AI驱动体积法自动评估结直肠癌肝转移化疗反应
本研究评估AI驱动体积法对比RECIST 1.1评估结直肠癌肝转移治疗反应的临床效用。nnU-Net分割管道在112例CT上验证，Dice为0.775±0.211；157例临床试验患者中，两种方法均对总生存有显著预后价值（p<0.0001），但在仅肝转移患者（n=43）中体积法有显著分层（-30%阈值p=0.0150；-50%阈值p=0.0409），RECIST 1.1未达统计学意义（p=0.2088）。结论认为AI体积评估提供显著预后信息，可能优于RECIST 1.1。
要点
- AI体积法在仅肝转移患者中预后分层优于RECIST 1.1
- nnU-Net分割性能随病灶大小变化，小病灶Dice仅0.566
- 体积法阈值-30%和-50%均对总生存有显著分层意义
核心结论
- AI驱动的CRLM体积评估提供显著预后信息，在肝局限转移患者中可能补充或超越RECIST 1.1。
研究设计与样本
- 研究开发并验证基于nnU-Net的AI分割管道，训练集为来自LiTS、MetaRec和MetaBrest三个数据集的476例CT扫描。
- 分割性能在MetaBrest的112例独立CT扫描上评估。
- 临床验证纳入PRODIGE 9-FFCD临床试验中157例CRLM患者，使用基线和3个月随访CT，比较RECIST 1.1与AI体积法的预后价值。
- 参考标准为总生存分析。
诊断标准与关键数值
- nnU-Net模型总体Dice相似系数为0.775±0.211，大病灶为0.899±0.046，中等病灶为0.821±0.135，小病灶为0.566±0.330。
- 在整体验证队列中，RECIST 1.1和体积评估对总生存均有显著预后价值（p<0.0001）。
- 在仅肝转移患者（n=43）中，体积评估在-30%阈值（p=0.0150）和-50%阈值（p=0.0409）显示显著预后分层，而RECIST 1.1未达统计学意义（p=0.2088）。
临床应用与阅片要点
- AI体积法可自动化实现三维评估，可能改善肝局限转移患者的预后评估和临床决策。
- 小病灶分割准确性较低（Dice 0.566），阅片时需注意小转移灶的体积测量可靠性。
局限性
- 研究为回顾性分析，临床验证队列来自单一线临床试验（PRODIGE 9-FFCD）。
- 仅肝转移亚组样本量较小（n=43）。
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）。
英文原文（摘要）· 点击展开
OBJECTIVES: This study evaluates the clinical utility of an artificial intelligence (AI)-driven volumetric approach for assessing treatment response in colorectal liver metastases (CRLM) compared to conventional RECIST 1.1 measurements.
METHODS: We developed and validated an AI segmentation pipeline using the nnU-Net framework trained on 476 CT scans from three datasets (LiTS, MetaRec, and MetaBrest). Performance was evaluated on 112 held-out CT scans from MetaBrest. For clinical validation, 157 patients with CRLM from the PRODIGE 9-FFCD clinical trial with baseline and 3-month follow-up CT scans were assessed using both RECIST 1.1 and AI-volumetric methods. Overall survival analysis was performed to compare the prognostic value of both approaches.
RESULTS: The nnU-Net model achieved a Dice similarity coefficient of 0.775 ± 0.211, with performance varying by lesion size (large: 0.899 ± 0.046; medium: 0.821 ± 0.135; small: 0.566 ± 0.330). In the overall validation cohort, both RECIST 1.1 and volumetric assessment demonstrated significant prognostic value for overall survival (p < 0.0001). In patients with liver-only metastases (n = 43), volumetric assessment showed significant prognostic stratification (p = 0.0150 at -30% threshold; p = 0.0409 at -50% threshold), while RECIST 1.1 failed to achieve statistical significance (p = 0.2088).
CONCLUSIONS: AI-driven volumetric assessment of CRLM provides significant prognostic information that complements or potentially surpasses conventional RECIST 1.1 measurements, particularly in patients with liver-limited metastatic disease. Automation through deep learning makes comprehensive 3D evaluation feasible in clinical routine.
KEY POINTS: Question Can AI-powered volumetric assessment of colorectal liver metastases improve prognostic stratification for overall survival compared with conventional RECIST 1.1 criteria? Findings AI-based volumetric tumor burden showed strong prognostic value for overall survival and provided superior risk stratification compared with RECIST 1.1 in liver-only metastatic disease. Clinical relevance Automated deep learning-based volumetric assessment enables comprehensive 3D evaluation of colorectal liver metastases, overcoming historical barriers and potentially improving prognostic assessment and clinical decision-making.
引用
- Abbas M, Andrade-Miranda G, Bourbonne V, et al. AI-driven volumetric approach for automatic chemotherapy response assessment in colorectal liver metastases. Eur Radiol. 2026. PMID: 42162286. doi:10.1007/s00330-026-12610-8
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。