Leveraging Attention-Based Deep Learning in Binary Classification for Early-Stage Breast Cancer Diagnosis.

Abstract

Background: Breast cancer diagnosis is a global health challenge, requiring innovative methods to improve early detection accuracy and efficiency. This study investigates the integration of attention-based deep learning models with traditional machine learning (ML) methods to classify histopathological breast cancer images. Specifically, the Efficient Channel-Spatial Attention Network (ECSAnet) is utilized, optimized for binary classification by leveraging advanced attention mechanisms to enhance feature extraction across spatial and channel dimensions. Methods: Experiments were conducted using the BreakHis dataset, which includes histopathological images of breast tumors categorized as benign or malignant across four magnification levels: 40×, 100×, 200×, and 400×. ECSAnet was evaluated independently and in combination with traditional ML models, such as Decision Trees and Logistic Regression. The study also analyzed the impact of magnification levels on classification accuracy, robustness, and generalization. Results: Lower magnification levels consistently outperformed higher magnifications in terms of accuracy, robustness, and generalization, particularly for binary classification tasks. Additionally, combining ECSAnet with traditional ML models improved classification performance, especially at lower magnifications. These findings highlight the diagnostic strengths of attention-based models and the importance of aligning magnification levels with diagnostic objectives. Conclusions: This study demonstrates the potential of attention-based deep learning models, such as ECSAnet, to improve breast cancer diagnostics when integrated with traditional ML methods. The findings emphasize the diagnostic utility of lower magnifications and provide a foundation for future research into hybrid architectures and multimodal approaches to further enhance breast cancer diagnosis.