AI多尺度预测透明细胞肾细胞癌微坏死与进展:多中心队列研究
本研究基于3个中心468例透明细胞肾细胞癌患者,提取瘤内及瘤周影像组学特征并训练2D、2.5D、3D深度学习模型,通过堆叠集成构建多模态预测系统MPS。MPS在测试集AUC为0.774,优于单一模型。SHAP分析显示瘤周1mm影像组学特征贡献最大,低风险组患者无进展生存期更佳。
要点
- MPS整合影像组学与深度学习,术前无创预测ccRCC肿瘤微坏死,测试集AUC达0.774。
- 瘤周1mm区域影像组学特征对微坏死预测贡献最大,提示勾画ROI时应包含瘤周区域。
- 低风险组患者预后更佳,可用于术前个体化治疗决策与手术管理参考。
核心结论
- MPS可较为稳健地术前无创预测ccRCC肿瘤微坏死及进展,支持个体化治疗决策。
- 多模态集成模型MPS在测试集AUC为0.774,优于单一影像组学或深度学习模型。
研究设计与样本
- 多中心队列研究,共纳入3家独立机构468例ccRCC患者,训练集198例、验证集85例、外部测试集185例。
- 从瘤内及瘤周区域提取影像组学特征,构建多种机器学习分类器进行评价。
- 基于ResNet34训练2D、2.5D、3D深度学习模型,使用相同感兴趣区ROI。
- 采用堆叠集成学习整合最优影像组学模型与最优深度学习模型,构建多模态预测系统MPS。
- 使用SHAP进行模型可解释性分析,使用Kaplan-Meier分析评估无进展生存期PFS。
关键影像表现/征象
- SHAP分析显示Rad-Peri-1mm模型的影像组学特征对肿瘤微坏死预测贡献最为显著。
- 瘤周区域特征在预测中具有重要价值,2.5D-Peri-5mm深度学习模型在测试集AUC为0.763。
诊断标准与关键数值
- Rad-Peri-1mm影像组学模型在测试集AUC为0.765,为影像组学中最佳表现。
- 2.5D-Peri-5mm深度学习模型在测试集AUC为0.763,展现较好的泛化能力。
- 基于支持向量机SVM堆叠策略的MPS在测试集AUC为0.774。
- Kaplan-Meier分析显示低风险组患者预后更好。
临床应用与阅片要点
- 日常阅片提示ccRCC时,可关注瘤周区域影像特征,瘤周特征对微坏死预测价值较高。
- MPS可作为术前评估工具辅助ccRCC手术管理决策,预测肿瘤微坏死及进展风险。
局限性
- 本文未报告研究的具体局限性内容。
- 研究设计为回顾性多中心队列,外部验证集来自不同机构。
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: Tumor micronecrosis (TM) represents a critical pathological feature influencing the prognosis and surgical management of clear cell renal cell carcinoma (ccRCC). This multicenter study aimed to develop and validate a Multimodal Predictive System (MPS) integrating radiomics and deep learning (DL) for accurate noninvasive assessment of TM and progression in ccRCC.
MATERIALS AND METHODS: A total of 468 ccRCC patients from three independent institutions were divided into a training set (n = 198), validation set (n = 85), and an external test set (n = 185). Radiomics features were extracted from intratumoral and peritumoral regions. Multiple machine learning (ML) classifiers were constructed and evaluated. In parallel, 2D, 2.5D, and 3D DL models based on ResNet34 were trained using the same regions of interest (ROIs). The optimal models from each modality were integrated via stacking ensemble learning to construct the MPS. The model interpretability was analyzed using SHapley Additive exPlanations (SHAP). Kaplan-Meier analysis was used to evaluate the Progression-free survival (PFS) of patients.
RESULTS: The Rad-Peri-1mm model achieved the best performance (AUC=0.765 in test set). Among DL models, the 2.5D-Peri-5mm model showed superior generalizability (AUC=0.763 in test set). The MPS integrating the optimal radiomics and optimal DL models via an support vector machine (SVM) stacking strategy outperformed single-modality models, achieving AUC of 0.774 in test set. SHAP analysis revealed that features derived from the Rad-Peri-1 mm model contributed most significantly to TM prediction. The Kaplan-Meier analysis revealed a better prognosis for patients in the low-risk group.
CONCLUSION: The MPS provides a robust, noninvasive tool for preoperative prediction of TM and progression in ccRCC, supporting individualized treatment decision-making.
引用
- Lv, Li, Wang, et al. Multiscale Prediction of Tumor Micronecrosis and Progression in Clear Cell Renal Cell Carcinoma Based on Artificial Intelligence: A Multicenter Cohort Study. Acad Radiol. 2026. PMID: 42265018. doi:10.1016/j.acra.2026.04.051
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。