Prediction Of Molecular Subtype Of Triple Negative Breast Cancer Based On Radiomics

Triple negative breast cancer(TNBC)is the most malignant molecular subtypes of breast cancer.Compared with other molecular subtype,triple-negative breast cancer is extremely prone to metastasis and recurrence.Immunohistochemistry is widely used to determine the molecular subtype of breast cancer in clinical.However,radiology uses non-invasive scanning technology to obtain the patient’s preoperative images.The application of radiomics can predict the targeted treatment response of patients,so more and more studies have begun to focus on the prediction of the molecular subtype of breast cancer.This paper proposes several radiomic features based on 2D radiomics and 3D radiomics,and constructs a classification model that combines the radiomic features extracted from the intra-tumoral and peri-tumoral regions to distinguish TNBC from other breast cancer molecular subtypes.In this study,120 cases of 2D image data containing four molecular subtypes and 101 cases of 3D image data containing three molecular subtypes were used to construct 2D and 3D radiomic features.The 2D radiomic features include gray-level statistical features,Gabor features,Laws features,Haralick features and Co LIAGe features.The 3D radiomic features include morphological texture features and 3D Co LIAGe features.Firstly,features were extracted from both inside and outside the tumor,the feature selection method of minimum redundancy maximum correlation(m RMR)was used to select effective features to build a classification model.Then,the combination of effective features from intra-tumoral and peritumoral region were used to build a classification model,the linear discriminative analysis classifier was used in this study.The experimental results have shown that the combination of features extracted from intra-tumoral and peri-tumoral region achieved the highest AUC value of 0.75±0.04 in distinguishing TNBC from Luminal A subtype.The AUC value in distinguishing TNBC from HER2,Luminal B and Non-TNBC was 0.69±0.04,0.64±0.05 and 0.72±0.04,respectively.The performance of the automatic segmentation model based on conditional generative adversarial networks was also evaluated in this study,the Dice coefficient of automatic segmentation reached 0.932.The radiomics extracted from the automatic segmentation region and the doctor’s manually annotation region was slightly different,which indicating that the tumor region segmented by the automatic segmentation model is useable in radiomics.The 3D radiomics experimental results achieve a maximum AUC value of 0.71±0.07,accuracy of 0.81±0.04,sensitivity and specificity of 0.77±0.10 and 0.84±0.08 in the TNBC and HER2 experiments.The AUC value for distinguishing TNBC and Luminal B experiments was 0.67±0.07,the accuracy,sensitivity and specificity were 0.77±0.05,0.85±0.10 and 0.72±0.10,respectively.The results of this study indicate that the radiomic features extracted from intra-tumoral and peri-tumoral regions can predict the molecular information of breast cancer.Meanwhile,radiomic features of the tumor surrounding areas that characterize the tumor microenvironment can predict the molecular subtypes of breast cancer.