Differential Diagnosis Of Benign And Malignant Breast Lesions And Prediction Of HER2 2+ Status And TILs Levels Of Breast Cancer On MRI Radiomics

Part Ⅰ.Prediction of benign and malignant phyllodes tumors of breast based on DCE-MRI texture analysisBackground&Objective:Phyllodes tumors（PTs）are fibroepithelial tumors that are histologically similar to fibroadenomas.PTs are subcategorized into benign,borderline,or malignant,with borderline/malignant PTs having a higher local recurrence rate.Histological grading of PTs is usually associated with prognosis.Improving the preoperative differential diagnosis of benign and malignant PTs is of great significance for clinical treatment planning and prognostic assessment of patients.The purpose of this study was to explore the value of texture analysis（TA）based on dynamic contrast enhanced MRI（DCE-MRI）in the differential diagnosis of benign PTs（BPTs）and borderline/malignant PTs（BMPTs）.Methods:The clinical and MRI data of 47 patients with histologically proven PTs,including 26 benign BPTs and 21 BMPTs,were enrolled in this retrospective study.Two senior radiologists evaluated the MRI findings of all tumors according to the Breast Imaging Reporting and Data System（BI-RADS）without knowing the pathological results,and the disagreements were resolved by consultation.The seventh post-contrast phase on DCE-MRI(DCE_phase7of all patients were selected,and MaZda 4.6 software was used to manually delineate the volume of interest（VOI）of the tumor layer by layer and extract a total of 314 texture features.Then,the variance threshold method,univariate analysis and univariate Logistic regression analysis were used to reduce the dimension of features.The differences of MRI imaging features and optimal texture features between BPTs and BMPTs were compared respectively.The variables with statistically significant differences were further included in multivariate Logistic regression analysis,and the imaging model,texture analysis model and the combined model of the two approaches were constructed respectively.Receiver operating characteristic（ROC）curve was used to evaluate the efficacy of different models in differentiating BPTs from BMPTs,and the area under curve（AUC）,sensitivity and specificity were calculated.Delong test was used to compare AUC differences among different models.Results:In terms of texture features,histogram features of Perc.90%,Mean and Variance;gray level co-occurrence matrix（GLCM）features of S（0,0,1）AngScMom and S（1,0,0）AngScMom;and gray run length matrix（RLM）features of Z_GLevNonU,45dgr_GLevNonU and 135dr_GLevNonU were significantly different between BPTs and BMPTs（all p<0.05）.Regarding MRI findings,however,only cystic wall morphology showed significant differences between the two groups（p<0.05）.The areas under the ROC curve（AUCs）of texture analysis model,imaging model,and the combined model were 0.886（95%CI,0.760-0.960）,0.687（95%CI,0.518-0.825）and 0.894（95%CI,0.754-0.970）,respectively;the sensitivity was 85.7%,61.1%and 94.4%,respectively;and the specificity was 80.8%,76.2%,and 76.2%,respectively.The Delong test showed that both the AUC of texture analysis model and the combined model were significantly higher than that of imaging model（all p<0.05）.The AUC of the combined model was higher than that of the texture analysis model,however,the difference was not statistically significant（p>0.05）.Conclusion:DCE-MRI TA has potential in differentiating BPTs from BMPTs,and its diagnostic efficiency was higher than that of conventional imaging analysis.Furthermore,the combination of these two approaches can further improve the diagnostic performance.Part Ⅱ.UE-MRI-based radiomics in the differential diagnosis of benign and malignant breast lesions with high T2WI signal intensityBackground&Objective:Various histopathologic components in benign and malignant breast lesions could increase T2 relaxation time and may generate hyperintense signal at T2-weighted imaging（T2WI）,including extensive necrosis,a cystic or microcystic component,an adipose or sebaceous component,duct ectasia,mucous secretion,interstitial mucous degeneration,stromal edema,and hemorrhagic changes.Accurate identification of benign and malignant breast lesions with high T2WI signal intensity before surgery is of great significance for guiding clinical treatment and evaluating prognosis.The purpose of this study was to explore the value of radiomics based on unenhanced MRI（UE-MRI）in the differential diagnosis of benign and malignant breast lesions with high T2WI signal intensity.Methods:The clinical and MRI data of 170 patients with breast lesions showing high T2WI signal intensity,were retrospectively analyzed,including 86 benign lesions and 84 malignant lesions confirmed by histopathology.All patients were randomly divided into training cohort（n=119）and validation cohort（n=51）according to a ratio of 7:3.3D-Slicer software was used to manually delineate the VOIs of the lesion on T2WI,diffusion-weighted（DWI）and apparent diffusion coefficient（ADC）,respectively.After extracting 1218 radiomics features from each sequence,features with inter-/intra-class correlation coefficients（ICCs）greater than 0.75 were retained.The Select k Best,max relevance min redundancy（mRMR）and gradient boosting decision tree（GBDT）algorithm were adopted to find the optimal features in the training cohort.Support vector machines（SVM）classifier was used to construct T2WI,DWI,ADC single-sequence-based radiomics models and multiparametric MRI（T2WI+DWI+ADC）model for differentiating benign and malignant breast lesions with high T2WI signal intensity.The radiomics models constructed in the training cohort were tested in the validation cohort.In addition,in the validation cohort,the results of subjective diagnosis based on UE-MRI and DCE-MRI images by two radiologists with different work experience were recorded.RadiAnt DICOM Viewer software was used to measure the overall ADC value and solid ADC value at the maximum slice of the lesion on the ADC map,and the differences of the two variables in benign and malignant lesions were compared.The performance of the optimal radiomics model,radiologist subjective diagnosis and ADC value were evaluated by receiver operating characteristic（ROC）curve,respectively.The Delong test was used to compare the AUCs between the optimal radiomics model and radiologist subjective diagnosis and ADC value,respectively.Results:In the single-sequence-based model,the AUC of the ADC model in the training and validation cohorts were 0.85 and 0.84,respectively,which were higher than DWI model（training cohort,0.77;validation cohort,0.72）and T2WI model（training cohort,0.68;validation cohort,0.63）;the AUC differences between ADC model and T2WI model was statistically significant in both training and validation cohorts（all p<0.05）.The model based on multi-sequence combination（T2WI+DWI+ADC）had the highest diagnostic efficiency,the AUC was 0.89 and 0.87 in the training and the validation cohorts,respectively.In the validation cohort,the AUCs of subjective diagnosis from the radiologist with 3-and 10-year experiences based on UE-MRI were 0.61 and 0.77,respectively,and those based on DCE-MRI were 0.71 and 0.88,respectively.The difference of ADC values of solid regions between benign lesions(1.39×10^-3mm²/s)and malignant lesions(1.15×10^-3 mm²/s)was statistically significant（p=0.008）.The best differential diagnostic efficacy was achieved when the ADC threshold was 1.248×10^-3 mm²/s,with an AUC of 0.74.The Delong test showed that AUC of the combined model were significantly higher than that of subjective diagnosis from the radiologist with 3-year experience based on UE-MRI and DCE-MRI,respectively（0.87 vs 0.61,p=0.001;0.87 vs 0.71,p=0.032）,but slightly lower than that of subjective diagnosis from the radiologist with 10-year experience based on DCE-MRI,which showed no statistical significance（0.87 vs 0.88,p>0.05）.The AUC of the combined model was higher than that of ADC,which showed no statistical significance（0.87 vs 0.74,p>0.05）.Conclusion:UE-MRI Radiomics has the potential to distinguish benign and malignant breast lesions with high T2WI signal intensity.The efficacy of multi-sequence（T2WI+DWI+ADC）based radiomics model was close to the subjective diagnosis of radiologist with 10-year experience based on DCE-MRI,and significantly higher than that of radiologist with 3-year experience based on UE-MRI and DCE-MRI,respectively.Part Ⅲ.Multi-parametric MRI-based radiomics in predicting the HER2 2+status of breast cancerBackground&Objective:Human epidermal growth factor receptor2（HER2）is an important biomarker for judging the molecular subtype of breast cancer（BC）,and its expression can usually be determined by immunohistochemistry（IHC）.Tumors with IHC scores of 1+or 0 was defined as HER2-negative,whereas IHC scores of 3+was defined as HER2-positive.Compared with HER2-negative breast cancer,HER2-positive breast cancer has a higher malignant biological behavior,however,some patients with HER2-positive breast cancer can benefit from anti-HER2-targeted therapy to improve survival.HER2 2+is a critical state and fluorescence in situ hybridization（FISH）testing is required to determine HER2 2+positive（HER2 gene amplification）or negative（HER2 gene amplification）status.However,FISH testing is expensive,time-consuming,and requires specialized equipment and technology.Therefore,identifying a cost-and time-effective alternative method for distinguishing HER2 2+status would be beneficial.The purpose of this study was to investigate the feasibility of multiparametric magnetic resonance imaging（MP-MRI）based radiomics in predicting the HER2 2+status in BC patients.Methods:The clinical and MRI data of 107 patients with HER2 2+breast cancer confirmed by histopathology were retrospectively analyzed,including 42 HER2 2+FISH positive and 65 HER2 2+FISH negative cases.The patients were divided into training group（n=74）and validation group（n=33）according to the ratio of 7:3.Univariate and multivariate Logistic regression analysis were used to analyze the conventional clinical-imaging features,and the clinical-imaging independent predictors with significant predictive value between BER2 2+positive and negative BCs were selected to construct the clinical-imaging model.The 3D-Slicer software was used to delineate the tumor VOIs on T2WI,DWI and the third and seventh post-contrast phases on DCE-MRI(DCE_{phase3 and DCEpase7},respectively,then,the radiomics features from different sequence were extracted.The training set were used to select the radiomics features of single sequence(T2WI,DWI,DCE₍phase3 and DCE_phase7and multi-sequence combination(T2WI+DWI,T2WI+DWI+DCE₍phase3 and T2WI+DWI+DCE_phase7,respectively.In the training group,the features from single-parametric MRI(T2WI,DWI,DCE₍phase3 and DCE_phase7and multi-parametric（MP）MRI(T2WI+DWI,T2WI+DWI+DCE₍phase3 and T2WI+DWI+DCE_phase7were selected.The multivariate Logistic regression classifier based on 5-fold cross-validation was used to construct predictive model,including single-parametric MRI model,multi-parametric MRI model and the combined model of multi-parametric MRI model and clinical-imaging model,to distinguish HER2 2+positive from HER2 2+negative BC.The performance of each model constructed by the training group was tested in the verification group.ROC curve was used to evaluate the predictive efficiency of each model.Delong test was used to compare AUC differences among different models.Result:Among the clinical-imaging features,Ki-67 status,tumor margin and peritumoral edema were statistically different between HER2 2+positive and negative BC（all p<0.05）.The AUC of the clinical-imaging model was 0.733 and 0.738 in the training and validation groups,respectively.In the single-parametric MRI model,the DCE_phase7 model had the highest efficiency,with AUCs of 0.874 and 0.835 in the training and validation groups,respectively,while the T2WI model had the lowest predictive efficiency,and the AUCs in the training and validation group were 0.740 and 0.573,respectively.The MP-MRI model constructed by the combination of T2WI,DWI and DCE-MRI sequences was better than that of any single-parametric MRI model.The AUC of MP-MRI model A（T2WI+DWI）,MP-MRI model B(T2WI+DWI+DCE_phase3)and MP-MRI model C(T2WI+DWI+DCE_phase7)in the training group were 0.903,0.930 and 0.931,respectively,and 0.850,0.869 and 0.873 in the validation group,respectively.In the combined model,the combined model C（MP-MRI model C+clinical-imaging model）had the highest prediction efficiency,and the AUCs in the training and validation group were 0.952 and 0.892,respectively.The Delong test showed that the AUC of each MP-MRI model or combined model was higher than that of the clinical-imaging model,and the AUC differences was statistically significant in the training group（all p<0.05）,but not in the validation group（all p>0.05）.Conclusion:The MP-MRI radiomics model has potential value in predicting HER2 2+status of BC,which showed higher predictive efficiency than that of single-parametric MRI model and conventional clinical-imaging model.Part Ⅳ.DCE-MRI radiomics nomogram in predicting CD8+TILs levels in HER2-positive breast cancerBackground&Objective:Tumor infiltrating lymphocytes（TIls）is an important part of tumor immune microenvironment,in which CD8+TILs play a key role in the host anti-tumor immune response.In patients with HER2-positive BC,high levels of CD8+TILs were strongly associated with better treatment response and longer survival.Therefore,accurate preoperative assessment of CD8+TILs level could be useful to guide immune therapy and predict the prognosis.The purpose of this study was to investigate the potential value of DCE-MRI radiomics nomogram in predicting CD8+TILs level in patients with HER2-positive BC.Methods:The clinical and MRI data of 126 patients with HER2-positive BC confirmed by postoperative pathology were retrospectively analyzed.All patients were randomly divided into training group（n=88）and validation group（n=48）according to a ratio of 7:3.According to the median level of CD8+TILs,all patients were divided into CD8+TILs high level group and CD8+TILs low level group.3D-Slicer software was used to manually delineate the VOIs of the lesion on the third post-contrast phases from DCE-MRI.Select k Best,mRMR,and least absolute shrinkage and selection operator（LASSO）algorithm were used for feature selection in the training group,and Rad-score was calculated to establish the radiomics model.Univariate and multivariate Logistic regression analysis were used to analyze the clinical-imaging characteristics,and the clinical-imaging independent influencing factors with significant predictive value were selected to construct the clinical-imaging model.Then,combined with Rad-score and clinical-imaging independent influencing factors,the combined model was constructed and the nomogram was established.Model validation was performed in the validation group.The Hosmer-Lemeshow test and calibration curve were used to evaluate the goodness of fit of the models.ROC curve was used to evaluate the performance of different models.The Delong test was used to evaluate the difference of AUCs between different models.The decision curve analysis（DCA）was used to evaluate the clinical application value of different models.Results:Univariate Logistic regression analysis showed that T stage,N stage,MRI maximum diameter,mass margin,internal enhancement pattern was statistically different between CD8+TILs high-level group and low-level group（all p<0.05）.Multivariate Logistic regression analysis further showed that the tumor margin and internal enhancement patterns were independent clinical-imaging predictors（all p<0.05）.The AUC,sensitivity and specificity of the clinical-imaging model were 0.785（95%CI:0.690-0.881）,76.2%and 71.7%in the training group,and 0.803（0.654-0.951）,77.8%and 80.0%in the validation group,respectively.In the training group,the Rad-scores of CD8+TILs high-level and low-level groups were 0.563±0.112 and 0.399±0.114,respectively,and the difference was statistically significant（p<0.001）.In the validation group,the Rad-scores of CD8+TILs high-level and low-level groups were 0.634 ± 0.147 and 0.438 ± 0.135,respectively,and the difference was statistically significant（p<0.001）.The AUC,sensitivity and specificity of the radiomics model in the training group were 0.853（95%CI:0.771-0.935）,85.7%and 82.6%,respectively,and the corresponding values in the validation group were 0.822（95%CI:0.686-0.958）,83.3%and 80.0%,respectively.The nomogram model had the highest predictive efficacy.The AUC value,sensitivity and specificity in the training group were 0.866（95%CI:0.792-0.941）,83.3%and 78.3%,respectively,and the corresponding values in the validation group were 0.886（95%CI:0.778-0.994）,77.8%and 90.0%,respectively.Delong test showed that in the training group,the AUC of the nomogram model was higher than that of the clinical-imaging model,and the difference was statistically significant（p<0.05）;In the validation group,there was no statistical significance in the AUC of pairwise comparison among the three models（all p>0.05）.Hosmer-Lemeshow test showed that the nomogram model in the training and validation groups had good goodness of fit（all p>0.05）.The calibration curve showed that the probability value of the high level of CD8+TILs predicted by the nomogram model was in good agreement with the actual value.The DCA curve showed that the nomogram model had a higher overall net benefit than that of clinical-imaging model and radiomics model in predicting the high level of CD8+TILs.Conclusion:DCE-MRI radiomics has certain application value in predicting the level of CD8+TILs in HER2-positive BC.A higher Rad-score predicts higher levels of CD8+TILs.The nomogram model based on DCE-MRI radiomics features and conventional clinical-imaging features showed the highest predictive performance for CD8+TILs level in HER2-positive BC.