Building an Extensive Database for Training Predictive Models in Image Classification of Mammography Views and Projections using Support Vector Machines
Fernanda Perez-Barcena, Rafael Bayareh-Mancilla, Yazmín Mariela Hernández-Rodríguez, Blanca Olivia Murillo-Ortiz, Oscar Eduardo Cigarroa-Mayorga · 2023
Breast cancer remains a significant cause of mortality among women worldwide. One of the early symptoms of breast cancer is morphological asymmetry. Therefore, prevention systems based on Machine Learning models for classifying breast asymmetry in mammograms can serve as a solution for early detection of breast cancer. This paper presents a method for classifying mammograms based on views and projections, with the aim of training machine learning models to detect asymmetries using Support Vector Machines. The database consisted of 54,668 mammograms, with 48% in Craniocaudal and 52% in Mediolateral Oblique projection, from women with an average age of 58 years ± 10, labeled with BIRADS 0, 1, and 2. Geometric features were extracted by generating masks for the mammograms. The features used for classifying the projection pair included Hu moments, asymmetry, pixel position, circularity, and ellipticity. For the orientation classification, the angle between the image center and the breast centroid was computed. This set of features served as the training information for the SVM model, which successfully classified the projection and orientation. Due to the nature of the features, the model was trained using a linear kernel with a regularization parameter C = 1. The classification performance was evaluated using accuracy metrics, achieving a result of 99% ± 1% and an F1 score = 0.99. The automatic classification approach enables the creation of a labeled database based on views and projections, enhancing the training of machine learning models for future research in this field.