基于递归特征消除支持向量机人工神经网络模型鉴别乳腺导管原位癌及其伴微浸润  

Artificial neural network model based on recursive feature elimination-support vector machine for differentiating ductal carcinoma in situ and complicated with microinvasion

作　　者：周晓平 杨蔚[2] 尹清云[3] 张朝林[4] 张宁妹[5] ZHOU Xiaoping;YANG Wei;YIN Qingyun;ZHANG Chaolin;ZHANG Ningmei(the First School of Clinical Medicine,Ningxia Medical University,Yinchuan 750004,China;Department of Radiology,General Hospital of Ningxia Medical University,Yinchuan 750004,China;Department of Medical Oncology,General Hospital of Ningxia Medical University,Yinchuan 750004,China;Department of Oncology Surgery,General Hospital of Ningxia Medical University,Yinchuan 750004,China;Department of Pathology,General Hospital of Ningxia Medical University,Yinchuan 750004,China)

机构地区：[1]宁夏医科大学第一临床医学院,宁夏银川750004 [2]宁夏医科大学总医院放射科,宁夏银川750004 [3]宁夏医科大学总医院肿瘤内科,宁夏银川750004 [4]宁夏医科大学总医院肿瘤外科,宁夏银川750004 [5]宁夏医科大学总医院病理科,宁夏银川750004

出　　处：《中国医学影像技术》2024年第9期1345-1350,共6页Chinese Journal of Medical Imaging Technology

基　　金：宁夏回族自治区重点研发计划项目(2022BEG03166)。

摘　　要：目的观察基于递归特征消除支持向量机(RFE-SVM)人工神经网络(ANN)模型鉴别乳腺导管原位癌(DCIS)及其伴微浸润(DCISM)的价值。方法回顾性纳入296例女性单发乳腺癌(DCIS 244例、DCISM 52例)作为训练集,另前瞻性收集120例女性单发乳腺癌(DCIS 87例、DCISM 33例)作为验证集;比较集间一般资料、乳腺钼靶及MRI表现,筛选最优特征子集并构建ANN模型;绘制受试者工作特征曲线,计算曲线下面积(AUC),评估ANN模型鉴别DCIS与DCISM的效能。结果位列前10的最优特征子集依次为Ki-67指数、最小表观弥散系数(ADC_(min))、核分级、ADC_(异质性)、病灶最长径、年龄、P63、病灶强化类型、病灶钙化状态及病灶坏死。ANN模型鉴别训练集DCIS与DCISM的准确率、敏感度、特异度、阳性预测值、阴性预测值及AUC分别为91.55%、63.46%、97.54%、84.62%、92.61%及0.950,在验证集分别为80.00%、69.70%、83.91%、62.16%、87.95%及0.896;其在训练集和验证集的校准曲线与理想曲线走行均基本一致(P=0.355、0.480),且具有较高临床净收益。结论RFE-SVM ANN模型可有效鉴别DCIS与DCISM。Objective To observe the value of artificial neural network(ANN)model based on recursive feature elimination-support vector machine(RFE-SVM)for differentiating ductal carcinoma in situ(DCIS)and DCIS complicated with microinvasion(DCISM).Methods Totally 296 female patients with single breast cancer(244 cases of DCIS and 52 cases of DCISM)were retrospectively collected as training set.Then 120 female patients with single breast cancer(87 cases of DCIS and 33 cases of DCISM)were prospectively enrolled as validation set.The general data,mammography and MRI findings were compared between sets.The optimal feature subsets for establishing ANN model were screened.Receiver operating characteristic curve was drawn,and the area under the curve(AUC)was calculated to evaluate the efficacy of ANN model for differentiating DCIS and DCISM.Results Ki-67 index,the minimum apparent diffusion coefficient(ADC_(min)),nuclear grade,ADC_(heterogeneity),maximum diameter of lesion,patient's age,P63,lesion enhancement type,calcification status and necrosis were the selected top 10 optimal feature subsets.The accuracy,sensitivity,specificity,positive predictive,negative predictive and AUC of ANN model for differentiating DCIS and DCISM was 91.55%,63.46%,97.54%,84.62%,92.61%and 0.950 in training set,respectively,while was 80.00%,69.70%,83.91%,62.16%,87.95%and 0.896 in validation set,respectively.The calibration curves of ANN model were consistent with the ideal curves in both training and validation set(P=0.355,0.480),which also expressed high clinical net benefit.Conclusion ANN model based on SVM-RFE could be used to differentiate DCIS and DCISM effectively.

关 键 词：乳腺肿瘤 磁共振成像 乳房X线摄影术 机器学习 

分 类 号：R737.9[医药卫生—肿瘤] R445.2[医药卫生—临床医学]