This paper presents a comprehensive study on the identification of key radiomic features for breast cancer classification using mammographic images from the CBIS-DDSM dataset. Four independent experiments were conducted, corresponding to two lesion types—masses and calcifications—across two standard mammographic views: craniocaudal (CC) and mediolateral oblique (MLO). Radiomic features were extracted from segmented regions of interest and evaluated through a diverse feature selection pipeline, including Variance Thresholding, Correlation Filtering, SelectKBest, Recursive Feature Elimination (RFE), Lasso regularization, Boruta, and Permutation Feature Importance. The objective was to identify features consistently selected across methods, under the hypothesis that such consensus features may serve as robust radiomic biomarkers. The selected features were used to train Random Forest classifiers and evaluated through cross-validation using metrics such as accuracy, AUC, and F1-score. Among all configurations, the calcification-CC subgroup demonstrated the most robust and consistent predictive performance across all metrics, achieving an AUC of 0.90, F1-score of 0.81, and accuracy of 0.86. These findings highlight the utility of consensus-driven feature selection and support its role in future multimodal breast cancer diagnostic frameworks.