Artificial neural network model based on recursive feature elimination-support vector machine for differentiating ductal carcinoma in situ and complicated with microinvasion
10.13929/j.issn.1003-3289.2024.09.015
- VernacularTitle: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
- Author:
Xiaoping ZHOU
1
;
Wei YANG
;
Qingyun YIN
;
Chaolin ZHANG
;
Ningmei ZHANG
Author Information
1. 宁夏医科大学第一临床医学院,宁夏银川 750004
- Keywords:
breast neoplasms;
magnetic resonance imaging;
mammography;
machine learning
- From:
Chinese Journal of Medical Imaging Technology
2024;40(9):1345-1350
- CountryChina
- Language:Chinese
-
Abstract:
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(ADCmin),nuclear grade,ADCheterogeneity,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.
