Skin Cancer Classification Using a Transfer Learning—Enhanced Hybrid Architecture with Multi-Scale Attention

Early and precise identification of melanoma from dermoscopic images is crucial for timely treatment, yet manual examination remains inconsistent because it is influenced by subjective visual interpretation. Deep learning has improved the reliability of lesion classification, but real-world challenges—such as considerable variation within the same lesion category, visual resemblance across different classes, and imbalanced datasets—still limit performance. To address these gaps, this work proposes an improved melanoma classification model based on EfficientNet-B3, enhanced with a Multi-Scale Attention Fusion Module (MSAFM). This module merges multi-scale convolutional branches with both spatial and channel-attention components to strengthen contextual awareness and emphasize clinically important patterns. Using the HAM10000 dataset, consisting of over 10,000 dermoscopic images, the model achieves 93.8% accuracy, a macro-F1 of 92.6%, and an AUC of 0.966. These findings suggest that the proposed framework provides a dependable foundation for computer-assisted dermatological diagnosis.