Mammography Classification: How Useful is Machine Learning? A Radiomics Study and Future Perspectives

This study investigates the effectiveness of machine learning (ML)-based radiomics in classifying mammographic lesions. Leveraging the publicly available CBIS-DDSM and the matRadiomics toolbox, Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM) models were tested across progressively larger training sets. LDA achieved the highest performance, demonstrating excellent discrimination between masses and calcifications (AUC of 97.08%, Accuracy of 95.63%). However, classification of benign versus malignant lesions yielded lower AUCs of 68.28% for microcalcifications and 61.53% for masses highlighting the limitations of traditional ML approaches. These findings point toward the need for more advanced methods, such as deep learning, in future research.