Prediction Of Lymph Node Metastasis In Breast Cancer By Radiomics Based On Multiparametric MRI

Objectives:The aim of this study was to establish a prediction model of lymph node metastasis in breast cancer by radiomics based on DCE and DWI MRI,and to compare the predictive performance of the different models.Methods:Preoperative MRI images,clinical data and pathological results of 254 breast cancer patients(99 positive and 155 negative)were retrospectively collected.First,the region of interest(ROI)of the primary breast cancer lesions was manually delineated on the DCE and DWI sequences.Then,using FAE software,1316 features were extracted based on the ROI of DCE and DWI,and the features were standardized by Z-score transformation.254 breast cancer patients were randomly divided into training set and test set in a ratio of7:3,respectively.Due to the unbalanced composition ratio of positive and negative patients,in order to improve the model training performance,the data balance of the training set was carried out by synthetic minority oversampling technique(SMOTE).Univariate logistic regression analysis was used for all radiomics features,and features with P <0.05 were retained.In order to select the most valuable radiomics features for the prediction of lymph node metastasis,the LASSO was used to carry out feature convergenc.Finally,multivariate logistic regression was used to build models.The receiver operating characteristic curve(receiver operator characteristic curve,ROC)of DCE model,DWI model,radiomics model and radiomics-clinical combined model were drawn,and the discriminatory performance of each prediction model was evaluated by Delong test.Model fit was assessed using calibration curves and Hosmer-Lemeshaw test.The differences between the models were statistically compared by the comprehensive discriminant improvement Index(Integrated Discrimination Improvement,IDI)and quantified the clinical benefit between the different models.A nomogram of the radiomics-clinical combination model was drawn to visualize the individualized adjuvant clinical prediction of breast cancer lymph node metastasis.Results:1.For the pathological results,only Ki-67 was statistically different between the groups(P <0.05).The predictive model was constructed from the clinically independent risk factor Ki-67,and its diagnostic performance AUC was 0.596 and 0.522 in the training set and testing set.2.The AUC,accuracy,sensitivity,and specificity of the DCE model training set were 0.770,74.7%,60.9%,and 83.5%,respectively,and the AUC,accuracy,sensitivity,and specificity of the test set were 0.81,73.7%,73.3%,and 78.3%.The AUC of the DWI model in the training set was 0.789,and the accuracy,sensitivity,specificity were 76.7%,71%,and 80.7%,respectively.The AUC of the testing set was 0.761,with accuracy,sensitivity,specificity,73.7%,66.7%,and 82.6%,respectively.The AUC of the DCE and DWI combination radiomics in the training set was 0.84,and the accuracy,sensitivity and specificity were 81.5%,76.8% and 85.3%,respectively.The AUC of the testing set was0.864,with accuracy,sensitivity,specificity,78.9%,83.3%,and 76.1%,respectively.The AUC of the radiomics-clinical combination model was 0.854 in the training set,with accuracy,sensitivity,specificity,79.8%,89.9%,and 69.7%.The AUC of the testing set was 0.872,with accuracy,sensitivity,specificity,80.3%,76.73%,and 84.8%.3.Pairwise comparison of different models: The difference between the joint model and the single sequence model in the training set was statistically significant(DCE vs Radiomics:Training set:De Long test:P=0.003,DWI vs Radiomics:training set:De Long test:P=0.044),and there was no statistical difference in the testing set.The radiomics-clinical combination model compared with the single sequence model was statistically significant in both sets.(DCE vs Radiomics-clinical:training set:De Long test:P=0.002,testing set:De Long test:P=0.034.DWI vs Radiomics-clinical:Training set:De Long test:P=0.001.testing set:De Long test:P=0.41).The results show that the prediction performance of the MRI radiomics model and radiomics-clinical combination model are better than the single sequential prediction model,and the radiomics-clinical combined risk factor model is the highest.4.The DCE and DWI model IDI was 0.0213,and the P-value for comparing the two models was 0.59214.The IDI of the combined sequence model and the radiomics model was 0.0174,and the P value of the two models was 0.11674,with no statistical difference.The radiomics-clinical combined model and the single sequence model IDI was 0.14295,P <0.0001,and the two models were statistically different.The combined radiomincs model had a gain of 14.2% over the single sequence model.5.The radiomics-clinical combined model visualizes the predicted risk of clinical risk factors,and the calibration curve and HL test indicate good model fit.Decision curve analysis shows that the application of radiomics to predict lymph node metastasis in breast cancer has a better clinical benefit than a single sequence radiomincs model when the threshold probability is in the range of 20%-90%.Conclusion:1.The radiomics model based on DCE and DWI MRI showed good predictive performance for lymph node metastasis in breast cancer.2.The prediction performance of combined radiomics model and radiomics-clinical combined model is better than that of single sequence model、clinical prediction model.3.A nomogram combining clinical risk factors with radiomics scores has some clinical value in visual prediction of lymph node metastasis in breast cancer.