Microcalcification Clusters (MCCs) are radiological markers of breast cancer. However, accurate diagnosis of these minute calcium deposits remains a significant challenge due to their spatial obscurity, particularly in dense breast tissues. Reducing false positives is crucial for detecting MCCs in mammograms, as traditional methods often yield erroneous cases. We present WCMA-Net (Wavelet-based Channel-wise Mamba Attention), an interpretable deep learning framework that integrates Discrete Wavelet Transform (DWT), channel and spatial attention, and a state-space Mamba attention mechanism to improve MCC detection. It isolates high-frequency diagnostic cues through wavelet decomposition and enhances feature discriminability via dual attention mechanisms. The Mamba attention further captures long-range dependencies and temporal-spatial dynamics, facilitating precise classification. We incorporate Gradient-weighted Class Activation Mapping (Grad1CAM) visualizations to explain model decisions and highlight diagnostically relevant regions. Experiments on two benchmark datasets demonstrate state-of-the-art performance, achieving AUCs of 0.99 and 0.96, with high sensitivity and specificity. Compared to Swin Transformer and Vision-Mamba models, WCMA-Net delivers superior accuracy with lower computing cost Giga Floating Point Operations Per Second (0.43 GFLOPs), suitable for real1time applications. The results establish WCMA-Net as a practical, interpretable system for breast cancer screening.
Rehman et al. (2026) studied this question.