放射科医生测量脑肿瘤时的测量变异性
本研究复现一项关键研究,让16名放射科医生用目测、直径测量和人工智能辅助分割三种方法测量12个近球形脑肿瘤,分析观察者间和观察者内变异性,并用计算机模拟模型研究测量误差对疾病进展误分类的影响。结果显示,以人工智能辅助分割为参考,目测和直径测量的误差均导致一定比例的进展或缓解误判,目测的误判率最高。
要点
- 以AI辅助分割为参考,目测和直径测量的观察者内与观察者间合并误差分别为13.6-22.2%和6.8-7.2%。
- 同一肿瘤重复测量中,目测在25.5%的实例中错误判定进展(20%增大阈值),直径测量为1.1%。
- 目测在12.3%的实例中错误判定缓解(30%缩小阈值),直径测量为0%。
核心结论
- 本研究提供了基于真实数据的测量误差最低预期值,测量误差对脑病灶疗效评价标准误分类的影响在目测时最为明显,未来研究应关注不同肿瘤类型的测量误差并评估其对患者随访期间疗效分类的影响。
研究设计与样本
- 复现研究,16名放射科医生测量12个近球形脑肿瘤,采用目测、直径测量和人工智能辅助分割三种方法,并以人工智能辅助分割作为参考,复现原研究的观察者间和观察者内变异性分析,并使用计算机模拟模型研究测量误差对进展性疾病误分类的影响。
诊断标准与关键数值
- 以人工智能辅助分割为参考,目测的观察者内和观察者间合并误差为13.6-22.2%,直径测量为6.8-7.2%。
- 同一肿瘤重复测量中,目测在25.5%的实例中错误判定进展(阈值:增大20%),直径测量为1.1%。
- 目测在12.3%的实例中错误判定缓解(阈值:缩小30%),直径测量为0%。
局限性
- 未来研究应关注不同肿瘤类型的测量误差并评估其对患者随访期间疗效分类的影响。
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
BACKGROUND: In oncology trials, response evaluation criteria are pivotal in developing new treatments. This study examines the influence of measurement variability in brain lesions on response classification, considering long-standing cut-offs for progression and response were determined before the era of submillimeter resolutions of medical imaging.
METHODS: We replicate a key study using modern radiological tools. Sixteen radiologists were tasked with measuring twelve near-spherical brain tumors using visual estimation (eyeballing), diameter measurements and artificial intelligence (AI) assisted segmentations. Analyses for inter- and intraobserver variability from the original were replicated. Additionally, we researched the effect of measurement error on the misclassification of progressive disease using a computer simulation model.
RESULTS: The combined effect of intra- and interobserver error varied between 13.6 and 22.2% for eyeballing and 6.8-7.2% for diameter measurement, using AI-assisted segmentation as reference. We observed erroneously declared progression (cut-off at 20% increase) in repeat measurements of the same tumor in 25.5% of instances for eyeballing and in 1.1% for diameter measurements. Response (cut-off at 30% decrease) was erroneously declared in 12.3% for eyeballing and in 0% for diameter measurements. The simulation model demonstrated a more pronounced impact of measurement error on cases with fewer total number of lesions.
CONCLUSIONS: This study provides a minimum expected measurement error using real-world data. The impact of measurement error on response evaluation criteria misclassification in brain lesions was most pronounced for eyeballing. Future research should focus on measurement error for different tumor types and assess its impact on response classification during patient follow-up.
引用
- van der Loo I, Bucho TMT, Hanley JA, et al. Measurement variability of radiologists when measuring brain tumors. Eur J Radiol. 2024. PMID: 39657547. doi:10.1016/j.ejrad.2024.111874
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。