Scholarlink

<p>The challenge of accurate wound classification lies in distinguishing etiologies with high visual similarity, a task for which existing models often sacrifice efficiency for accuracy. Addressing this, we introduce ERYXNet, a novel architecture that is not a straightforward fusion but a carefully engineered integration through scaling and custom pruning of components from EfficientNet, ResNet, and the YOLOv8 classifier. ERYXNet is designed to capture multi-scale features efficiently, from localized textures to global context. When trained from scratch on a dataset encompassing nine wound types, a fake wound (e.g., tattoos and pigments), and normal skin, ERYXNet achieves classification accuracies of 95.32% and 92.31% in our experiments, setting a new state-of-the-art benchmark without employing transfer learning under the same conditions. It also proved more efficient, requiring fewer FLOPs and parameters than state-of-the-art alternatives. This balance of high accuracy and low computational demand makes ERYXNet a practical solution for real-world clinical applications.</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

ERYXNet: a lightweight yet robust architecture for multi-type wound classification

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Trishaani Acharjee,
  • Rajdeep Chatterjee,
  • Biswaranjan Acharya,
  • Fernando Moreira

摘要

The challenge of accurate wound classification lies in distinguishing etiologies with high visual similarity, a task for which existing models often sacrifice efficiency for accuracy. Addressing this, we introduce ERYXNet, a novel architecture that is not a straightforward fusion but a carefully engineered integration through scaling and custom pruning of components from EfficientNet, ResNet, and the YOLOv8 classifier. ERYXNet is designed to capture multi-scale features efficiently, from localized textures to global context. When trained from scratch on a dataset encompassing nine wound types, a fake wound (e.g., tattoos and pigments), and normal skin, ERYXNet achieves classification accuracies of 95.32% and 92.31% in our experiments, setting a new state-of-the-art benchmark without employing transfer learning under the same conditions. It also proved more efficient, requiring fewer FLOPs and parameters than state-of-the-art alternatives. This balance of high accuracy and low computational demand makes ERYXNet a practical solution for real-world clinical applications.