Scholarlink

Artificial intelligence (AI) in dermatology has enormous potential to improve patient care and diagnostic precision. With an emphasis on accuracy and interpretability, this work offers a thorough examination of many deep learning architectures for the categorisation of skin lesions. Eight cutting-edge CNN models–Xception, ResNet50, AlexNet, EfficientNetV2L, MobileNetV2, DenseNet201, DenseNet121, and VGG16–are implemented and contrasted. Metrics like F1-score, recall, accuracy, and precision are used to assess their performance. We use interpretability approaches like Grad-CAM and LIME to increase transparency by giving model decisions visual explanations. With a 95.30% testing accuracy and a 95% F1-score, our results show that the Xception model performs better than the EfficientNetV2L and DenseNet variations, which also show competitive accuracy and resilience. The study concludes with a web-based implementation that integrates these models with visualization tools, offering a practical solution for clinical applications and paving the way for enhanced AI-driven dermatological diagnostics.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Skin Cancer Detection Using Interpretable Deep Learning: A Comparative Analysis of CNN Architectures

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Archana Kotangale,
  • Selvin Furtado,
  • V. Anusha,
  • Khushboo Jain,
  • Manmeet Kaur Huda

摘要

Artificial intelligence (AI) in dermatology has enormous potential to improve patient care and diagnostic precision. With an emphasis on accuracy and interpretability, this work offers a thorough examination of many deep learning architectures for the categorisation of skin lesions. Eight cutting-edge CNN models–Xception, ResNet50, AlexNet, EfficientNetV2L, MobileNetV2, DenseNet201, DenseNet121, and VGG16–are implemented and contrasted. Metrics like F1-score, recall, accuracy, and precision are used to assess their performance. We use interpretability approaches like Grad-CAM and LIME to increase transparency by giving model decisions visual explanations. With a 95.30% testing accuracy and a 95% F1-score, our results show that the Xception model performs better than the EfficientNetV2L and DenseNet variations, which also show competitive accuracy and resilience. The study concludes with a web-based implementation that integrates these models with visualization tools, offering a practical solution for clinical applications and paving the way for enhanced AI-driven dermatological diagnostics.