| Objective: To establish a comprehensive model to predict whether the patients with breast cancer will achieve pathologic complete response(pCR)after neoadjuvant chemotherapy(NAC)based on pretreatment MRI and biopsy whole slide image(WSI),and the comprehensive model was verified using data from two institutions.Methods: In total,this study retrospectively collected 331 patients with pathologically confirmed invasive breast cancer from January 2016 to December 2021 in two institution(Institution Ⅰ: n=259;Institution Ⅱ: n=72)who underwent NAC.Ultrasound-guided needle biopsy and dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI)were performed within 1 month before NAC.According to Miller-Payne(MP)grading of surgical specimens,the patients were divided into the pCR group and non-pCR group.Radiological features from original lesions in MRI and clinicopathologic data of all patients were analyzed.The independent clinical predictive factors for pCR to NAC in patients with breast cancer were obtained after selection by univariate and multivariate logistic analysis,which established the clinical model.Manually delineated the primary focus of breast cancer as the region of interest(ROI)on the sequence images of first postcontrast images of the first phase of enhanced MRI,DWI and ADC mapping.Radiomics features were extracted based on ROI.Radiomics features were selected through Pearson product-moment correlation coefficient(Pearson’s r)and the least absolute shrinkage and selection operator(LASSO)regression analysis,then radiomics prediction model were established.Pathological features from all WSIs of patients extracted through deep learning established the deep learning pathological model(DLPM),which was combined with radiomics and clinical features to build a comprehensive model and draw a nomogram.259 patients(78 in the pCR group and 181 in the non-pCR group)included in Institution Ⅰ were randomly divided into training(n =213)and testing cohort(n = 92)in a ratio of 7:3.72 patients included in Institution Ⅱ(18in the pCR group and 54 in the non-pCR group)were used as an independent external validation cohort.The performance of each model was evaluated by area under the curves(AUC)of the receiver operating characteristic curves(ROC),accuracy,sensitivity and specificity.Calibration curves and decision curve analysis(DCA)were used to evaluate nomogram’s value of the clinical application.The Delong test was applied to compare the AUC values of each model.Results: Univariate and multivariate logistic analysis screened 5 clinical features to build clinical model.Radscore were made up by 19 radiomic features through the LASSO method.Ten deep learning features were selected through Alex Net to establish pathological deep learning model.In addition,using univariate analysis and multivariate logistic analysis screen five clinical features including ER and PR expression status,HER-2 expression status,time signal intensity curve(TIC)pattern,average ADC value and breast gland density as independent clinical predictors of pCR and 3 deep learning features,which combined with Radscore to establish comprehensive model.Hence,a total of 4 models were established to predict the pCR of patients with breast cancer after NAC.The results showed that nomogram had good prediction efficiency with the AUC values of 0.95,0.84 and 0.83,respectively in the training,test and external validation cohorts.Delong test showed that the difference of AUC value between the comprehensive model and other models was statistically significant(p < 0.05).DCA showed that nomogram obtained the highest clinical benefit among 4 models.Conclusions: The comprehensive nomogram model based on multiparametric MRI radiomics,clinical features and puncture WSI deep learning features can predict the pCR to NAC in breast cancer patients.The integrated model of radiomics-clinical featuresdeep learning pathological features perform better than that of the separate radiomics,deep learning and clinical features model. |
PDF Full Text Request