Recent Deep Learning Architectures for Mammography Image Seg-mentation: A Narrative Review

Abstract

Mammography remains the frontline imaging modality for early breast cancer detection, yet accurate segmentation of mammographic images poses significant challenges due to variability in breast tissue density, lesion appearance, and acquisition conditions. This review synthesizes advancements in mammography image segmentation based on deep learning, hybrid architectures, and advanced transformer-based models. Drawing from recent studies, we categorized methodologies into convolutional encoder–decoder structures, hybrid CNN–transformer frameworks, multi-task learning strategies, domain generalization approaches. Comparative analyses on benchmark datasets such as INbreast and CBIS-DDSM were conducted, employing performance metrics including Dice similarity coefficient, Intersection over Union (IoU), accuracy, sensitivity, and specificity. Comprehensive tables highlight method strengths, limitations, and evaluation results. We additionally provide an overview of dataset characteristics and discuss future trends in clinical deployment. Comparative analyses reveal the superiority of hybrid CNN-Transformer architectures, which excel by integrating local and global features. However, challenges such as computational cost and data dependency remain barriers to clinical implementation. Future trends focus on self-supervised learning and domain generalization to develop efficient and reliable tools.