| Objective:Breast cancer is the most common malignancy tumor in women,and its incidence is increasing year by year,which seriously endangers women’s health.With the development of individualized precision medicine,the expression status of tumor im munohistochemical indexs have great significance to the formulation of clinical treatment and the evaluation of prognosis.ER、PR、HER-2 and Ki67 are the most common and important immunohistochemical indicators in breast cancer,and they mainly detected by pathological examination at present,which is not only invasive and time-consuming,but also may have false negative results due to the heterogeneity of breast cancer.Radiomics has the advantages of simple,non-invasive and comprehensive prediction of immunohistochemical indexes in breast cancer.Therefore,this study mainly to explore the value of radiomics models based on dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI)by using different algorithms in preoperative prediction of ER 、PR、 HER-2 and Ki67 status in breast cancer,to help the clinical preoperative the prediction of immunohistochemical status in breast cancer.Materials and methods:Clinical,pathological and imaging data of 124 female patients with breast cancer were collected in the Hospital of Shandong University of traditional Chinese Medicine from January 2019 to November 2022.Doctor draw the outline of each images from the third phase of DCE-MRI by using the software of 3D-slicer,saved as the region of interest(ROI).Radiomics features were extracted by the software of pyradiomics in Python.1032 radiomics features were extracted from the third phase of DCE-MRI images.Regard the expression level of ER、PR、HER-2 and Ki67 in breast cancer by immunohistochemistry tested as the reference standard,divided the ER、PR、HER-2 and Ki-67 into two groups.First of all,one-way ANOVA,Spearman correlation analysis and logistic regression with L2 penalty were used as the basis model to select the best radiomics features for each immunohistochemical index.Then,four kinds of machine learning algorithms contained logistic regression(LR),support vector machine(SVM),random forest(RF)and Light GBM(LGB),were used to construct models to predict ER、PR 、HER-2 and Ki67 expression respectively,which were verified in the test set simultaneously.The receiver operating characteristic curve(ROC)was used to evaluate the prediction efficiency for each immuno histochemical index of four models,and the corresponding area under the curve(AUC)was calculated.The stability of the model for each immunohistochemical index was evaluated by the stratified five-fold cross-validation ROC curve,and the clinical values of four models for each immunohistochemical index were analyzed by decision curve analysis(DCA).The best model was selected to predict ER、PR、HER-2 and Ki67 expression in breast cancer.Results:According to the expression level of ER,PR,HER-2 and Ki-67,8,6,7 and 6 key features were selected respectively to construct the models from the third phase of DCEMRI,the predictive effectiveness of each immunohistochemical index based on four models constructed by LR、SVM、RF and LGB algorithms,the specific information is as follows:(1)The ER of the AUCs based on four algorithms of LR、SVM、RF and LGB to construct the models in the test set were 0.628、0.744、0.750、0.783,respectively.The model constructed by LR algorithm to prediction the ER status,which accuracy,sensitivity,specificity were 0.622、0.654 and 0.545,respectively.The model constructed by SVM algorithm to prediction the ER status,which accuracy,sensitivity,specificity were 0.703、0.731 and 0.636,respectively.The model constructed by RF algorithm to prediction the ER status,which accuracy,sensitivity,specificity were 0.730,0.769 and 0.636,respectively.The model constructed by LGB algorithm to prediction the ER status,which accuracy,sensitivity,specificity were 0.757,0.769 and 0.727,respectively.The average AUCs of the stratified five-fold cross-validation based on these four algorithms to construct the models were 0.742、0.768、0.791、0.809,respectively.(2)The PR of the AUCs based on four algorithms of LR、SVM、RF and LGB to construct the models in the test set were 0.688、0.767、0.728、0.737,respectively.The model constructed by LR algorithm to prediction the PR status,which accuracy,sensitivity,specificity were 0.622,0.652 and 0.571,respectively.The model constructed by SVM algorithm to prediction the PR status,which accuracy,sensitivity,specificity were 0.730,0.739 and 0.714,respectively.The model constructed by RF algorithm to prediction the PR status,which accuracy,sensitivity,specificity were 0.676,0.609 and 0.786,respectively.The model constructed by LGB algorithm to prediction the PR status,which accuracy,sensitivity,specificity were 0.703,0.739 and 0.692,respectively.The average AUCs of the stratified five-fold cross-validation based on these four algorithms to construct the models were 0.770、0.796、0.684、0.737,respectively.(3)The HER-2 of the AUCs based on four algorithms of LR、SVM、RF and LGB to construct the models in the test set were 0.769、0.824、0.810、0.843,respectively.The model constructed by LR algorithm to prediction the HER-2 status,which accuracy,sensitivity,specificity were 0.703,0.692 and 0.727,respectively.The model constructed by SVM algorithm to prediction the HER-2 status,which accuracy,sensitivity,specificity were 0.757,0.731 and 0.818,respectively.The model constructed by RF algorithm to prediction the HER-2 status,which accuracy,sensitivity,specficity were 0.757,0.769 and 0.727,respectively.The model constructed by LGB algorithm to prediction the HER-2 status,which accuracy,sensitivity,specificity were 0.811,0.808 and 0.818,respectively.The average AUCs of the stratified five-fold cross-validation based on these four algorithms to construct the models were 0.771、0.806、0.796、0.825,respectively.(4)The Ki67 of the AUCs based on four algorithms of LR、SVM、RF and LGB to construct the models in the test set were 0.724、0.815、0.762、0.796,respectively.The model constructed by LR algorithm to prediction the Ki67 status,which accuracy,sensitivity,specificity were 0.676,0.708 and 0.615,respectively.The model constructed by SVM algorithm to prediction the Ki67 status,which accuracy,sensitivity,specificity were 0.784,0.792 and 0.769,respectively.The model constructed by RF algorithm to prediction the Ki67 status,which accuracy,sensitivity,specificity were 0.730,0.750 and 0.692,respectively.The model constructed by LGB algorithm to prediction the Ki67 status,which accuracy,sensitivity,specificity were 0.757,0.792 and 0.692,respectively.The average AUCs of the stratified five-fold cross-validation based on these four algorithms to construct the models were 0.773、0.819、0.795、0.787,respectively.Conclusion:(1)These four algorithms contained LR、SVM、RF and LGB were used to construct models from DCE-MRI images have a certain predictive efficiency of ER 、PR、HER-2 and Ki67 status in breast cancer,which can be used to help the clinical preoperative the prediction of immunohistochemical status in breast cancer.(2)Models constructed by different algorithms have predictive effectiveness in immunohistochemical status of breast cancer from DCE-MRI images.The model constructed by LGB algorithm have a best preoperative performance for ER and HER-2 status in breast cancer.The model constructed by SVM algorithm have a best preoperative performance for PR and Ki67 status in breast cancer. |
PDF Full Text Request