Scholarlink

<p>Skin cancer is a significant global health issue, with early and accurate detection being essential for effective treatment and improved patient outcomes. In recent years, multimodal machine learning approaches that combine dermoscopic images with patient metadata have shown considerable promise in automating skin lesion classification. However, challenges such as high intra-class variation, inter-class similarity, and the need for precise localization of clinically relevant features continue to hinder performance. In this work, we introduce a multimodal approach that processes both dermoscopic images and patient metadata. Prior to training, the images undergo preprocessing using Contrast Limited Adaptive Histogram Equalization (CLAHE) and data augmentation techniques to enhance contrast and increase sample diversity. The preprocessed images are passed through a multi-scale gated residual block, which integrates a spatial gating mechanism to extract hierarchical, multi-scale features while preserving fine spatial details. A soft attention mechanism is then applied to emphasize clinically relevant regions. The resulting image features are fused with patient metadata, with various fusion techniques tested to assess their impact on classification performance and identify the most effective configuration. We evaluated our approach on the widely used HAM10000 dataset. Experimental results demonstrate that our model, using element-wise multiplication as the primary fusion method, achieves an average accuracy of 94.02% and an AUC of 99.31%.</p>

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 novel multimodal deep learning architecture integrating a multi-scale gated residual block and soft attention for skin lesion classification

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • Mohamed Amine Ibrahimi,
  • Khalid Abbad,
  • Rachid Ben Abbou

摘要

Skin cancer is a significant global health issue, with early and accurate detection being essential for effective treatment and improved patient outcomes. In recent years, multimodal machine learning approaches that combine dermoscopic images with patient metadata have shown considerable promise in automating skin lesion classification. However, challenges such as high intra-class variation, inter-class similarity, and the need for precise localization of clinically relevant features continue to hinder performance. In this work, we introduce a multimodal approach that processes both dermoscopic images and patient metadata. Prior to training, the images undergo preprocessing using Contrast Limited Adaptive Histogram Equalization (CLAHE) and data augmentation techniques to enhance contrast and increase sample diversity. The preprocessed images are passed through a multi-scale gated residual block, which integrates a spatial gating mechanism to extract hierarchical, multi-scale features while preserving fine spatial details. A soft attention mechanism is then applied to emphasize clinically relevant regions. The resulting image features are fused with patient metadata, with various fusion techniques tested to assess their impact on classification performance and identify the most effective configuration. We evaluated our approach on the widely used HAM10000 dataset. Experimental results demonstrate that our model, using element-wise multiplication as the primary fusion method, achieves an average accuracy of 94.02% and an AUC of 99.31%.