Background/Introduction Skin lesion classification poses a critical diagnostic challenge in dermatology, where early and accurate identification has a direct impact on patient outcomes. While deep learning approaches have shown promise using dermatoscopic images alone, the integration of clinical metadata remains underexplored despite its potential to enhance diagnostic accuracy. Methods We developed a novel multimodal data fusion framework that systematically integrates dermatoscopic images with clinical metadata for the classification of skin lesions. Using the HAM10000 dataset, we evaluated multiple fusion strategies, including simple concatenation, weighted concatenation, self-attention mechanisms, and cross-attention fusion. Clinical features were processed through a customized Multi-Layer Perceptron (MLP), while images were analyzed using a modified Residual Networks (ResNet) architecture. Model interpretability was enhanced using Gradient-weighted Class Activation Mapping (Grad-CAM) visualization to identify the contribution of clinical attributes to classification decisions. Results Cross-attention fusion achieved the highest classification accuracy, demonstrating superior performance compared to unimodal approaches and simpler fusion techniques. The multimodal framework significantly outperformed image-only baselines, with cross-attention effectively capturing inter-modal dependencies and contextual relationships between visual and clinical data modalities. Discussion/Conclusions Our findings demonstrate that integrating clinical metadata with dermatoscopic images substantially improves the accuracy of skin lesion classification. However, challenges, including class imbalance and the computational complexity of advanced fusion methods, require further investigation.
Das et al. (Mon,) studied this question.