FCC-FMLO and FLeft-FRight: two novel multi-view fusion techniques for breast density assessment from mammograms

Breast density classification presents a crucial risk factor for breast cancer. The breast density evaluation via the breast imaging reporting and data system (BI-RADS) presents several challenges for medical professionals due to their variable diagnostics and workload. To address these challenges, researchers have developed computer-aided diagnostic systems using machine learning and deep learning techniques. In this research, we propose two novel multi-view approaches for breast density classification: Fusion Left-Right (FLeft-FRight) and Fusion Cranial Caudal-Mediolateral Oblique (FCC-FMLO). To evaluate the proposed methods, we conducted an extensive comparative analysis comparing multiple feature extractors and classifiers. Furthermore, we introduced dilated convolution and channel-wise attention mechanisms to enhance the feature extraction process. We carefully considered all these components and developed a robust model based on the FCC-FMLO and the DenseNet201+LGBM architecture. This combination achieved a macro F1-score of 69.1% on the public VindrMammo dataset. Furthermore, we validated its efficiency and explainability with the GRAD-CAM visualization technique.