Accurate and reliable breast cancer detection from mammographic images remains a critical challenge due to subtle lesion appearance, high intra-class variability, and class imbalance inherent in clinical datasets. To address these issues, this study proposes Swin-BreastNet, an explainable and optimization-driven deep learning framework for binary classification of benign and malignant breast lesions from full-field digital mammograms. The proposed approach leverages the hierarchical Swin Transformer model to effectively capture fine-grained local texture patterns and long-range contextual dependencies through Shifted Window Multi-head Self-Attention (SW-MSA). A key novelty of this work lies in the integration of Harris Hawks Optimization (HHO) for automated hyperparameter tuning of the proposed transformer architecture. Unlike conventional manual or grid-based tuning strategies, HHO formulates hyperparameter selection as a population-based global optimization problem, directly maximizing validation Area Under the Curve (AUC) and enabling robust exploration–exploitation trade-offs. A principled image preprocessing pipeline, including resolution normalization, intensity scaling, and structured data augmentation, is employed to reduce acquisition variability and enhance generalization. The model is validated using the INbreast dataset, achieving a detection accuracy of 95.6% and a mean Average Precision (mAP) of 93.9%. Explainability is enhanced using attention rollout visualizations and SHapley Additive exPlanations (SHAP) values to provide insight into decision-making processes. The proposed Swin-BreastNet demonstrates better diagnostic performance and interpretability, making it a suitable candidate for reliable breast cancer screening tools in clinical settings.
Abugabah et al. (Tue,) studied this question.