<![CDATA[The primary objective of this study is to design a cascaded classification framework that integrates deep-learning representations with handcrafted and clinical features to enhance the reliability and accuracy of breast cancer detection in mammographic screening. A multi-source mammography dataset comprising four databases was used to ensure diversity and reduce bias. The proposed system operates in two stages. In the first stage, transfer learning models (VGG16, ResNet50, and EfficientNet_B0) were evaluated using ROC-AUC, PR-AUC, calibration curves, and bootstrap confidence intervals. EfficientNet_B0, which achieved the best balance between discrimination and calibration, was selected as the feature extractor. In the second stage, the malignancy probability was combined with Haralick texture features, patient age, and breast density, and classified using SVM, Random Forest, MLP, Decision Tree, and Logistic Regression. Model robustness was verified through multi-run experiments (five random seeds) and subgroup analyses by age and density. Among the CNN models, EfficientNet_B0 yielded the best performance (accuracy = 0.9438, ROC-AUC = 0.944, PR-AUC = 0.960). In the second stage, although Random Forest achieved the highest accuracy (0.9556 ± 0.002), SVM obtained the highest mean ROC-AUC (0.980 ± 0.001) with stable accuracy (0.9539 ± 0.001) and the most significant p-values, indicating superior robustness and generalization. The proposed cascaded framework effectively combines deep, handcrafted, and clinical features to improve mammogram classification performance. The SVM-based model demonstrates strong calibration, stability, and subgroup consistency, highlighting its potential for deployment in computer-aided mammography screening systems that assist radiologists in early breast cancer detection.]]>
