基于多尺度3D-CT特征预测肾细胞癌组织亚型和WHO/ISUP分级模型的构建与验证  

A Predictive Model of Pathological Subtypes and WHO/ISUP Grade Based on Multi-Scale 3D-CT Features in Renal Cell Carcinoma

作　　者：吴翔 杨焕程 管亿欣 游伟樊 彭卓毅 钟雪儿 袁阳光 张锟 冯晓荣 蒲学佳 刘翰林 WU Xiang;YANG Huancheng;GUAN Yixin(Department of Radiology,Shenzhen Third Hospital,Affiliated to Shenzhen University(Shenzhen Luohu Hospital Group),Shenzhen,Guangdong Province 529000,P.R.China)

机构地区：[1]深圳大学附属第三医院(深圳市罗湖医院集团)放射科,518000 [2]汕头大学医学院深圳罗湖临床学院(深圳市罗湖医院集团) [3]中山大学附属第八医院放射科 [4]深圳市中医院放射科

出　　处：《临床放射学杂志》2024年第6期971-976,共6页Journal of Clinical Radiology

基　　金：广东省基础与应用基础研究基金项目(编号:2019A1515110038)。

摘　　要：目的 探讨基于多尺度3D-CT特征构建的模型能否为肾细胞癌(RCC)组织亚型和WHO/ISUP分级提供稳健的预测。方法 回顾性分析来自4个医疗中心的507例经术后病理证实为RCC的患者,将其分为训练集(中心1~3,346例)、内部验证集(中心1~3,87例)和外部测试集(中心4,74例)。构建的预测模型包含以下模块:由3D-UNet构建的肾脏-肿瘤语义分割模型、基于感兴趣区域的多尺度特征提取器以及由XGBoost算法构建的2个分类器。另外,采用五折交叉验证和多中心外部测试的策略验证和测试模型的稳定性和泛化能力。最后,通过计算单一特征的SHAP量探索每个特征对模型决策的贡献。结果 无论RCC组织亚型还是WHO/ISUP分级预测,相比单期相CT模型,多期相CT模型的准确率和效能都更加优秀。在预测RCC组织亚型模型中,五折交叉验证的AUROC分别为0.86、0.85、0.88、0.88和0.89,最优模型的内部验证集和外部测试集AUROC分别为0.89和0.75。而对于WHO/ISUP分级的预测模型,五折交叉验证的AUROC分别为0.81、0.82、0.79、0.73和0.81,最优模型的内部验证集和外部测试集AUROC分别为0.82和0.73。在模型的可解释性分析上,一阶统计量和灰度矩阵特征分别为RCC组织亚型预测模型贡献量排名第一、二的特征;而在WHO/ISUP分级预测模型中一阶统计量则发挥了最重要的作用。结论 基于多尺度3D-CT特征构建的模型可为术前评估RCC组织亚型和WHO/ISUP分级提供一个强有力的预测。Objective To investigate whether a model constructed based on multi-scale 3D-CT features can provide a robust prediction of renal cell carcinoma(RCC)pathological subtypes and WHO/ISUP grade.Methods A retrospective analysis of 507 patients with postoperative pathologically confirmed RCC from four medical centers were divided into a training set(centers 1-3,346 cases),an internal validation set(centers 1-3,87 cases),and an external test set(center 4,74 cases).The prediction model constructed in this study contained the following modules:a kidney-tumor semantic segmentation model constructed by 3D-UNet,a multi-scale features extractor based on region of interest(ROI),and two classifiers constructed by the XGBoost algorithm.In addition,we used a strategy of five-fold cross-validation and multi-center external testing to validate and test the stability and generalization ability of the models.Finally,we also explored the contribution of each feature to the model decision by calculating the amount of SHAP for a single feature.Results The accuracy and efficacy of the multi-phase CT model were superior compared to the single-phase CT model for both RCC pathological subtype and WHO/ISUP grade prediction.In the predicted RCC pathological subtype model,the AUROC for the five-fold cross-validation were 0.86,0.85,0.88,0.88,and 0.89,respectively,and the AUROC for the internal validation set and external test set of the optimal model were 0.89 and 0.75,respectively.While for the WHO/ISUP grade prediction model,the AUROC for the five-fold cross-validation were 0.81,0.82,0.79,0.7 and 0.81,respectively,and the AUROC for the internal validation set and external test set of the optimal model were 0.89 and 0.75,respectively.In terms of model solvability analysis,the first-order statistics and gray-scale matrix features were the first and second features of the RCC organizational subtype prediction models,respectively;while the first-order statistics in the WHO/ISUP hierarchical prediction model played the most important role in the W

关 键 词：肾细胞癌 多尺度3D-CT特征 机器学习 组织亚型 WHO/ISUP分级 

分 类 号：R737.11[医药卫生—肿瘤] R730.44[医药卫生—临床医学]