Improving Breast Cancer Diagnosis using Attention-Enhanced Hybrid CNN–Transformer Model

Breast cancer is the most common cancer among women and the most frequently diagnosed cancer worldwide. Recent advancements in deep learning have led to significant improvements in tumor detection from breast ultrasound (BUSI) images, enhancing the diagnostic accuracy of breast cancer screening. Although deep convolutional neural networks (CNNs) and transformer-based architectures have individually yielded promising results, challenges such as low contrast, spatial variability, and irregular tumor shapes continue to hinder the robustness of current methods. Therefore, in this study, a novel hybrid CNN–Transformer framework is proposed to improve discriminative feature extraction for BUSI cancer analysis. The network employs a dual-branch architecture, integrating features extracted from both CNN and transformer models. In the first branch, the Swin Transformer is combined with a Triplet Attention to strengthen its ability to learn long-range dependencies and global contextual information. The Triple Attention module processes feature maps along three orthogonal axes, enabling a more effective representation of both spatial and channel-level relationships. The second branch incorporates the Efficient Net architecture augmented with an Efficient Channel Attention (ECA) module, which facilitates adaptive channel-level feature recalibration. This design allows the model to emphasize diagnostically salient regions within ultrasound images. High-level features from both branches are fused for final classification. Experimental results on the BUSI dataset demonstrate that the proposed architecture achieves superior performance, with 97.4% accuracy, 97.9% precision, 97.9% sensitivity, and a 97.9% F1-score. These outcomes confirm the effectiveness of the proposed hybrid CNN–Transformer design in improving automated breast cancer diagnosis using ultrasound imaging.