AE-CCAF: Cross-Attention Guided Ensemble Learning for Robust Multi-Class Skin Lesion Classification
Proper and early classification of skin lesions is crucial for the effective diagnosis of melanoma, but remains challenging due to class imbalance, class similarity, and artefacts in dermoscopic images. This paper proposes AE-CCAF, an adaptive ensemble framework that combines heterogeneous deep backbones with a cross-attention fusion mechanism to robustly classify multi-class skin lesions. The model uses complementary representations from EfficientNet, Swin Transformer, and ConvNeXt, along with a cross-attention module to dynamically weight inter-feature interactions. As a measure to tackle data imbalance, an asymmetric focal loss using class-sensitive weighting is added. Extensive experiments on the ISIC 2018 benchmark indicate that AE-CCAF achieves a macro-AUC of 0.9487 and a balanced accuracy of 83.7%, outperforming current state-of-the-art algorithms. The proposed solution enhances the sensitivity for clinically urgent cases of melanoma while maintaining very high specificity across all levels. These findings point to the success of attention-guided ensemble learning to provide reliable and scalable dermatological diagnosis.