Xv Yingjie, Xiao Bangxin, Wei Zongjie, Cao Youde, Jiang Qing, Li Feng, Lv Fajin, Peng Canjie, Li Xingshu, Xiao Mingzhao
Department of Urology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China (Y.X., B.X., Z.W., C.P., M.X.).
Department of Basic Medical Sciences, University of Chongqing Medical University, Chongqing, China (Y.C.).
Acad Radiol. 2025 May;32(5):2739-2750. doi: 10.1016/j.acra.2024.11.072. Epub 2025 Jan 9.
To develop and externally validate interpretable CT radiomics-based machine learning (ML) models for preoperative Ki-67 expression prediction in clear cell renal cell carcinoma (ccRCC).
506 patients were retrospectively enrolled from three independent institutes and divided into the training (n=357) and external test (n=149) sets. Ki67 expression was determined by immunohistochemistry (IHC) and categorized into low (<15%) and high (≥15%) expression groups. Radiomics features were extracted from segmented tumor regions in the corticomedullary phase (CMP) CT images using the "PyRadiomics" package. The least absolute shrinkage and selection operator (LASSO) regression was applied to select the most relevant radiomics features for Ki-67 expression, which were then used to train five ML models. Models' performances were evaluated via the receiving operator curve analysis and compared using Delong test. Calibration and decision curve analyses assessed the models' clinical utility. Kaplan-Meier analysis and Log-rank tests were conducted to determine the prognostic value of radiomics-predicted Ki-67 expression status. The optimal model was interpreted using SHapley Additive exPlanations (SHAP).
Eight radiomics feature were selected to build models using Random forest (RF), eXtreme Gradient Boosting (XGBoost), Logistic regression (LR), Support vector machine (SVM), and K-nearest neighbor (KNN). The RF model exhibited the best performance, achieving the highest area under the curve (AUC) in both the training (0.910, 95% confidence interval [CI]: 0.881-0.936) and external test (0.885, 95% CI: 0.826-0.934) sets, as confirmed by Delong test (all P values＜0.05). Calibration and decision curves further demonstrated the superior clinical utility of the RF model. Both IHC-based and RF-predicted high Ki-67 expression groups were significantly associated with a higher risk of tumor recurrence in the training and external test sets (all P values＜0.05). The prediction process of the RF model was uncovered in the globe and individualized terms using the SHAP.
The interpretable CT radiomics-based RF classifier exhibited robust predictive performance in assessing Ki-67 expression levels preoperatively, offering valuable prognostic insights and aiding clinical decision-making in ccRCC patients.
开发并外部验证基于可解释CT影像组学的机器学习（ML）模型，用于预测透明细胞肾细胞癌（ccRCC）术前Ki-67表达情况。
回顾性纳入来自三个独立机构的506例患者，分为训练集（n = 357）和外部测试集（n = 149）。通过免疫组织化学（IHC）检测Ki67表达，并分为低表达（<15%）和高表达（≥15%）组。使用“PyRadiomics”软件包从皮质髓质期（CMP）CT图像中的分割肿瘤区域提取影像组学特征。应用最小绝对收缩和选择算子（LASSO）回归选择与Ki-67表达最相关的影像组学特征，然后用于训练五个ML模型。通过接受者操作曲线分析评估模型性能，并使用德龙检验进行比较。校准和决策曲线分析评估模型的临床实用性。进行Kaplan-Meier分析和对数秩检验，以确定影像组学预测的Ki-67表达状态的预后价值。使用SHapley加性解释（SHAP）对最优模型进行解释。
选择了八个影像组学特征，使用随机森林（RF）、极端梯度提升（XGBoost）、逻辑回归（LR）、支持向量机（SVM）和K近邻（KNN）构建模型。RF模型表现最佳，在训练集（0.910，95%置信区间[CI]：0.881 - 0.936）和外部测试集（0.885，95% CI：0.826 - 0.934）中均获得最高曲线下面积（AUC），德龙检验证实（所有P值＜0.05）。校准和决策曲线进一步证明了RF模型卓越的临床实用性。在训练集和外部测试集中，基于IHC和RF预测的高Ki-67表达组均与更高的肿瘤复发风险显著相关（所有P值＜0.05）。使用SHAP在全局和个体层面揭示了RF模型的预测过程。
基于可解释CT影像组学的RF分类器在术前评估Ki-67表达水平方面表现出强大的预测性能，为ccRCC患者提供了有价值的预后见解并有助于临床决策。
