RA-BCNet: Residual attention breast cancer network for high-precision breast cancer detection

Background Breast cancer remains a major global health concern, and early and accurate detection is critical for improving clinical outcomes. However, variations in lesion characteristics and imaging conditions in Digital Breast Tomosynthesis (DBT) images present substantial challenges for automated breast tumor classification. Objective This study aims to develop a robust and interpretable deep learning framework, termed RA-BCNet-PReLU, by integrating residual attention mechanisms with Parametric Rectified Linear Unit (PReLU) activation for accurate multi-class breast tumor classification. Methods The proposed framework employs a structured processing pipeline comprising Min–Max normalization, stratified sampling, and targeted image augmentation for data preprocessing and generalization. Lesion segmentation is performed using a convolutional neural network (CNN), with segmentation further optimized using the Dynamic Hybrid Grey Wolf–Sunflower Optimization (DHGWSFO) algorithm. Subsequently, handcrafted shape, color, and spatial descriptors are integrated with deep features extracted using ResNet-50 to obtain a comprehensive lesion representation. The fused features are then classified using a residual attention-based CNN with PReLU activation. The model was evaluated on a DBT mammography dataset and externally validated on the CBIS-DDSM dataset. Model reliability was assessed using five-fold cross-validation, ANOVA, paired t -tests, and Wilcoxon signed-rank tests, while Grad-CAM was employed to improve model interpretability. Results The proposed RA-BCNet-PReLU achieved a mean classification accuracy of 96.87% ± 0.42 on the DBT dataset and 98.72% ± 0.35 on the CBIS-DDSM dataset, with corresponding AUC values of 0.967 and 0.981, respectively. Statistical analyses demonstrated the reliability of the obtained results, while Grad-CAM visualizations highlighted clinically relevant regions contributing to the model's predictions. Conclusion The proposed RA-BCNet-PReLU framework demonstrates high classification performance, robustness, and interpretability across DBT and external mammography datasets. Its integration of optimized lesion segmentation, hybrid feature representation, residual attention, and explainable visualization indicates its potential as an effective computer-aided tool for automated breast cancer detection and clinical decision support.