Efficient Channel Attention Networks for Accurate Breast Cancer Histopathological Image Classification

Breast cancer is one of the most common and deadly cancers affecting women worldwide. Early and accurate diagnosis is essential for effective treatment. Histopathological image analysis is the standard method for diagnosing breast cancer. However, manual evaluation takes a lot of time, is very subjective, and can vary between different observers.

Deep convolutional neural networks (CNNs) have improved computer-aided diagnostic systems, but they often have trouble capturing complex features without significantly increasing computational costs. To address these issues, this study presents a new, lightweight transfer learning framework focused on the binary classification of benign and malignant breast tumors. We used the well-known BreaKHis dataset, which contains 7,909 histopathological images. We applied a rigorous stratified data split: 70% for training, 15% for validation, and 15% for testing, along with effective spatial and color augmentations to improve the model's generalization.

Our approach integrates an Efficient Channel Attention (ECA) module into popular, pre-trained CNN backbones, specifically ResNet101 and DenseNet169. The ECA mechanism allows for local cross-channel interaction without reducing dimensionality. This dynamically enhances the representation of important morphological features while keeping computational demands low. Extensive evaluations show that this integration leads to outstanding diagnostic performance. Our proposed ResNet101_ECA architecture achieved a peak test accuracy of 96.80%, with a macro precision of 96.22%, a macro recall of 96.35%, and a macro F1-score of 96.29%. In addition, the DenseNet169_ECA model reached a competitive overall accuracy of 96.63%. These results demonstrate that localized channel attention effectively identifies key histopathological features, providing a highly accurate, automated, and feasible solution for clinical pathology workflows.