基于自动乳腺容积扫描瘤内及不同瘤周区域影像组学与机器学习准确预测乳腺癌Ki-67
本研究纳入668例乳腺癌患者(训练集486例、测试集182例),基于自动乳腺容积扫描图像提取瘤内及2、4、6、8、10 mm瘤周区域影像组学特征,构建机器学习模型预测Ki-67。瘤内联合6 mm瘤周区域的XGBoost模型预测准确性最高,训练集AUC为0.957、测试集AUC为0.920,SHAP分析提示6 mm瘤周特征比瘤内特征更重要。
要点
- 瘤内联合6 mm瘤周区域的XGBoost模型预测Ki-67准确性最高,测试集AUC为0.920。
- SHAP分析显示6 mm瘤周影像组学特征比瘤内特征更重要。
- 不同瘤周区域尺寸(2、4、6、8、10 mm)中,6 mm为预测Ki-67的最佳瘤周大小。
核心结论
- 基于自动乳腺容积扫描的瘤内联合6 mm瘤周区域影像组学特征构建的XGBoost模型,能实现最准确的术前Ki-67预测。
研究设计与样本
- 研究纳入668例乳腺癌患者,分为训练集486例和测试集182例;在训练集中针对瘤内及2、4、6、8、10 mm瘤周区域构建机器学习模型,并在测试集中验证以确定预测Ki-67的最佳瘤周区域。
诊断标准与关键数值
- 瘤内联合6 mm瘤周区域的XGBoost模型预测准确性最高,训练集AUC为0.957,测试集AUC为0.920;校准曲线证实可靠性,决策曲线分析显示最高净获益。
临床应用与阅片要点
- 作者认为,采用自动乳腺容积扫描影像组学分析时,纳入瘤内及6 mm瘤周区域特征可提高术前Ki-67预测准确性。
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
RATIONALE AND OBJECTIVES: Current radiomics research primarily focuses on intratumoral regions and fixed peritumoral areas, lacking optimization for accurate Ki-67 prediction. This study aimed to develop machine learning (ML) models to analyze radiomic features from Automated Breast Volume Scanning (ABVS) images of different peritumoral region sizes to identify the optimal size for accurate preoperative Ki-67 prediction.
MATERIALS AND METHODS: A total of 668 breast cancer patients were enrolled and divided into training (486) and testing (182) cohorts. In the training cohort, ML models were developed for intratumoral and peritumoral regions (2, 4, 6, 8, and 10 mm). Relevant Ki-67 features for each ROI were identified, and different models were compared to determine the optimal one. These models were validated using a testing cohort to find the most accurate peritumoral region for Ki-67 prediction. SHAP (Shapley Additive Explanations) analysis was performed to identify key radiomic features from the optimal model.
RESULTS: The Extreme Gradient Boosting (XGBoost) model for the intratumoral region combined with the 6 mm peritumoral region achieved the highest predictive accuracy, with an AUC of 0.957 in the training cohort and 0.920 in the testing cohort. Calibration curves confirmed reliability, and decision curve analysis demonstrated the highest net benefit. SHAP indicated that 6 mm peritumoral radiomic features are more significant than intratumoral features.
CONCLUSION: The XGBoost model using ABVS-derived radiomic features from both the intratumoral and 6 mm peritumoral regions provides the most accurate preoperative Ki-67 prediction.
引用
- Hu B, Xu Y, Gong H, et al. Radiomics Analysis of Intratumoral and Various Peritumoral Regions From Automated Breast Volume Scanning for Accurate Ki-67 Prediction in Breast Cancer Using Machine Learning. Acad Radiol. 2024. PMID: 39256084. doi:10.1016/j.acra.2024.08.040
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。