Deep Learning-Based Automated Reports for Breast Density Assessment in Mammography Images

Breast density is a key factor in mammographic screening, as high-density tissue increases cancer risk and can obscure lesions, reducing diagnostic sensitivity. This work presents a deep learning-based framework for automated breast density assessment in mammography, combining BI-RADS classification and dense fibroglandular tissue segmentation. To improve data diversity and mitigate class imbalance, we construct a fusion dataset from different sources and introduce manually refined classification and segmentation annotations. The framework generates structured reports that combine visualizations and quantitative metrics, providing interpretable outputs to support analysis of breast density patterns. In addition to standard classification and segmentation tasks, we include spatial and morphological metrics, such as tissue distribution across quadrants, shape regularity, proximity to anatomical landmarks, and enhanced patch-based visualizations. These reports aim to complement and clarify BI-RADS predictions, especially in borderline cases. Extensive experiments using several model architectures demonstrate strong performance on the internal dataset and reasonable generalization across external datasets. The proposed system provides a reproducible and interpretable approach for breast density analysis and may serve as a foundation for further validation toward standardized and explainable reporting in breast imaging.