The Value Of A Combined Model Based On Mammography And MRI To Predict Axillary Lymph Node Metastasis In Breast Cancer

Objective: To build a combined model using features extracted from mammography and MRI for preoperative noninvasive prediction of axillary lymph node metastasis(ALNM),and validated using data from both institutions.Methods: A total of 492 women with primary breast cancer were enrolled(Institution 1: n= 420;Institution 2: n = 72).A retrospective study was conducted according to certain inclusion and exclusion criteria,and patients were divided into training cohort,validation cohort,internal test cohort and external test cohort by time and institution.Clinicopathological features and features described in the report of dynamic contrast enhanced magnetic resonance imaging(DCE-MRI)were collected from all patients.Obtain the craniocaudal(CC),mediolateral oblique(MLO)and diffusion weighted imaging(DWI)in MRI of all patients and enhance the first sequence image.A total of 330 patients with primary breast cancer from August 2013 to August 2020 in institution 1 were randomly divided into training cohort(n = 231)and validation cohort(n = 99)at a ratio of 7:3.The clinicopathological features and features described in MRI reports in the training cohort were analyzed by univariate analysis.Region of Interest(ROI)was delineated by 3D Slicer software,features were extracted by Pyradiomics software,and statistical analysis was performed by R 3.6.3 software.A Radiomics score(Radscore)is calculated by selecting features using the least absolute shrinkage and selection operator(LASSO)regression method.Multivariate logistic regression was used to analyze the statistically significant differences in the training cohort and Radscore.Clinical model,radiomics model and combined model were established respectively.The combined model was represented by a nomogram,the effectiveness of the nomogram was verified by correction curve,and the validation cohort was used to verify.The function of the model was evaluated by receiver operating characteristic curve(ROC),and the area under curve(AUC)was compared with the clinical model and the radiomics model.The p value of integrated discrimination improvement(IDI)and De Long method were used to compare the area under the curve values(AUC)of clinical models and combined models.The application value of decision curve analysis(DCA)model in clinic.Results: Univariate analysis found that tumor histological type,tumor maximum diameter,lesion shape,lesion edge,ADC value and MRI reported ALN status in the training cohort were associated with ALNM in breast cancer patients.Multivariate logistic regression showed that MRI reported ALN status and Radscore were independent predictors of ALNM.Finally,three models for preoperative noninvasive prediction of ALNM in breast cancer were established,including clinical model,radiomics model and combined model.The clinical model included a feature of ALN status reported by MRI,and the results showed that there was no significant difference in the probability of suspected positive lymph nodes in MRI among the 4 samples(P = 0.662).The AUC of diagnosis was listed as 0.786 in the training cohort,0.811 in the validation cohort,0.780 in the internal test cohort,and 0.837 in the external test cohort.The radiomics model based on mammography and DCE-MRI contained a total of 24 features with a diagnostic threshold of 0.469.The AUC of Radscore for ALNM was 0.846,0.762,0.899 and 0.793 in the training cohort,validation cohort,internal test cohort and external test cohort,respectively.Calibration curves show that the combined model has good calibration and is better than the radiomics model(area under the curve in the training cohort [AUC]: 0.886 vs.0.846;0.826 vs.0.762 in the validation queue;0.925 vs.0.899 in the internal cohort;the external cohort was 0.902 vs.0.793).IDI results showed that the diagnostic efficacy of the combined model was significantly improved in all cohorts except the validation cohort,and the AUC of the combined model and the clinical model was ranked 0.886 and 0.786 in the training cohort(IDI = 0.125,95%CI: 0.081-0.170,P < 0.001),and the validation cohort was 0.826 and 0.811(IDI =0.012,95%CI:-0.008-0.032,P = 0.241),0.925 and 0.780 for internal test cohort(IDI =0.271,95%CI: 0.177-0.365,P < 0.001),0.902 and 0.837 in the external test cohort(IDI =0.070,95%CI: 0.006-0.135,P = 0.033).Conclusion: MRI reported ALN status and Radscore calculated by radiomics were independent predictors of ALNM of breast cancer.The value of combined model based on mammography and MRI for preoperative noninvasive prediction of breast cancer ALNM was superior to that of clinical model and radiomics model.It was found that combined model has good stability after multi-center data verification.