Advanced Breast Cancer Classification Using a Multi-Branch Spectral Channel Attention Network in Histopathology Imaging

In the field of medical image analysis, deep learning has revolutionized the diagnosis and classification of diseases, notably breast cancer. Current methods predominantly utilize convolutional neural networks (CNNs) to analyze spatial features within histopathology images. Despite their efficacy, these methods often overlook the rich frequency domain information that texture-based images inherently possess. Recognizing this gap, our research introduces the Multi-Branch Spectral Channel Attention Network (MbsCANet), a novel architecture designed to enhance breast cancer classification by integrating frequency domain analysis with spatial domain insights. This approach leverages a dual-channel mechanism that combines low and high-frequency features extracted through the two-dimensional discrete cosine transform (DCT). By doing so, MbsCANet not only preserves phase information but also enriches the contextual data crucial for accurate classification. We rigorously tested our model using the BreakHis dataset, achieving unprecedented accuracy levels of 99.01% at image-level and 98.87% at patient-level classification. These results significantly outperform traditional spatial-domain-focused models. The visualization of the model's processing layers further validates its effectiveness, presenting a compelling case for the integration of spectral analysis in medical imaging diagnostics.