Gradient-guided Multi-attention Fusion: An Artificial Intelligence-enhanced Melanoma Diagnosis System Integrating Convolutional Neural Network and Vision Transformer Architectures

Abstract Background: We propose Gradient-guided Multi-attention Fusion (GGMAF), a novel artificial intelligence-enhanced melanoma diagnosis system that integrates convolutional neural networks (CNNs) and vision transformers (ViTs) to address the limitations of conventional methods in skin lesion analysis. Methods: The proposed framework dynamically combines local texture features from CNNs with global contextual representations from ViTs through a gradient-guided attention mechanism, which prioritises diagnostic regions while suppressing artefacts in dermoscopic images. To enhance resilience, the system employs a multiscale fusion module to hierarchically aggregate features from both pathways, followed by a learnable ensemble strategy that adaptively weights predictions based on image characteristics. Results: Our approach significantly improves diagnostic accuracy by simultaneously capturing fine-grained details and structural patterns, with inference times compatible with clinical workflows. The architecture interfaces seamlessly with existing medical systems, providing both malignancy probabilities and interpretable heat maps for clinical decision support. Conclusion: Experimental validation demonstrates superior performance compared to state-of-the-art methods, particularly in challenging cases with ambiguous lesion boundaries or heterogeneous textures. Furthermore, the gradient-based attention mechanism offers inherent explainability, aligning model decisions with clinically salient features. These findings highlight the potential of GGMAF as a dependable assistive tool and advance the field of computational dermatology by establishing a robust, interpretable and scalable solution for early melanoma detection and contributing to better patient outcomes in dermatologic care.