基于超声和临床特征决策树模型预测乳腺癌分子亚型  

Prediction of molecular subtypes of breast cancer using decision tree model based on ultrasonic and clinical features

作　　者：施维 夏威[1] 黄敏[2] 常才 王宇 游小慧 顾华芸[2] 董智芬 郭建锋[2] SHI Wei;XIA Wei;HUANG Min;CHANG Cai;WANG Yu;YOU Xiaohui;GU Huayun;DONG Zhifen;GUO Jianfeng(Department of Medical Imaging,Suzhou Institute of Biomedical Engineering and Technology,Chinese Academy of Science,Suzhou 215163,Jiangsu Province,China;Department of Ultrasound,Affiliated Suzhou Hospital of Nanjing Medical University,Suzhou Municipal Hospital,Suzhou 215001,Jiangsu Province,China;Department of Ultrasound,Fudan University Shanghai Cancer Center,Shanghai 200032,China)

机构地区：[1]中国科学院苏州生物医学工程技术研究所医学影像技术研究室,江苏苏州215163 [2]南京医科大学附属苏州医院,苏州市立医院超声科,江苏苏州215001 [3]复旦大学附属肿瘤医院超声科,上海200032

出　　处：《肿瘤影像学》2022年第3期249-257,共9页Oncoradiology

基　　金：广东省重点领域研发计划(2019B010152001);江苏省卫计委“六个一”人才项目(LGY2017009);苏州市科技局项目(SYS201767);苏州市科技计划项目(SYG201908)。

摘　　要：目的:建立不同亚型浸润性乳腺癌的超声决策树预测模型,并分析模型的临床价值。方法:回顾并分析420例经病理学检查证实的浸润性乳腺癌患者,其中管腔A(Luminal A,LA)型患者137例、管腔B(Luminal B,LB)型患者157例、人表皮生长因子受体2过表达(human epidermal growth factor receptor 2 over-expression,HER2+)型患者61例和三阴性乳腺癌(triple-negative breast cancer,TNBC)型患者65例。使用方差分析和Fisher精确概率检验统计分析患者的超声特征和临床特征,将差异有统计学意义的特征纳入决策树模型以预测乳腺癌分子亚型。采用受试者工作特征(receiver operating characteristic,ROC)曲线的曲线下面积(areaundercurve,AUC)评价模型的预测效果。结果:对于4种乳腺癌分子亚型,找到7种差异有统计学意义(P<0.05)的临床和超声特征:临床分期、肿瘤最大径、肿瘤内部回声变化、肿瘤后方回声变化、钙化形态、钙化部位、有无转移性淋巴结。LA、LB、HER2+和TNBC型决策树模型的训练集AUC分别为0.731、0.708、0.722和0.877。此外,与高年资超声科医师相比,决策树模型鉴别TNBC型的灵敏度(81.0%)较高。结论:基于超声和临床特征构建的决策树模型可以准确地预测乳腺癌分子亚型。Objective: To establish the ultrasonic decision tree model for predicting different subtypes of invasive breast cancer,and to analyze the clinical value of the model. Methods: A total of 420 invasive breast cancer patients were analyzed retrospectively,which were confirmed by pathology. All patients were divided into 4 molecular subtypes: Luminal A(LA) type(n=137), Luminal B(LB) type(n=157), human epidermal growth factor receptor 2 over-expression(HER2+) type(n=61) and triple-negative breast cancer(TNBC) type(n=65). The ultrasonic and clinical features of patients were evaluated by analysis of variance and Fisher exact probability test, and the statistically significant features were included in the decision tree model for molecular subtype identification of breast cancer. The corresponding area under the receiver operating characteristic(ROC) curves(AUC) were calculated to assess the performance of each decision tree. Results: Seven features, including clinical stage, maximum diameter, echo patterns,posterior acoustic features, calcification pattern, the position of calcification, and lymph node metastasis, were statistically significant differences(P<0.05) among four molecular subtypes. The AUC of LA, LB, HER2+ and TNBC type decision tree models were 0.721,0.708, 0.722, 0.877 respectively in training sets. Moreover, the sensitivity of the decision tree model(81.0%) for TNBC type was higher than that of the senior sonographer. Conclusion: The decision tree built by ultrasonic and clinical features has favorable value in the prediction of breast cancer molecular subtypes with high diagnostic accuracy.

关 键 词：乳腺癌 分子亚型 超声 决策树 

分 类 号：R737.9[医药卫生—肿瘤] R445.1[医药卫生—临床医学]