摘要
目的探讨基于动态对比增强磁共振成像(dynamic contrast-enhanced magnetic resonance imaging,DCE-MRI)肿瘤异质性定量和深度学习(deep learning,DL)预测乳腺癌新辅助化疗(neoadjuvant chemotherapy,NAC)病理完全缓解的价值。材料与方法回顾性收集2019年1月至2025年1月在皖南医学院第一附属医院179例经病理证实为乳腺癌的患者临床及影像资料,其中58例患者NAC后病理完全缓解(pathological complete response,pCR),121例患者NAC后病理非完全缓解(non-pathological complete response,non-pCR)。按照7∶3的比例将患者随机分为训练组(n=125)和验证组(n=54)。所有患者均在NAC前行MRI检查,使用ITK-SNAP软件逐层手动勾画感兴趣区(region of interest,ROI)并进行三维融合,使用高斯混合模型(gaussian mixture model,GMM)进行聚类分析及贝叶斯信息准则(bayesian information criterion,BIC)确定肿瘤病灶亚区,并计算肿瘤内异质性分数(intratumoral heterogeneity score,ITH-score),建立生境组学模型。使用Python软件PyRadiomics包提取肿瘤整体的传统影像组学特征,使用ViT(vision transformer)DL模型提取DL特征,采用最小冗余最大相关(minimum redundancy maximum relevance,mRMR)和最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)回归方法进行特征降维、筛选,分别构建传统影像组学模型、DL模型,并根据模型中特征权重计算出每例患者的定量得分。采用多因素logistic回归分析构建临床模型及联合模型。绘制受试者工作特征(receiver operating characteristic,ROC)曲线评估各个模型的预测效能,采用DeLong检验比较各模型效能,使用决策曲线分析(decision curve analysis,DCA)分析模型的临床收益。采用SHAP(Shapley Additive exPlanations)方法分析联合模型中各特征的重要性。结果临床模型、传统影像组学模型、DL模型、生境组学模型及联合模型在训练组中预测NAC后pCR的曲线下面积(area under the curve,AUC)[95%置信区间(confidence interval,CI)]:分别为0.864(0.832~0.895)、0.776(0.745~0.807)、0.728(0.703~0.752)、0.823(0.785~0.881)、0.943(0.903~0.983),在验证组中0.732(0.684~0.781)、0.634(0.589~0.679)、0.757(0.720~0.791)、0.750(0.690~0.840)、0.875(0.821~0.929),以联合模型预测效果最佳。DCA结果显示联合模型的临床获益高于临床模型及其他影像组学模型。在SHAP方法中,ITH-score重要性高于分子分型,其SHAP值越大,预测结果越倾向于pCR。结论基于DCE-MRI的异质性定量分析及DL的联合模型对乳腺癌患者NAC后pCR具有优越的预测效能,对早期预测NAC后pCR具有一定的临床应用价值,有助于乳腺癌的临床诊疗管理。
Objective:To explore the value of dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI)-based tumor heterogeneity quantification integrated with deep learning(DL)in predicting the pathological complete response of neoadjuvant chemotherapy(NAC)for breast cancer.Materials and Methods:The clinical and imaging data of 179 patients with pathologically confirmed breast cancer at the First Affiliated Hospital of Wannan Medical College from January 2019 to January 2025 were retrospectively collected.Among them,58 patients achieved pathological complete response(pCR)after NAC,and 121 patients achieved non-pathological complete response(non-pCR).The patients were randomly divided into a training group(n=125)and a validation group(n=54)at a ratio of 7∶3.All patients underwent MRI examination before NAC.The ITK-SNAP software was used to manually delineate the region of interest(ROI)layer by layer and perform three-dimensional fusion.The Gaussian mixture model(GMM)was used for cluster analysis,and the Bayesian information criterion(BIC)was used to determine the sub-regions of the tumor lesions.The intratumoral heterogeneity score(ITH-score)was calculated,and a habitat imaging model was established.The PyRadiomics package in Python software was used to extract the traditional radiomics features of the whole tumor,and the ViT deep learning model was used to extract the deep learning features.The minimum redundancy maximum relevance(mRMR)and least absolute shrinkage and selection operator(LASSO)regression methods were used for feature dimensionality reduction and screening.A traditional radiomics model and a deep learning model were constructed respectively,and the quantitative score of each patient was calculated according to the feature weights in the models.Multivariate logistic regression analysis was used to construct a clinical model and a combined model.Receiver operating characteristic(ROC)curves were drawn to evaluate the predictive efficacy of each model.The DeLong test was used to compare the efficacy of each model,and decision curve analysis(DCA)was used to analyze the clinical benefits of the models.The SHAP method was used to analyze the importance of each feature in the combined model.Results:The AUC[95%(confidence interval,CI)]values of the clinical model,traditional radiomics model,deep learning model,habitat imaging model,and combined model in predicting pCR after NAC in the training group were 0.864(0.832 to 0.895),0.776(0.745 to 0.807),0.728(0.703 to 0.752),0.823(0.785 to 0.881),and 0.943(0.903 to 0.983)respectively,and in the validation group were 0.732(0.684 to 0.781),0.634(0.589 to 0.679),0.757(0.720 to 0.791),0.750(0.690 to 0.840),and 0.875(0.821 to 0.929)respectively.The combined model had the best predictive performance.The DCA results showed that the clinical benefit of the combined model was higher than that of the clinical model and other radiomics models.In the SHAP method,the importance of the ITH-score was higher than that of the molecular subtype.The larger the SHAP value,the more the prediction result tended to pCR.Conclusions:The combined model based on DCE-MRI heterogeneity quantitative analysis and deep learning demonstrates superior predictive performance for pCR in breast cancer patients after NAC,which holds clinical application value for early prediction of pCR after NAC and contributes to clinical diagnosis and treatment management of breast cancer.
作者
张晴
陈基明
吴莉莉
丁俊
叶慧
夏怡
江璇
焦南林
ZHANG Qing;CHEN Jiming;WU Lili;DING Jun;YE Hui;XIA Yi;JIANG Xuan;JIAO Nanlin(Medical Imaging Center,the First Affiliated Hospital of Wannan Medical College,Wuhu 241001,China;Department of Pathology,the First Affiliated Hospital of Wannan Medical College,Wuhu 241001,China)
出处
《磁共振成像》
北大核心
2026年第1期42-50,共9页
Chinese Journal of Magnetic Resonance Imaging
基金
皖南医学院中青年科研基金项目(编号:WK2024ZQNZ60)。
关键词
乳腺癌
新辅助化疗
磁共振成像
生境成像
深度学习
影像组学
breast cancer
neoadjuvant chemotherapy
magnetic resonance imaging
habitat imaging
deep learning
radiomics
