摘要
目的探讨基于数字乳腺断层摄影(DBT)的影像组学模型预测乳腺癌Ki-67表达状态和生存预后的价值,并进行模型可解释性分析。方法该研究为横断面回顾性研究。收集2019年1月至2020年8月复旦大学附属肿瘤医院532例和2021年12月至2022年5月上海交通大学附属瑞金医院137例经术后病理证实为浸润性乳腺癌患者(分别为队列1、队列2)的临床、病理资料及DBT资料。分别在DBT图像的头尾位和内外斜(MLO)位对患者的病灶进行勾画,构建掩膜矩阵并提取影像组学特征。将队列1按照8∶2的比例分为训练集425例和测试集107例,队列2作为外部验证集。经特征降维和筛选后,将保留的特征放入逻辑回归、支持向量机和极端梯度提升(XGBoost)3个机器学习模型,利用受试者操作特征曲线和曲线下面积(AUC)评估3种模型预测Ki-67表达状态的效能,并选出最优模型。进一步将训练集、验证集和外部测试集的患者分成Luminal型亚组和非Luminal型亚组,比较2个亚组中最优模型预测Ki-67表达状态的效能。采用最优模型生成Ki-67评分,并与无病生存期(DFS)相关联。使用Shapley加性解释(SHAP)来解释模型。结果最终从DBT病灶提取10个影像组学特征纳入模型,3种模型中XGBoost模型预测Ki-67表达状态效能最好,为最优模型,AUC值在训练集、测试集和外部验证集分别为0.89、0.77和0.74。Luminal型亚组训练集、测试集和外部验证集中XGBoost模型预测Ki-67表达状态的AUC分别为0.90、0.78和0.77,优于非luminal型亚组(训练集、测试集和外部验证集的AUC分别为0.88、0.76和0.68)。根据XGBoost模型生成患者的Ki-67评分,使用X-tile软件确定最佳生存阈值(0.25分),训练集和测试集中Ki-67评分>0.25分是较短DFS的危险因素(训练集:HR=2.76,95%CI 1.25~6.12,P<0.05;测试集:HR=5.07,95%CI 1.16~22.19,P<0.05)。SHAP可解释性分析显示,来自MLO位的小波特征在最终筛选出的特征中占比最多,形态学特征虽然占比不多但权重较大。结论基于可解释的DBT影像组学模型能有效预测乳腺癌患者的Ki-67表达状态和生存预后情况,为个体化诊疗提供非侵入性评估工具。
ObjectiveTo explore the value of radiomics models based on digital breast tomosynthesis(DBT)in predicting Ki-67 expression and survival prognosis in breast cancer,and to conduct an analysis of model interpretability.MethodsThis was a cross-sectional retrospective study.A total of 532 patients from Fudan University Shanghai Cancer Center from January 2019 to August 2020 were retrospectively enrolled as Cohort 1,and 137 breast cancer patients from Ruijin Hospital,Shanghai Jiao Tong University from December 2021 to October 2022 were enrolled as Cohort 2.All patients were pathologically confirmed invasive breast cancer.Clinical and pathological data as well as DBT images were collected for each patient.For each patient′s lesion,segmentation was performed on both the craniocaudal and mediolateral oblique(MLO)views of the DBT images to construct mask matrices and extract radiomics features.Cohort 1 was divided into a training set of 425 cases and a testing set of 107 cases in an 8∶2 ratio,while Cohort 2 served as the external validation set.After feature dimensionality reduction and selection,the retained features were input into three machine learning models:logistic regression,support vector machine,and extreme gradient boosting(XGBoost).The area under the receiver operating characteristic curve(AUC)was used to evaluate the performance of the three models in predicting Ki-67 expression,and the optimal model was selected.Patients in the training,testing,and external validation sets were further divided into Luminal and non-Luminal subgroups,and the performance of the optimal model in predicting Ki-67 expression status was compared between the two subgroups.The optimal model was used to generate a Ki-67 score,which was then correlated with disease-free survival(DFS).The shapley additive explanations(SHAP)were used to interpret the model.ResultsA total of 10 radiomics features extracted from DBT lesions were ultimately included in the models.Among the three models,the XGBoost model demonstrated the best performance in predicting Ki-67 expression status and was selected as the optimal model,with AUC values of 0.89,0.77,and 0.74 in the training set,testing set,and external validation set,respectively.In the Luminal subgroup,the XGBoost model achieved AUCs of 0.90,0.78,and 0.77 in the training,testing,and external validation sets,respectively,and outperforming the non-Luminal subgroup,whose AUCs were 0.88,0.76,and 0.68.Ki-67 scores were generated for each patient using the XGBoost model,and the optimal prognostic cutoff(0.25)was determined with X-tile software.A Ki-67 score>0.25 was associated with shorter disease-free survival in both the training set(HR=2.76,95%CI 1.25‒6.12,P<0.05)and the testing set(HR=5.07,95%CI 1.16‒22.19,P<0.05).SHAP revealed that wavelet features derived from the MLO view accounted for the largest proportion of the finally selected features.Although morphological features had a smaller proportion,they carried greater weight.ConclusionThis study developes an interpretable DBT radiomics model to predict Ki-67 expression status and survival outcomes in breast cancer patients,providing a non-invasive assessment tool for personalized diagnosis and treatment.
作者
李佳蔚
李金辉
伏秋燚
贺盛恒
范明
姜婷婷
厉力华
彭卫军
顾雅佳
柴维敏
尤超
Li Jiawei;Li Jinhui;Fu Qiuyi;He Shengheng;Fan Ming;Jiang Tingting;Li Lihua;Peng Weijun;Gu Yajia;Chai Weimin;You Chao(Department of Radiology,Fudan University Shanghai Cancer Center,Department of Oncology,Shanghai Medical College,Fudan University,Shanghai 200032,China;Department of Radiology,Ruijin Hospital,Shanghai Jiaotong University School of Medicine,Shanghai 200025,China;Institute of Biomedical Engineering and Instrumentation,Hangzhou Dianzi University,Hangzhou 310018,China)
出处
《中华放射学杂志》
北大核心
2026年第1期70-78,共9页
Chinese Journal of Radiology
基金
上海市卫生健康委员会面上项目(202240241)
国家癌症中心攀登基金重点项目(NCC201909B06)。
