基于自动分割的急诊CT影像组学预测原发性脑干出血出院功能结局及院内卒中后肺炎
本研究回顾性纳入372例原发性脑干出血患者,采用三维U-Net自动分割血肿,构建影像组学模型预测出院不良功能结局和院内卒中后肺炎。自动分割影像组学模型与手动分割模型性能相当,联合预后模型在训练集、内部测试集和时间内部验证集的AUC分别为0.926、0.923、0.858,优于常规影像模型;肺炎联合模型AUC分别为0.828、0.742、0.681,改善不一致。SHAP显示血肿分类是最有影响力的预后预测因子。
要点
- 自动分割影像组学模型性能与手动分割相当,可用于急诊NCCT早期风险分层。
- 联合预后模型对出院不良功能结局判别良好,AUC最高达0.926。
- 血肿分类是影响预后的最重要预测因子,肺炎模型仍为探索性。
核心结论
- 基于自动分割的急诊NCCT影像组学对原发性脑干出血早期风险分层可行,联合预后模型判别良好且可能补充常规CT评估。
- 肺炎联合模型性能改善不一致,仍需进一步优化和验证。
研究设计与样本
- 回顾性研究,纳入372例原发性脑干出血患者,分为训练集182例、内部测试集79例和时间内部验证集111例。
- 使用三维U-Net自动分割血肿,比较手动与自动分割影像组学模型,并将自动分割特征与常规影像预测因子整合构建联合模型。
- 参考标准为出院不良功能结局和院内卒中后肺炎。
诊断标准与关键数值
- 自动分割影像组学模型在三个数据集上与手动分割模型性能相当。
- 联合预后模型在训练集、内部测试集和时间内部验证集的AUC分别为0.926、0.923、0.858,AUC高于常规影像模型且Brier评分更低。
- 联合肺炎模型在三个数据集的AUC分别为0.828、0.742、0.681,未见一致改善。
- SHAP分析显示血肿分类是最有影响力的预后预测因子。
临床应用与阅片要点
- 急诊NCCT自动分割影像组学可辅助早期识别原发性脑干出血不良功能结局高风险患者。
- 血肿分类对预后评估有重要价值,阅片时应关注血肿异质性特征。
- 肺炎预测模型尚不成熟,不建议单独用于临床决策。
局限性
- 回顾性研究设计。
- 肺炎联合模型性能改善不一致,仅为探索性,需进一步优化和验证。
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: Primary brainstem hemorrhage (PBSH) is associated with high mortality, disability, and uncertain early outcomes. Emergency non-contrast CT (NCCT) is the earliest imaging examination, but conventional assessment may not fully capture hematoma heterogeneity. This study aimed to develop an automated segmentation-based emergency NCCT radiomics framework for predicting poor functional outcome at discharge and in-hospital post-stroke pneumonia (PSP) in PBSH.
METHODS: This retrospective study included 372 patients with PBSH, comprising a training set (n = 182), an internal test set (n = 79), and a temporal internal validation set (n = 111). Hematomas were segmented using a three-dimensional U-Net. Manual and automated segmentation-based radiomics models were compared, and automated segmentation-derived radiomics features were integrated with conventional radiological predictors to construct combined models for poor functional outcome and PSP.
RESULTS: The automated segmentation-based radiomics model showed comparable performance to the manual segmentation-based model across all three datasets. The combined prognostic model achieved AUCs of 0.926, 0.923, and 0.858 in the training, internal test, and temporal internal validation sets, with higher AUCs and lower Brier scores than the conventional radiological model. The combined PSP model (AUCs: 0.828, 0.742, 0.681) showed no consistent improvement. SHAP identified hematoma classification as the most influential prognostic predictor.
CONCLUSIONS: Automated segmentation-based NCCT radiomics is feasible for early risk stratification in PBSH. The combined prognostic model showed good discrimination and potential complementary value beyond conventional CT assessment. The PSP model remains exploratory and requires further refinement and validation.
引用
- Li Y, Wang J, Liang Y, et al. Automated segmentation-based emergency CT radiomics for predicting functional outcome at discharge and in-hospital post-stroke pneumonia in primary brainstem hemorrhage. Eur J Radiol. 2026. PMID: 42880155. doi:10.1016/j.ejrad.2026.113290
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。