Differential Diagnosis Of Benign And Malignant Breast Lesions And Prediction Of Axillary Lymph Node Status In Breast Cancer Patients Based On DCE-MRI And Radiomic Analysis

The first part DCE-MRI for Differentiation between Granulomatous Lobular Mastitis and Ductal Carcinoma in SituObjective: Granulomatous lobular mastitis(GLM)is a rare benign breast disease and non-inflammatory GLM is often misdiagnosed as breast carcinoma.Ductal carcinoma in situ(DCIS)is a noninvasive malignancy characterized by a proliferation of malignant epithelial duct cells confined to mammary ducts without demonstrable evidence of invasion through the basement membrane into the surrounding stroma.According to the presence or absence of calcifications on mammography,DCIS can be classified as calcified or non-calcified subtypes.It is difficult to differentiate GLMs from non-calcified DCIS depending on mammographic findings because almost all GLMs are absent of calcifications.Moreover,ultrasound is helpless for the differential diagnosis.Dynamic contrast-enhanced MRI(DCE-MRI)demonstrates an excellent diagnostic performance to clarify equivocal breast findings detected on conventional imaging.Therefore,the purpose of this study is to investigate whether DCE-MRI is an effective modality for the differentiation between clinically non-inflammatory GLM and non-calcified DCIS and identify DCE-MRI characteristics contributing to the differential diagnosis.Methods: DCE-MRI characteristics of 33 clinically non-inflammatory GLM and 36 non-calcified DCIS were retrospectively analyzed in the study.Internal enhancement of non-mass enhancement(NME)lesion was divided into clustered enhanced ring(absence/presence),and then clustered enhanced ring(presence)was further classified as small and large ring based on the optimal cutoff value.The distribution of NME can be divided into two types: the distribution along the duct(linear/segmental)and the distribution not along the duct(non-linear/segmental).The 5th Breast Imaging and Data System(BI-RADS)MRI descriptors were used for assessing the other DCE-MRI characteristics.Univariate analyses were used to select variables with significant differences.Multivariate analysis was used to identify the independent predictors of the differential diagnosis.The discriminative abilities of different predictors and their combination were compared by area under the receiver operating characteristic(ROC)curves(AUCs).Results: Although NME is the main manifestation of both entities,NME was seen more commonly in clinically non-inflammatory GLM than in non-calcified DCIS(p=0.003).DCE-MRI characteristics with significant differences(p<0.05)in univariate analyses included NME size,clustered enhanced ring(absence/presence),ring size,initial increase and kinetic characteristics for the differentiation between these two entities presenting as NME lesion.Clustered enhanced ring(presence)was further classified as small(≤7mm)or large ring(>7mm)based on the maximum value of Youden index with AUC of 0.878(95%CI,0.734-0.961).Multivariate analysis revealed that internal enhancement and initial increase were identified as significant independent predictors.The AUCs of internal enhancement,initial increase and their combination were 0.825(95%CI,0.699-0.914),0.700(95%CI,0.561-0.816)and 0.867(95%CI,0.748-0.943),respectively.Conclusion: DCE-MRI can act as an effective non-invasive method to differentiate clinically non-inflammatory GLM from non-calcified DCIS.Large ring was more highly suggestive of clinically non-inflammatory GLM,while no or small ring was more highly predictive of non-calcified DCIS.The second part DCE-MRI and Radiomic Analysis for Differentiation between Pure Mucinous Breast Carcinomas and Fibroadenomas with Strong High-Signal Intensity on T2WIObjective: pure mucinous breast carcinomas(PMBCs)typically show strong high-signal intensity on T2 WI due to rich extracellular mucus.It is difficult to discriminate PMBCs from fibroadenomas(FAs)with strong high-signal intensity on T2WI(T2-SHi)for radiologists.However,accurate preoperative diagnosis is crucial because the lesions require very different therapeutic approaches.Although previous studies revealed that DCE-MRI aided in the differential diagnosis of the two entities,the results of multivariate analyses were expressed as odds ratios.Additionally,differentiation PMBCs from FAs with T2-SHi based on radiomic features has not been investigated.Therefore,the purpose of this study is to develop DCE-MRI nomogram and determine the application value of radiomic analysis for differentiating PMBCs from FAs with T2-SHi.Methods: DCE-MRI features of 64 PMBCs and 137 FAs with T2-SHi were analyzed retrospectively.The BI-RADS classification from the original report was recorded.DCE-MRI features with statistical difference in univariate analysis were included in multivariate logistic regression analysis to establish DCE-MRI model for the differential diagnosis and develop DCE-MRI nomogram that is easy to use in clinical practice.An open-source software named ITK-SNAP(Version 3.6.0)was used to manually segment the lesions on the final phase(the 8th phase)of DCE-MRI.GE Analysis Kit(AK)software was used to extract the radiomic features.All lesions were divided(7:3)randomly into the development(n=141)and validation cohorts(n=60).First,Mann Whitney U test was performed on all of the radiomic features of the lesions in the training cohort.Subsequently,based on the 116 features with statistical differences according to univariate analysis,the least absolute shrinkage and selection operator(LASSO)was used for feature selection and principal component analysis(PCA)was used to reduce the dimension of features,respectively.LASSO radiomic signature was developed based on the six features selected by LASSO and PCA radiomic signature was developed based on the nine rotated principal components both by multivariate Logistic regression method.The radiomic signatures established by the training cohort were tested by the validation cohort.AUC,sensitivity,specificity,accuracy,PPV and NPV were used to evaluate DCE-MRI model and two radiomic signatures.Delong test was used to compare AUCs of DCE-MRI model and two radiomic signatures.The calibration curves were drawn to show the consistency between the predictive value and the true value.Decision curve analysis(DCA)was conducted to determine and compare the clinical usefulness of DCE-MRI model and radiomic signatures by calculating the net benefits at different threshold probabilities.For all analyses,P < 0.05 was considered statistically significant and all tests were bilateral.Statistical analysis was performed with R software(version 3.6.1).Results: Sensitivity,specificity,accuracy,PPV and NPV calculated according to the BI-RADS classification from the original report were 76.56%,73.00%,74.13%,56.98% and 86.96%,respectively.Multivariate analysis showed that age,margin,delayed enhancement pattern,enhancing internal septation and extent of lobulation were independent predictors for differentiating PMBCs from FAs with T2-SHi.AUC,sensitivity,specificity,accuracy,PPV and NPV of DCE-MRI model were 96.24%,87.50%,94.89%,92.54%,88.89% and 94.20%,respectively.In the training and validation cohorts,LASSO radiomic signature showed 94.40% and 98.63% of AUCs and PCA radiomic signature showed 94.49% and 97.95% of AUCs.Pairwise comparisons among DCE-MRI model and LASSO,PCA radiomic signatures were done with no significant difference(p>0.05).Compared with LASSO and PCA radiomic signatures,DCE-MRI model showed better consistency between predicted value and true value and greater clinical net benefit.Conclusion: DCE-MRI model was superior to the BI-RADS classification from the original report and radiomic signatures in the differential diagnosis of PMBCs and FAs with T2-SHi.Nomogram made DCE-MRI model act as an easy-to-use and understand tool for doctors and improved the accuracy of differential diagnosis,so as to provide reliable objective basis for individualized clinical decision-making without additional cost.The third part DCE-MRI and Radiomic Analysis for Preoperative Prediction of Axillary Lymph Node Status in Clinically Node-negative Early Breast Cancer PatientsObjective: Sentinel lymph node biopsy(SLNB)has become a standard procedure for axillary lymph node(ALN)staging and treatment in clinically node-negative early breast cancer patients.In the breast cancer patients with clinically negative ALN,metastases are detected in 33.2%-39% SLNs,which means that approximately 60%–70% patients have no direct benefit from SLNB.The prior studies demonstrated that the predictive ability of radiomic features based on MRI deriving from primary breast cancer lesions was generally better than that of model according to clinicopathological features for predicting axillary lymph node metastasis(ALNM).However,to our best knowledge,there is no report on the prediction of ALNM based on MRI radiomics for the patients with clinically negative ALN.Therefore,the purpose of this study is to develop and validate a model based on the DCE-MRI radiomics deriving from primary breast cancer lesions to predict ALN status for clinically node-negative early breast cancer patients,in order to avoid SLNB for patients with low risk of ALNM.Methods: The clinicopathological and DCE-MRI data of 260 invasive breast cancer patients with T1-2 and clinically negative ALN were retrospectively analyzed.Among 260 patients in this study,71 patients had positive ALN and 189 patients had negative ALN.A total of 851 radiomic features based on primary breast cancer that could be divided into four categories were extracted from T1 WI,T2WI and the third postcontrast phase of DCE sequence(CE3)using the open-source software named Pyradiomics.All lesions were divided(7:3)randomly into the development(n=182)and validation cohorts(n=78).Radiomic features of T1 WI,T2WI and CE3 were selected separately in the training cohort.For feature selection of combinations of three sequences,all features were combined first and selected then.Multivariate Logistic regression method was used to develop radiomic signatures of T1 WI,T2WI,CE3 and multiparametric MRIrespectively,and the performance of each signature was validated in the validation cohort.ROC curve,calibration curve and decision curve analysis were used to assess the radiomic signatures.For all analyses,P < 0.05 was considered statistically significant and all tests were bilateral.Statistical analysis was performed with R software(version 3.6.1).Results: There was no statistically significant difference in clinicopathological and DCE-MRI features between ALNM positive and negative groups(p>0.05).In the training cohort,AUC of multiparametric radiomic signature was 0.78,which was higher than that of T1 WI,T2WI and CE3 radiomic signature with significant statistical difference(p<0.05);in the validation cohort,although there was no significant difference between multiparametric radiomic signature and each individual T1 WI,T2WI and CE3 radiomic signature(p>0.05),AUC of 0.75 for multiparametric radiomic signature was still the highest numerically among all radiomic signatures.According to the calibration curve,the predicted and true values of multiparametric radiomic signature showed good consistency.Based on decision curve analysis,the net benefit of using multiparametric radiomic signature to predict ALN status was greater than treat-all or treat-none scheme when the threshold probability ranged from 0.08 to 0.92.Conclusion: Multiparametric radiomic signature was superior to radiomic signatures based on separate T1 WI,T2WI and CE3.Multiparametric radiomic signature has the potential to act as an effective tool to predict ALN status for patients with clinically node-negative early breast cancer,which can aid in individualized decision-making to avoid overtreatment for patients with low risk of ALNM.