LGGC-Net: a local-global graph and color attention-based lightweight CNN for skin cancer classification

Developing clinically deployable AI systems for skin cancer classification remains challenging due to limited robustness, lack of interpretability and constrained computational resources in hospitals. Although many deep learning models report high accuracy, their large sizes, extensive training requirements and low generalizability hinder practical deployment. In this study, we propose LGGC-Net, a lightweight CNN that incorporates LGGC (Local, Global Graph, and Color) attention to enhance discriminative feature learning while maintaining computational efficiency. Experimental results demonstrate that LGGC attention consistently improves performance across all evaluated CNN backbones. The LGGC-Net model was assessed on external image sets with diverse skin tones to ensure robustness and generalizability under domain shift conditions. Ablation studies analyzed the contribution of individual attention components and explainability was examined using Gradient-weighted Class Activation Mapping++ (Grad-CAM++) and SHapley Additive exPlanations (SHAP). With only 0.81 million parameters, LGGC-Net achieved 88.05% accuracy in 50 epochs, corresponding to 1.761 accuracy per epoch and 108.7 accuracy per million parameters in binary classification. In multiclass settings, it attained 76.1% accuracy on the unseen HAM10000 dataset, with 1.52 accuracy per epoch and 94.0 accuracy per million parameters. In both cases, the area under the curve exceeded 0.93. LGGC-Net consistently outperformed existing methods on deployment-oriented metrics, maintaining stable accuracy. These results indicate that LGGC-Net is an effective, interpretable and potentially deployment-ready solution for practical skin cancer classification.