Evaluating Lightweight CNN Models with CBAM for Explainable AI Skin Cancer Classification using GradCAM and ScoreCAM

Skin cancer continues to be a prevalent and deadly malignancies cancer globally, necessitating early and accurate diagnosis. Lightweight CNN has emerged as effective tools for automated skin lesion classification, demonstrating significant potential due to their computational efficiency and scalability. Their ability to efficiently process and analyze medical imaging data makes them particularly suitable for this application in dermatology. However, their limited interpretability and varying performance across architectures remain critical concerns in clinical deployment. We compare three lightweight backbones MobileNetV2, ShuffleNetV2, and Efficient-NetV2B0 - each integrated with CBAM to improve attention localization. Classification performance was assessed via accuracy, precision, recall, F1-score, and AUC-ROC. We also apply two visualization techniques, GradCAM and ScoreCAM, to evaluate the transparency and understandability of model decision-making processes. Experiments were conducted on the ISIC 2024 SLICE-3D dataset, with dermoscopic images across four lesion types. Results show that EfficientNetV2B0 with CBAM outperforms other models in terms of classification accuracy (97%). The findings suggest that EfficientNetV2B0 with CBAM offers a compelling balance between accuracy and interpretability. This research highlights the important role of integrating attention mechanisms with explainable AI frameworks for developing trustworthy and deployable clinical diagnostic tools.