Attention-Guided Fusion of EfficientNet-B0 and Swin Transformer for Cervical Cancer Classification

The interpretation of colposcopy images is a critical yet subjective component of cervical cancer screening. To enhance this process, we propose a novel hybrid deep learning framework for the classification of cervical lesions. Our model integrates EfficientNet-B0, adept at extracting localized hierarchical features, with a Swin-Tiny Transformer, which excels at modeling long-range dependencies and global context. Moving beyond basic fusion techniques, we introduce a novel cross-attention fusion mechanism, augmented with channel and spatial attention modules. This design selectively highlights the most discriminative inter-feature relationships while maintaining computational efficiency. Evaluated on the International Agency for Research on Cancer (IARC) colposcopy image dataset, our framework achieves an accuracy of 94.76%, significantly outperforming a concatenation-based fusion model (83.99%). This represents an absolute improvement of 10.77 percentage points and captures 67.3% of the residual performance margin toward perfect ac-curacy. The model also demonstrates robust performance across other metrics, including a precision of 94.68%, recall of 94.82%, F1-score of 94.74%, and a Cohen’s Kappa of 89.48%. These results indicate that our approach can enhance both the accuracy and reliability of cervical cancer screening, offering valuable support for clinical decision-making.