DCE-MRI And Radiomics In Predicting Tumor Response Of Neoadjuvant Therapy In Breast Cancer

ObjectiveTo explore whether the pretreatment DCE-MRI and radiomics signatures were associated with pathologic complete response(pCR)to Neoadjuvant therapy(NAT)in breast cancer.MethodsA retrospective review of 73 patients with breast invasive carcinomas proved by biopsy between June 2015 and October 2018.All the patients underwent breast MRI before NAT in SIEMENS Verio 3.0T MR imaging system.The postoperative pathological results after NAT were divided into two groups: pCR(G5)and non-pCR(G1-G4)according to the Miller&Payne histological grading system.The pretreatment clinical and pathological data were analyzed,including age,menstrual status,histological grade,the expression status of estrogen receptor(ER),progestogen receptor(PR),HER-2 and Ki-67.Then MRI features were also collected,including maximum diameter,enhancement pattern,morphology,margin,internal enhancement characteristic,TIC curve,and whether the axillary lymph node and skin was involved.Two radiologists segmented lesions manually in DCE-MRI images and extracted the radiomic features,then using LASSO Cox regression to reduce dimensions of the features.All the patients were randomly divided into training set and validation set according to 8:2,three supervised learning algorithms LR,RF and KNN were used to classify the features.The prediction performance of model in training set was assessed using receiver operating curve(ROC),then using maximal Youden index to determine the optimal critical value.The area under the ROC curve(AUC),sensitivity,specificity,precision and F1 score of training set and validation set were calculated to evaluate the effects of the radiomics model.Results1.There was no significant difference in age,menstrual status,histological grade,the expression status of HER-2 and Ki-67 between pCR and non-pCR group(P>0.05).The number of patients with ER(+)/PR(+)in pCR group were significantly less than that in non-pCR group,and there were statistically significant differences between these two group(P<0.05).In pCR group,the largest proportion is triple negative type(36.4%),while in non-pCR group,the largest proportion is type Luminal B(58.8%).The molecular subtype distribution show significant difference between two groups(P=0.016).2.The pretreatment MR image showed 90.9%(20/22)lesions in pCR group presented as mass and the rest 9.1%(2/22)presented as non-mass enhancement.While in non-pCR group,the proportion of the mass was 49.0%(25/51),and the proportion of non-mass enhancement was 51.0%(26/51).The difference between two groups was statistically significant(P<0.001).There was no statistically significant difference in maximum mass diameter,morphology,margin,internal enhancement characteristic,TIC curve,or whether the axillary lymph node and skin were involved between two groups(P>0.05).3.Radiomics analysis: LASSO Cox regression selected 4 radiomic features which best relevant with pCR.When training with LR classifier,the AUC value of the training set was 0.871(sensitivity was 0.76,specificity was 0.68),the AUC value of the validation set was 0.745(sensitivity was 0.80,specificity was 0.73).When training with RF classifier,the AUC value of the training set was 1.000(sensitivity was 1.00,specificity was 0.95),the AUC value of the validation set was 0.909(sensitivity was 0.80,specificity was 0.82).When training with KNN classifier,the AUC value of the training set was 0.821(sensitivity was 0.71,specificity was 0.75),the AUC value of the validation set was 0.791(sensitivity was 0.80,specificity was 0.73).Conclusion1.The immunohistochemistry presented ER(-)/ PR(-)or molecular subtype presented HER-2(+)/TNBC in breast cancer is more likely to achieve pCR.2.In pretreatment breast MRI features,only “mass” can be used as a predictor of NAT response,which is more likely to achieve pCR.3.The radiomic features from pretreatment DCE-MRI could be the predictor of NAT response.The prediction model based on radiomic features could predict pCR.Compared with the traditional MRI features,the model based on radiomic features had the best performance in evaluation of the response to NAT.