Application of Radiomics Analysis on Mammography for Differentiating Benign and Malignant Masses

Abstract

Radiomics is emerging as a promising quantitative tool for extracting imaging features. The diagnostic accuracy of mammography may be altered by the reader’s experience. To build a radiomics model to differentiate benign and malignant mammographic masses and to evaluate whether the diagnostic performance of mammography could be improved by radiomics. In this retrospective study, 101 patients were included in the training set, and 36 patients were included in the test set. A total of 127 radiomics features were extracted from each mammographic mass using both craniocaudal (CC) and mediolateral oblique projection (MLO) images. RELIEF algorithm and Mann–Whitney-U test were used for feature selection. Seven machine learning algorithms were applied to construct a predictive model. Machine learning algorithms were trained by stratified tenfold cross-validation on the training set. The classification performance of the radiomics models was compared with the diagnostic predictions of two radiologists with different experience levels by using receiver operating characteristic curve (ROC) analysis. A total of 15 radiomics features remained after the pre-processing. “XGBoost” presented the best differentiation ability among the seven machine learning methods and yielded 0.871 area under the curve (AUC), 0.888 accuracy, 0.895 sensitivity, and 0.875 specificity. The breast radiologist yielded 0.901 (AUC), 0.916 accuracy, 0.956 sensitivity, and 0.846 specificity. The inexperienced radiologist yielded 0.819 (AUC), 0.833 accuracy, 0.869 sensitivity, and 0.769 specificity. The radiomics-based model shows favorable outcomes in differentiating benign and malignant mammographic masses. Quantifiable assessment of mammography could enhance radiology practice, especially for inexperienced radiologists. © 2025 Elsevier B.V., All rights reserved.