Advancing ultrasound breast cancer classification via a dual-backbone fusion model integrating multi-scale attention mechanisms and transformer encoder
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE, vol.181, 2026 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 181
- Publication Date: 2026
- Doi Number: 10.1016/j.engappai.2026.115745
- Journal Name: ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Applied Science & Technology Source, Compendex, INSPEC, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO)
- Istanbul University Affiliated: Yes
Abstract
Breast cancer remains a leading cause of cancer-related mortality among women, highlighting the need for early and accurate tumor assessment from ultrasound images. Existing methods for breast ultrasound classification often rely on a single backbone network, provide limited feature diversity, and offer insufficient interpretability for clinical use. This study proposes an applied artificial intelligence framework for breast ultrasound decision support, combining a 50-layer Residual Network (ResNet-50) branch enhanced with the Convolutional Block Attention Module (CBAM), coordinate attention, anatomy attention, and a transformer encoder, with an EfficientNet-B4 branch enhanced with CBAM and a projection module. High-dimensional features are fused using five strategies -Tensor-Convolutional Neural Network (Tensor-CNN) fusion, cross-attention, bilinear pooling, transformer-based fusion, and gated fusion -and classified using an ensemble deep Random Vector Functional Link (edRVFL) network and six standard machine learning classifiers. Experiments on the Breast Ultrasound Images dataset (BUSI, n = 780), the University of Castilla-La Mancha breast ultrasound dataset (BUS-UCLM, n = 683), and a combined dataset (n = 1463) used five-fold cross-validation with image-level splitting for BUSI and patient-wise grouping for BUS-UCLM. The strongest BUSI result achieved 0.9705 f 0.0397 accuracy, 0.9704 f 0.0400 F1-score, and 0.9939 f 0.0133 area under the receiver operating characteristic curve (AUC); the strongest BUS-UCLM result reached 0.9563 f 0.0372 accuracy, 0.9539 f 0.0425 F1-score, and 0.9852 f 0.0210 AUC; and the combined dataset achieved 0.9578 f 0.0336 F1-score and 0.9918 f 0.0099 AUC across five-fold cross-validation. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations support qualitative interpretability by highlighting lesion-relevant regions.