Evaluation of Data Balancing Methods for the Classification of Digital Mammography Images with Benign and Malignant Breast Lesions Using Machine Learning

Breast cancer is highly prevalent and a leading cause of cancer-related death in women. Early detection through mammographic imaging is critical but challenging due to subjectivity among doctors and the complex clinical context. Additionally, image datasets commonly exhibit class imbalances, posing a greater challenge compared to classification problems in other fields. In this work, we explore various class balancing techniques to enhance the predictive performance of machine learning models. We use the publicly available dataset "The mini-MIAS database of mammograms" to train SVM and CNN models (Suckling et al. in The mammographic image analysis society digital mammogram database. University of Essex, 1994 [1]), comparing their performance with and without class balancing preprocessing and ensemble methods to determine their impact on sensitivity and specificity in classification. This is done using metrics such as accuracy, F1-score, sensitivity, and specificity. The experiments presented lay the foundation for addressing issues with imbalanced datasets in the context of automated detection of anomalies in mammograms. These findings can be extended to test other class-balancing strategies and data preprocessing approaches.