Correlation Between Radiomics In Different Regions And Invasiveness And Recurrence Risk Of Invasive Breast Cancers Based On DCE-MRI

Objective:Our study aimed to assess the utility of radiomics models constructed with radiomics features from the intratumoral,peritumoral,transition zone,and intratumoral combined peritumoral regions based on dynamic enhanced magnetic resonance images(DCE-MRI)in predicting the aggressiveness and recurrence risk of breast cancer.Methods:A total of 389 mass-like invasive breast cancer patients confirmed by postoperative pathology in our hospital from September 2018 to June 2022 were retrospectively included.The intratumoral volume of interest(VOI)of breast cancer lesion was delineated by ITK-SNAP software.The peritumoral region was expanded 4 mm outward along the tumor edge using the expansion algorithm,and the transition zone was expanded 4 mm outside the tumor and within the tumor along the edge of the VOI using the expansion and erosion algorithms,and the intratumoral combined peritumoral region included the intratumoral and peritumoral 4 mm regions.1967 radiomics features of intratumoral,peritumoral,transition zone,and intratumoral combined peritumoral regions were sequentially extracted.Then,feature dimensionality reduction was performed.Firstly,the Mann-Whitney U test was used to test the differences of 1967 features,and then the least absolute shrinkage and selection operator(LASSO)and Maximum correlation minimum redundancy(mRMR)were sequentially used to obtain 30 features.Finally,the Akaike information criterion(AIC)was used for feature optimization.Based on the selected radiomics features from the intratumoral,peritumoral,transition zone,and combined intratumoral and peritumoral regions,five breast cancer aggressiveness indicators and one recurrence risk prediction tasks were established:(1)axillary lymph node(ALN)metastasis vs non-metastasis;(2)low histological grade(Grade Ⅰ/Ⅱ)vs high(Grade Ⅲ);(3)Human epidermal growth factor receptor 2(HER2)positive vs negative;(4)Hormone receptor(HR)positive vs negative;(5)Ki-67 high expression(>20%)vs low expression(≤20%);(6)High recurrence risk vs non-high recurrence risk.The radiomics models of the intratumoral,peritumoral,transition zone,and intratumoral combined peritumoral regions were established for each prediction task using logistic regression classifier.The area under the receiver operating characteristic curve(AUC),accuracy,specificity(SPE),sensitivity(SEN)and decision curve analysis(DCA)were used to evaluate each radiomics model.Delong’s test p-value was used to compare performance differences among different models.Results:For breast cancer aggressiveness indicators,the peritumoral region models showed the highest AUC for the prediction tasks of ALN,histological grading,HR and Ki67 with AUCs of 0.730/0.725,0.774/0.742,0.771/0.745 and 0.766/0.736 for the training and validation cohorts,respectively.The intratumoral region models in predicting HER2 showed higher diagnostic performance with AUC values of 0.744 and 0.708 in the training and validation cohorts,respectively.For the prediction of recurrence risk,the AUC value of 0.743/0.732 for the peritumoral region model was higher than that of 0.720/0.703 in the intratumoral region model,0.735/0.714 in the transitional region model,and 0.732/0.703 in the intratumoral combined peritumoral region model.In the prediction tasks of five aggressiveness indicators and one recurrence risk of breast cancer,the diagnostic performance(AUC value)of different regional models did not show significant differences.Conclusion:Radiomics based on different regions of DCE-MRI is valuable in noninvasive prediction of both invasiveness and recurrence risk of breast cancer.In the prediction tasks of ALN,histological grade,HR,Ki-67 and recurrence risk,the diagnostic efficiency(AUC value)of the peritumoral region model was higher than that of the intratumoral,transitional region and intratumoral combined peritumoral region models;in the prediction task of HER2,the diagnostic efficiency(AUC value)of the intratumoral region model was higher than that of the peritumoral,transition zone and intratumoral combined peritumoral region models.There was no statistical difference between the diagnostic performance(AUC values)of the different regional models in any of the prediction tasks.It is recommended to select personalized tumor regions of interest in a targeted manner according to different study purposes without loss of diagnostic efficiency.