。
- 患者为接受根治性放疗的头颈癌患者,随机分配至手动勾画(对照)或使用内部开发分割模型的AI辅助勾画(干预)。
- 主要终点为盲法多学科靶区会议中归一化新增路径长度(APL)的编辑程度。
主要发现
- APL中位差异(AI辅助减对照,负值利于干预)GTV-T为-0.02[-0.32;0.07],GTV-N为0.06[-0.01;0.14]。
- 会议时长、自报勾画时间及感知复杂度组间相似。
- 未报告不良事件。
关键数值与推荐
- GTV-T的APL中位差异-0.02[-0.32;0.07]与无相关效应一致。
- GTV-N的APL中位差异0.06[-0.01;0.14]趋向更多编辑,但补充轮廓指标未显示。
临床怎么用
- AI辅助GTV勾画未显示效率获益,广泛采用前需进一步研究。
- GTV-N的APL趋向更多编辑,提示需关注AI辅助淋巴结勾画的同行评审编辑。
局限性
- 探索性随机对照试验,单中心设计。
- 样本量83例纳入分析。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: AI-assisted contouring has demonstrated efficiency and consistency benefits for organs-at-risk in radiotherapy. However, its impact on gross tumor volume (GTV) delineation remains unexplored. The aim of this study was to evaluate AI-assisted GTV contouring in head and neck cancer radiotherapy in an exploratory randomized controlled clinical trial. METHODS: In this prospective, single-centre, randomized controlled evaluation, patients referred for curative-intent radiotherapy of head-neck cancer were randomized to manual contouring (control) or AI-assisted contouring (intervention) of gross tumor volumes using an in-house developed segmentation model. The primary endpoint was the extent of editing during a blinded multidisciplinary target conference, using normalized Added Path Length (APL). Secondary endpoints included conference duration, self-reported contouring time, and perceived complexity of the cases. RESULTS: Between Sept 11, 2023, and Jan 20, 2025, 85 patients were enrolled; 83 were included in the analysis (41 control, 42 AI-assisted). The median difference in APL (AI-assisted minus control, negative favoring the intervention) was -0.02 [-0.32; 0.07] for GTV-T and 0.06 [-0.01; 0.14] for GTV-N. Conference duration, contouring time, and perceived complexity were similar between groups. No adverse events were reported. CONCLUSION: The median difference in APL between groups was consistent with no relevant effect for GTV-T. For GTV-N, however, AI-assisted contouring tended towards more editing during peer review for APL, but not for supplementary contour metrics. No efficiency benefits were demonstrated. Overall, these findings do not indicate a concern, but warrants further research before widespread adoption.
引用
- Rasmussen ME, Ren J, Nijkamp JA, et al. A prospective randomized evaluation of AI assisted GTV contouring in head-and-neck radiotherapy. Radiother Oncol. 2026. PMID: 42967540. doi:10.1016/j.radonc.2026.111854
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。