基于多参数MRI的影像组学融合模型对乳腺癌腋窝淋巴结转移的术前预测价值  

Preoperative predictive value of a multiparametric MRI radiomics for axillary lymph node metastasis of breast cancer

作　　者：尹智 崔艳芬[2] 任嘉梁 杨晓棠[2] Yin Zhi;Cui Yanfen;Ren Jialiang;Yang Xiaotang(Department of Medical Imaging,Shanxi Medical University,Taiyuan 030000,China;Department of Medical Imaging,Shanxi Province Cancer Hospital,Shanxi Hospital Affiliated to Cancer Hospital,Chinese Academy of Medical Sciences,Cancer Hospital Affiliated to Shanxi Medical University,Taiyuan 030000,China;Department of Pharmaceuticals Diagnostics,GE Healthcare China,Beijing 100000,China)

机构地区：[1]山西医科大学医学影像学院,太原030000 [2]山西省肿瘤医院/中国医学科学院肿瘤医院山西医院/山西医科大学附属肿瘤医院医学影像科,太原030000 [3]北京通用电气公司药品管理科,北京100000

出　　处：《中华解剖与临床杂志》2023年第12期773-781,共9页Chinese Journal of Anatomy and Clinics

基　　金：国家自然科学基金(82171923、82001789);山西省卫健委“四个一批”科技兴医创新计划(2020TD09、2021XM51)。

摘　　要：目的探讨基于多参数MRI的影像组学融合模型在乳腺癌术前预测腋窝淋巴结(ALN)转移的应用价值。方法回顾性队列研究。纳入山西省肿瘤医院2020年8月—2021年9月经病理证实的272例乳腺癌患者的多参数MRI及临床病理资料。患者均为女性,年龄28~79(53.0±10.9)岁,其中ALN阳性107例、ALN阴性165例。按照7∶3的比例随机将患者分为训练组(191例)和验证组(81例)。从T2加权像(T2WI)、表观弥散系数(ADC)图和增强T1加权像(cT1WI)序列中提取影像组学特征。采用单因素逻辑回归、相关性分析和Boruta算法3个步骤进行特征选择,然后采用支持向量机(SVM)、随机森林(RF)和逻辑回归(LR)3种机器学习方法构建影像组学模型,并基于最优模型计算每位患者的影像组学分数(Radscore)。同时,通过多因素逐步回归分析筛选乳腺癌ALN转移的独立危险因素并构建临床模型。最后,联合Radscore和临床独立危险因素构建融合模型,并绘制列线图。采用受试者操作特征曲线、校准曲线和决策曲线(DCA)来评价模型对乳腺癌ALN转移的预测性能及临床效益。结果训练组和验证组患者肿瘤位置比较,差异有统计学意义(P<0.05);训练组中ALN阳性与ALN阴性患者间的肿瘤位置、MRI评估淋巴结状态比较,验证组中ALN阳性与ALN阴性患者间的雌激素受体、分子亚型及MRI评估淋巴结状态比较,差异均有统计学意义(P值均<0.05)。基于多参数MRI降维选择后,得到了6个与ALN转移呈显著相关的影像组学特征(P值均<0.05)。在训练组和验证组中,SVM、RF和LR模型均表现出很好的预测能力,AUC分别为0.784、0.826、0.703和0.733、0.817、0.703,其中RF模型效能最高。单因素、多因素回归分析显示,MRI评估淋巴结状态是乳腺癌ALN转移的独立预测因子[比值比(95%可信区间)=10.909(5.210~24.511),P<0.001],采用这一指标构建临床模型。联合Radscore和MRI评估ALN状态的融合模型在训�Objective This study aims to explore the preoperative predictive value of a combined model of multiparametric MRI and radiomics for predicting axillary lymph node(ALN)metastasis in patients with breast cancer.Methods Retrospective cohort study was conducted.This study included the MRI and clinico pathological information of 272 patients with pathologically confirmed breast cancer in Shanxi Province Tumor Hospital from August 2020 to September 2021.All patients were female,aged 28−79(53.0±10.9)years old,including 107 ALN-positive and 165 ALN-negative cases.Patients were randomly divided into a training cohort(191 cases)and a validation cohort(81 cases)at 7∶3 ratio.Radiomics features were extracted from T2-weighted imaging,apparent diffusion coefficient,and enhanced T1-weighted imaging sequences.Three steps(single-factor logistic regression,correlation analysis,and Boruta algorithm)were performed to select features.Based on the selected features,three machine-learning models including support vector machine(SVM),random forest(RF),and logistic regression(LR)were used to construct the radiomics model,and then the radiomics score(Radscore)of each patient based on the optimal model was obtained.A clinical model was built using multivariate stepwise regression based on the independent clinical risk factors.Finally,a combined model was constructed by joining the Radscore with clinical independent risk factors,followed by developing a nomogram.Receiver operating characteristic curve,calibration curve,and decision curve(DCA)were used to evaluate the predictive performance and clinical benefit.Results Significant differences existed in tumor location between the training and validation cohorts(P<0.05).Tumor location and MRI-defined ALN status had statistically significant differences between ALN-positive and ALN-negative patients in the training cohort,estrogen receptor,molecular subtypes,and MRI-defined ALN status between ALN-positive and ALN-negative patients in the verification cohort(all P values<0.05).Based on mul

关 键 词：乳腺肿瘤 磁共振成像 影像组学 机器学习 腋窝淋巴结转移 

分 类 号：R445.2[医药卫生—影像医学与核医学] R737.9[医药卫生—诊断学]