Prediction of benign and malignant breast masses using digital mammograms texture features

Abstract

Background: Breast cancer is one of the most common malignant disease for women. Mammography is the preferred method for breast cancer detection. The purpose is to investigate the feasibility and accuracy of texture features extracted from digital mammograms at predicting benign and malignant breast mass using Radiomics. Methods: 494 digital mammograms data who diagnosed as breast masses (Benign: 251 Malignant: 243) by mammography were enrolled. Enrol criteria: breast masses classified as BI-RADS 3, 4, and 5 and at last confirmed by histopathology. Lesion area was marked with a rectangular frame on the Cranio-Caudal (CC) and MedioLateral Oblique (MLO) images at the 5M workstation. The rectangular regions of interest (ROI) was segmented and 456 radiomics features were extracted from every ROI. Extracted features were dimensioned by Maximum Relevance Minimum Redundancy (MRMR) and Lasso algorithm. Post-dimension features were classified using Support Vector Machine (SVM). 70% of the data as a training set and the other 30% as a testing set. The reliability of the Classifier was evaluated by the 10-fold cross-validation. The classification accuracy was evaluated by the accuracy and sensitivity and AUC. Results: Both the MRMR and Lasso screened 30 radiomics features respectively. 10- fold cross-validation showed that their accuracy were 88.70% and 86.71%, respectively. In testing sets, Through the MRMR algorithm, the classifier achieves an accuracy of 92.00% and a sensitivity of 91.10% and AUC of 95.10%. Through the lasso dimension reduction algorithm, the classifier achieves an accuracy of 83.26% and a sensitivity of 75.90% and AUC of 89.38%. Conclusions: Radiomics texture features from digital mammograms may be used for benign and malignant prediction. This method offer better accuracy and sensitivity. It is expected to provide an auxiliary diagnosis for the imaging doctors.