Scholarlink

This study introduces a multimodal model for early skin cancer detection, leveraging image data and metadata from the ISIC 2024 dataset. Our method integrates CNNs and Transformer-based models to extract features from single-lesion images cropped from 3D Total Body Photographs (3D-TBP). These images, resembling close-up smartphone photos, are integrated with metadata analyzed using tree-based classifiers to enhance diagnostic accuracy. This study applies advanced sampling and stratified group cross-validation to address dataset class imbalance, ensuring broad model generalization across patient groups. The model achieves a partial area under the ROC curve (pAUC) of 0.18380 on cross-validation and 0.17295 on the private test set, maintaining high true positive rates (TPR \(\ge \) 80%). It ranks first for pAUC on the private test set, demonstrating potential for early detection and improved patient outcomes.

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

A Multimodal Approach Integrating Images and Their Metadata for Skin Cancer Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Lam Hung Nguyen,
  • Thang Cap,
  • Huong Bui,
  • Tuong Le

摘要

This study introduces a multimodal model for early skin cancer detection, leveraging image data and metadata from the ISIC 2024 dataset. Our method integrates CNNs and Transformer-based models to extract features from single-lesion images cropped from 3D Total Body Photographs (3D-TBP). These images, resembling close-up smartphone photos, are integrated with metadata analyzed using tree-based classifiers to enhance diagnostic accuracy. This study applies advanced sampling and stratified group cross-validation to address dataset class imbalance, ensuring broad model generalization across patient groups. The model achieves a partial area under the ROC curve (pAUC) of 0.18380 on cross-validation and 0.17295 on the private test set, maintaining high true positive rates (TPR \(\ge \) 80%). It ranks first for pAUC on the private test set, demonstrating potential for early detection and improved patient outcomes.