Abstract

Differentiation of Breast Cancer Immunohistochemical Status Using Digital Mammography Radiomics Features

Purpose: Discriminating breast cancer Hormonal Receptor (HR), human epidermal growth factor receptor (Her2) and Triple Negative (TN) status using mammography radiomic features. Materials and Methods: We used an open-source database enrolling 71 patients with confirmed breast cancer. It includes bilateral mammograms Craniocaudal (CC) and Mediolateral Oblique (MLO) as well as the breast cancer molecular status such as HR, Her2 and TN. We extracted a set of 486 quantitative descriptors from the original and the wavelets of the CC and the MLO mammograms. Using the training set (n_train=48), we performed the features selection following two steps: (i) first, univariable feature selection had been implemented with correlation statistical test to eliminate redundancy between mammogram features. (ii) In second part, we used Support Vector Machine Recursive Feature Elimination (SVM-RFE) method in 10-folds Cross-Validation repeated 10 times. Also, we applied the Synthetic Minority Oversampling Technique to tackle the issue of imbalanced classes. After that, we carried out three binary molecular classification (HR vs non-HR, Her2 vs non-Her2, TN vs non-TN) using Logistic Regression. These classifications were performed using respectively CC and MLO features individually and in two combinations: sum “CC+MLO” and concatenation “CC and MLO”. After the validation step (ntest=17), Accuracy and Under Receiver Operating Characteristic curve (AUC) were adopted to assess the proposed model performance. Results: Accuracies and AUCs recorded for three molecular classes in validation step were respectively ranging from 0.69/0.75 to 0.88/0.90, 0.52/0.53 to 0.64/0.63 and 0.70/0.70 to 0.79/0.77 for Her2, HR, TN. The best performances achieved for HR and Her2 classification were CC image features and “CC and MLO” features for TN. There is a strong representation of wavelets features based in the features selected. Conclusion: Our results suggest that mammographic quantitative features especially wavelet-based could be used to differentiate the breast cancer molecular subtype

• Full-Text
• PDF