基于影像特征建立乳腺癌腋窝淋巴结转移机器学习预测模型  

A machine learning model for predicting axillary lymph node metastasis in breast cancer based on imaging features

作　　者：吴泽琪 马梦伟 刘仁懿 曾凤霞 陈卫国[1] 秦耿耿[1] WU Zeqi;MA Mengwei;LIU Renyi;ZENG Fengxia;CHEN Weiguo;QIN Genggeng(Department of Radiology,Nanfang Hospital,Southern Medical University,Guangzhou 510515,China)

机构地区：[1]南方医科大学南方医院放射科,广州510515

出　　处：《国际医学放射学杂志》2023年第3期255-260,共6页International Journal of Medical Radiology

基　　金：国家自然科学基金(82171929);国家重点研发计划(2019YFC0121903);吴阶平医学基金会临床科研专项资助基金课题(320.6750.2021-06-36)。

摘　　要：目的探索乳腺癌原发灶的X线及超声征象术前无创预测腋窝淋巴结(ALN)转移的价值。方法回顾性收集320例经病理证实为浸润性乳腺癌女性病人的临床及影像资料。根据ALN病理结果是否有转移灶,将病人分为ALN阳性组(154例)和阴性组(166例)。数据集以8∶2的比例随机分为训练集(256例)和测试集(64例)。提取乳腺癌原发灶的全视野数字化乳腺X线摄影(FFDM)和超声(US)影像征象,选择最优算法分别构建基于FFDM、US征象以及两者联合的ALN转移预测模型,根据受试者操作特征(ROC)曲线下面积(AUC)评估模型效能并选取最优模型。采用SHAP可解释技术可视化最优模型决策过程,解析乳腺癌ALN转移的影像预测因子。2组病人FFDM和超声影像征象比较采用Mann-Whitney U检验、卡方检验或Fisher确切概率检验。采用Delong检验比较各模型的AUC。结果基于FFDM、US以及联合特征构建的3种模型,其AUC值有递增趋势,但AUC差异均无统计学意义(均P>0.05)。SHAP可解释技术显示,乳腺癌ALN转移的前5个重要影像预测因子包括2个US征象(肿块最大径、强回声光点)以及3个FFDM征象(皮肤增厚、乳头回缩、乳腺密度散在纤维腺体类)。其中,US显示肿块较大、FFDM上表现为皮肤增厚或乳头回缩时,模型更倾向于预测ALN转移为阳性;而在US上未发现强回声光点或在FFDM上乳腺腺体类型表现为散在纤维腺体类时,模型更倾向于预测ALN阴性。结论基于FFDM及US征象构建的可解释机器学习模型能较好地预测乳腺癌ALN转移,有望成为术前无创预测乳腺癌ALN转移的新手段。Objective To explore the value of mammography and ultrasonography signs of primary breast cancer for preoperative non-invasive prediction of axillary lymph node(ALN)metastasis.Methods The clinical and imaging data of 320 female patients with pathologically confirmed invasive breast cancer were analyzed retrospectively.The cases were divided into ALN-positive group(n=154)and ALN-negative group(n=166)according to the pathological results of ALN.The sample was randomly divided into a training set(n=256)and a testing set(n=64)in a proportion of 8∶2.Full-field digital mammography(FFDM)and ultrasound(US)features of primary breast cancer were extracted,and the optimal algorithms were selected to build the ALN metastasis prediction models based on FFDM,US,and combined features.Receiver operating characteristic curves and the area under the curve(AUC)were used to evaluate each model’s performance,and to select the optimal model.The SHapley Additive exPlanation(SHAP)value was used to visualize the decision process of the optimal model and identify the imaging predictors of axillary lymph node metastasis in breast cancer.We compared the age,FFDM,and US imaging features between the ALN-positive and ALN-negative groups using the Mann-Whitney U test,chi-square test,and Fisher’s exact test.We used the Delong test to compare the AUCs between different models.Results The AUCs of the FFDM,US,and the combined models showed an increasing trend,but there was no significant difference in AUC between any two groups(all P>0.05).According to the SHAP value,the top five imaging predictors of ALN metastasis contain two US features,including the maximum tumor diameter and strong punctate echo,and three FFDM features,including skin thickening,nipple retraction,and scattered areas of fibro-glandular density.The model’s output was more likely to be ALN metastasis positive when the tumor was larger in US or the skin thickening or nipple retraction in FFDM.However,the output of the model was more likely to be ALN metastasis negative when

关 键 词：乳腺肿瘤 淋巴转移 乳腺X线摄影术 超声检查 机器学习 

分 类 号：R737.9[医药卫生—肿瘤] R445.1[医药卫生—临床医学] R814