基于MRI生境影像组学预测肾细胞癌区域淋巴结转移的研究  

MRI-based habitat radiomics for evaluating lymph node metastasis in renal cell carcinoma

作　　者：白旭 傅旭 徐鸿昊 周少鹏 贾通宇 易思成 赵厚铭 刘波 刘鑫 刘海丽 穆学涛[7] 张萌萌 奇丽霞 叶慧义[1] 马鑫[3] 王海屹[1] Bai Xu;Fu Xu;Xu Honghao;Zhou Shaopeng;Jia Tongyu;Yi Sicheng;Zhao Houming;Liu Bo;Liu Xin;Liu Haili;Mu Xuetao;Zhang Mengmeng;Qi Lixia;Ye Huiyi;Ma Xin;Wang Haiyi(Department of Radiology,the First Medical Center of Chinese PLA General Hospital,Beijing 100853,China;School of Engineering Medicine,Beihang University,Beijing 100191,China;Department of Urology,the Third Medical Center of Chinese PLA General Hospital,Beijing 100039,China;CT-MRI Room,Ordos Central Hospital,Ordos 017000,China;Department of Radiology,Chinese PLA 920 Hospital,Kunming 650118,China;Department of Radiology,the Fourth Medical Center of Chinese PLA General Hospital,Beijing 100048,China;Department of Radiology,the Third Medical Center of Chinese PLA General Hospital,Beijing 100039,China;Department of Radiology,the Fifth Medical Center of Chinese PLA General Hospital,Beijing 100039,China)

机构地区：[1]解放军总医院第一医学中心放射诊断科,北京100853 [2]北京航空航天大学医学科学与工程学院,北京100191 [3]解放军总医院第三医学中心泌尿外科医学部,北京100039 [4]鄂尔多斯市中心医院CT-MRI室,鄂尔多斯017000 [5]解放军联勤保障部队第九二〇医院放射科,昆明650118 [6]解放军总医院第四医学中心放射诊断科,北京100048 [7]解放军总医院第三医学中心放射诊断科,北京100039 [8]解放军总医院第五医学中心放射科,北京100039

出　　处：《中华放射学杂志》2025年第4期384-392,共9页Chinese Journal of Radiology

基　　金：国家自然科学基金(82271951,U24A20755)。

摘　　要：目的通过肾脏MRI生境影像组学方法构建机器学习模型,评估其术前预测肾细胞癌(RCC)区域淋巴结转移的效能。方法本研究为横断面研究,回顾性收集2010年1月至2023年8月来自解放军总医院4个医学中心接受肿瘤及淋巴结切除手术的220例RCC患者,其中淋巴结转移组65例、淋巴结未转移组155例。采用分层抽样随机方法,以8∶2的比例将第一医学中心的175例患者分为训练集(140例)和内部测试集(35例),第三、第四、第五医学中心的45例RCC患者为外部测试集。依据肾脏MRI皮髓质期强化程度及T 2WI信号强度,将RCC原发灶划分为15个生境子区,分析不同子区体积占比。基于训练集MRI生境子区的影像组学特征,利用多种机器学习算法构建影像组学模型,包括极端随机树(ET)、梯度提升决策树(GBDT)、随机森林(RF)及支持向量机(SVM)模型,选取最优模型,联合淋巴结短径构建联合模型,采用受试者操作特征曲线评估各模型预测RCC淋巴结转移的效能。结果淋巴结未转移组的肿瘤高强化高信号区体积占比较淋巴结转移组更高,分别为0.05±0.09和0.02±0.03,差异有统计学意义(t=3.00,P=0.003)。基于15个生境影像组学特征构建的机器学习模型中,在内部测试集及外部测试集中,SVM模型的曲线下面积(AUC)分别为0.85(95%CI 0.72~0.98)、0.82(95%CI 0.67~0.98),均大于ET、GBDT和RF模型的AUC值,SVM模型表现最佳。在内部测试集和外部测试集中,结合SVM模型与淋巴结短径的联合模型AUC分别为0.94(95%CI 0.85~1.00)及0.89(95%CI 0.78~1.00),淋巴结短径分别为0.81(95%CI 0.66~0.96)、0.81(95%CI 0.68~0.94)。联合模型在内部测试集和外部测试集的诊断灵敏度分别为91.7%、85.7%,特异度分别为78.3%、67.7%。结论基于RCC的MRI生境影像组学与淋巴结短径的联合模型具有良好的区域淋巴结转移术前评估能力。ObjectiveTo evaluate the efficacy of preoperative prediction of regional lymph node(RLN)metastasis in renal cell carcinoma(RCC)using a machine learning model based on habitat imaging radiomics from renal MRI.MethodsThis cross-sectional study retrospectively analyzed 220 patients with RCC who underwent nephrectomy and RLN dissection at four medical centers of Chinese PLA General Hospital from January 2010 to August 2023.The cohort included 65 patients with RLN metastasis and 155 without.A stratified random sampling method was used to divide 175 patients from the first medical center into a training set(n=140)and an internal test set(n=35)in an 8∶2 ratio,while 45 patients from the third,fourth,and fifth medical centers constituted the external test set.The primary RCC lesions were categorized into 15 habitat subregions based on corticomedullary-phase enhancement and T 2WI signal intensity on MRI,and the volume fractions of different subregions were analyzed.In the training cohort,radiomics features derived from the habitat subregions were used to construct a radiomics model employing various machine learning algorithms,including extremely random trees(ET),gradient boosting decision trees(GBDT),random forest(RF),and support vector machine(SVM).The optimal model was selected and combined with RLN short-axis diameter to develop a combined model.The efficacy of each model in predicting RLN metastasis was evaluated using the receiver operating characteristic(ROC)curve.ResultsThe volume fraction of hyper-enhanced hyper-intense regions in the non-metastatic group was significantly higher than that in the metastatic group(0.05±0.09 vs.0.02±0.03;t=3.00,P=0.003).Among the machine learning models constructed using 15 optimal habitat radiomics features,the SVM model demonstrated the best performance,with area under the ROC curve(AUC)values of 0.85(95%CI 0.72-0.98)in the internal test set and 0.82(95%CI 0.67-0.98)in the external test set,surpassing those of the ET,GBDT,and RF models.The combined model,integrating the SVM mo

关 键 词：磁共振成像 肾细胞癌 淋巴结转移 影像组学 生境分析 

分 类 号：R737.11[医药卫生—肿瘤]