Enhancing Interpretability in Mammography Pathology Detection Using Deep Convolutional Features and Self-Organizing Maps

Breast cancer remains a prominent global health concern with its critical emphasis on early detection. Mammography plays a crucial role in the early-stage identification of breast abnormalities. However, the interpretative process poses challenges due to its complexity and time-consuming nature, particularly for radiologists. Considering prior research findings, this study introduces an innovative methodology for the detection and interpretation of mammography pathology, utilizing the comprehensive CBIS-DDSM database. The proposed approach comprises two key phases: Firstly, Convolutional Neural Networks are meticulously trained using transfer learning and from-scratch methodologies, to address two critical classification tasks: distinguishing calcification from mass lesions and discriminating benign calcifications from malignant ones. The study encompasses a diverse array of convolutional architectures, including EfficientNets, Inceptionv3, MobileNet, ResNet50, and VGG19, exploring an extensive range of hyperparameters beyond previous research. Secondly, features are extracted from the trained convolutional networks and employed to train Self-Organizing Maps. This method establishes robust rules based on label maps and U-matrices, enabling three outcomes: i) identification of the optimal convolutional network; ii) identification of the most effective features, by considering topological separation between classes; iii) interpretation of these features and their interplay with the convolutional network's performance. The results are compelling, with an accuracy of 0.928 ± 0.008 achieved in distinguishing calcification from mass lesions, surpassing state-of-the-art results. In the case of distinguishing benign from malignant calcifications, an accuracy of 0.750 ± 0.004 is attained.