Enhanced Feature Selectivity in MobileNetV2 for Skin Cancer Detection through Scaled Dot-Product Attention

Abstract

This study presents an enhanced MobileNetV2-based model architecture for efficient and accurate skin cancer classification, suitable for mobile deployment. Addressing the limitations of current MobileNet V2 and V3 architectures, we integrate a Scaled Dot-Product Attention mechanism to improve feature selectivity while maintaining computational efficiency. Our model was trained and evaluated on a balanced subset of the ISIC 2019 dataset, employing data augmentation and class-balancing techniques to mitigate dataset imbalances and enhance generalization. Comparative experiments demonstrated that the proposed model outperformed standard MobileNet V2 and V3 models in validation accuracy, particularly in distinguishing melanoma, basal cell carcinoma, and squamous cell carcinoma. Future research will focus on further optimizing the model’s architecture by incorporating additional MobileNet V3 features, expanding testing on diverse datasets, and exploring alternative attention mechanisms. This work supports the development of mobile-accessible, AI-driven diagnostic tools, contributing to early skin cancer detection and potentially improving public health outcomes.