EG-VAN: A Global and Local Attention-Based Dual-Branch Ensemble Network With Advanced Color Balancing for Multi-Class Skin Cancer Recognition

The accurate diagnosis of skin cancer remains challenging due to factors like intra-class variation and inter-class similarity in dermoscopic images. Despite their sophistication, existing methods struggle to generalize well across diverse cases, and their computational complexity often limits real-world applicability. To address these challenges, we introduce a novel dual-path architecture in the Efficient Global-Vision Attention-Net (EG-VAN) for early detection of skin cancer. The dual-path design uniquely combines (1) EfficientNetV2S, which processes multi-scale features with high efficiency, and (2) a modified ResNet50 enhanced with a Spatial-Context Group Attention (SCGA) module and Non-Local Block. The SCGA module captures local-contextual relationships in horizontal and vertical directions to focus on subtle lesion details, directly addressing intra-class variation, while the Non-Local Block captures long-range dependencies to resolve inter-class similarity. This synergistic dual-path fusion ensures comprehensive feature extraction, enabling EG-VAN to outperform single-path architectures. Further, we introduce a Multi-Scale Feature Fusion (MFF) module to aggregate hidden features across scales, improving generalization across skin cancer types. Additionally, our pioneering color balancing method, inspired by Gray World and Retinex theory, recalibrates color distribution to highlight lesion regions, enhancing the dual-path network’s input quality. We validate our EG-VAN through extensive experiments on the 9-class skin cancer dataset. Our experiments show improvement in performance by achieving an accuracy of 98.20%, a recall of 95.74%, and an F1-Score of 96.68%.