Early Preoperative Prediction Of Pathological Response In Breast Cancer Neoadjuvant Chemotherapy Based On Mri Radiomics Analysis

Objective This study was to investigate the contrast-enhanced MRI(CE-MRI)radiomics analysis in prediction of pathological complete response(pCR)to neoadjuvant chemotherapy(NAC)in breast cancerMethods Fifty-five female subjects with biopsy-confirmed locally advanced breast cancer(aged 25 to 70 years old)who underwent CE-MRI scan prior to NAC were collected retrospectively.According to pathological evaluation after NAC,patients were categorized into 17 pCR patients and 38 non-pCR patients.Texture analysis was performed using pre-NAC MRI on 3DQI platform(https://www.3dqi.harvard.edu).A set of 781 volumetric texture features(including shape,histogram,gray level co-occurrence matrix,grayscale run matrix,grayscale region)was extracted in CE-MRI in the segmented volume of each lesion,the periphery of each segmented lesion,and the wavelet-decompsoed domains of each segmented lesion.Random Forest(RF)was investigated to the prediction of pCR to NAC using the extracted CE-MRI textures.Feature selection was performed by Boruta algorithm.A 10-fold cross-validation method was applied to validate the RF performance by receiver operating characteristic(ROC)curves analysis.We compared the RF prediction performances for all 55 breast cancer patients,34 patients with mass enhanced type breast cancer,and 21 patients with non-mass enhancement breast cancer,respectively.Results For all 55 breast cancer patients,five important textural features(LHH_RL_srlgle,HHH_RL_sre,HHH_RL_srlgle,LLH_GLCM_infoCorr1,HLL_RL_gln)were selected for the prediction model with the area under curve(AUC)of ROC 0.737(79.6% sensitivity and54.5% specificity).For 34 mass enhanced and 21 non-mass enhanced breast cancer patients,RF models selected three textural features(HHL_RL_lgre,LLL_RL_srhgle,LLH_RL_hgre)and four textural features(HHH_GLCM_infoCorr1,LLL_RL_sre,SHAPE_compact2,and HLL_RL_gln),respectively,with the AUC values of ROC 0.861(90.1% sensitivity and65.8% specificity)and 0.926(89.2% sensitivity and 68.0% specificity).Conclusion.CE-MRI texture analysis improves the prediction performance of pCR to NAC in breast cancer.Both prediction models categorized by two mass-enhancement types of breast lesions(mass vs non-mass)outperform significantly the non-categorized model.