MRI影像组学:一种用于子宫内膜癌患者风险分层的机器学习方法
本研究回顾性纳入两家机构的133例子宫内膜癌患者,基于术前T2加权MRI全肿瘤三维标注提取影像组学特征,经筛选后构建支持向量机模型。模型在训练集和外部测试集的准确率分别为0.71和0.72,结论认为全病灶T2w影像组学在识别低危子宫内膜癌患者方面结果令人鼓舞且泛化性良好。
要点
- 全肿瘤T2w影像组学联合机器学习模型在训练集和外部测试集的准确率分别为0.71和0.72。
- 经稳定性、方差和相关性筛选后,LASSO与递归特征消除最终保留4个特征构建预测模型。
- 研究结论认为全病灶T2w影像组学在识别低危子宫内膜癌患者方面结果令人鼓舞且泛化性良好。
核心结论
- 全病灶T2w影像组学在识别低危子宫内膜癌患者方面结果令人鼓舞,且表现出良好的泛化性。
研究设计与样本
- 回顾性纳入两家机构共133例子宫内膜癌患者(机构1为104例、机构2为29例),依据分期和分级分为低危组和高危组;对T2加权图像进行全肿瘤三维标注,使用基于PyRadiomics且既往验证过的流程提取特征并筛选,采用支持向量机算法在机构1数据上经2折交叉验证评估,再以机构1全部数据训练并在机构2外部测试集测试。
关键影像表现/征象
- T2加权图像经三维标注获得整个肿瘤的感兴趣体积,并据此提取影像组学特征。
诊断标准与关键数值
- 共提取1197个影像组学特征,排除不稳定、低方差和互相关特征后,经LASSO与递归特征消除识别出4个特征用于构建预测机器学习模型。
- 模型在训练集和测试集的准确率分别为0.71和0.72。
临床应用与阅片要点
- 作者认为全病灶T2w影像组学在识别低危子宫内膜癌患者方面结果令人鼓舞且泛化性良好。
局限性
- 原文未报告局限性。
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
PURPOSE: To investigate radiomics and machine learning (ML) as possible tools to enhance MRI-based risk stratification in patients with endometrial cancer (EC).
METHOD: From two institutions, 133 patients (Institution1 = 104 and Institution2 = 29) with EC and pre-operative MRI were retrospectively enrolled and divided in two a low-risk and a high-risk group according to EC stage and grade. T2-weighted (T2w) images were three-dimensionally annotated to obtain volumes of interest of the entire tumor. A PyRadiomics based and previously validated pipeline was used to extract radiomics features and perform feature selection. In particular, feature stability, variance and pairwise correlation were analyzed. Then, the least absolute shrinkage and selection operator technique and recursive feature elimination were used to obtain the final feature set. The performance of a Support Vector Machine (SVM) algorithm was assessed on the dataset from Institution 1 via 2-fold cross-validation. Then, the model was trained on the entire Institution 1 dataset and tested on the external test set from Institution 2.
RESULTS: In total, 1197 radiomics features were extracted. After the exclusion of unstable, low variance and intercorrelated features least absolute shrinkage and selection operator and recursive feature elimination identified 4 features that were used to build the predictive ML model. It obtained an accuracy of 0.71 and 0.72 in the train and test sets respectively.
CONCLUSIONS: Whole-lesion T2w-derived radiomics showed encouraging results and good generalizability for the identification of low-risk EC patients.
引用
- Mainenti PP, Stanzione A, Cuocolo R, et al. MRI radiomics: A machine learning approach for the risk stratification of endometrial cancer patients. Eur J Radiol. 2022. PMID: 35231806. doi:10.1016/j.ejrad.2022.110226
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。