Attention-Gateway U-Net for Mammographic Mass Segmentation: An Empirical Study

Context: Accurate segmentation of breast masses in mammographic images is a critical step for early breast cancer detection and effective clinical decision-making. However, this task remains challenging due to variability in mass appearance, low contrast with surrounding tissues, and complex anatomical structures. Objective: This study proposes a novel deep learning framework based on an attention-guided U-Net architecture, specifically designed for the segmentation of suspicious masses in mammograms. The primary objective is to investigate the impact of integrating attention gates into the U-Net structure and to evaluate how these mechanisms influence the accuracy and robustness of mass segmentation. Method: Attention gates are embedded within the skip connections and function by dynamically highlighting relevant features from the encoder while suppressing less informative regions. This selective focus enables the model to more accurately delineate the often subtle boundaries of breast masses. Results: The proposed model is evaluated on the publicly available INbreast mammography dataset, and its performance is compared against the baseline U-Net and other state-of-the-art segmentation methods. Quantitative results demonstrate that the attention-enhanced U-Net significantly outperforms its counterparts in segmentation accuracy, particularly in challenging cases involving dense breast tissue or ill-defined masses. Conclusions: This study highlights the effectiveness of attention mechanisms in enhancing mammographic mass segmentation and represents a valuable step toward more intelligent and reliable Computer-Aided Diagnosis (CAD) systems for breast cancer.