<![CDATA[Early detection of skin cancer is critical for improving clinical outcomes, yet access to dermatological expertise remains limited in many regions. Smartphone-based artificial intelligence offers a promising alternative for accessible skin lesion screening; however, practical deployment requires an appropriate balance between diagnostic performance and computational efficiency. This study compared three deep learning architectures—MobileNetV2, EfficientNetB0, and Swin Transformer—to identify the most suitable model for smartphone-based skin lesion classification. A comparative experimental design was conducted using transfer learning with ImageNet-pretrained models and the ISIC 2019 dermoscopic image dataset comprising eight skin lesion categories. Model performance was evaluated based on predictive capability and on-device resource utilization following Android implementation. The findings reveal a clear trade-off between classification performance and computational efficiency. EfficientNetB0 achieved the strongest predictive performance, whereas MobileNetV2 consistently demonstrated superior computational efficiency, requiring the least processing resources, memory consumption, energy usage, and inference time during smartphone deployment. Although Swin Transformer delivered competitive predictive capability, its substantially higher computational demands reduced its practicality for resource-constrained devices. These results indicate that lightweight convolutional architectures remain the most appropriate choice for real-time smartphone-based skin cancer screening. This study provides an evidence-based framework for selecting deep learning architectures that balance diagnostic accuracy with deployment efficiency in mobile health applications.]]>
Copyrights © 2026
