Scholarlink

<p>Accurate diagnosis of skin lesion diseases is challenging, particularly for newly emerging conditions such as Monkeypox, due to their high visual similarity with other skin disorders. This study addresses this issue by leveraging transfer learning techniques to classify Monkeypox using three open-access datasets containing Monkeypox and visually similar lesions. Two hybrid approaches are proposed: one combines NASNetMobile and ResNet50V2 with a Prototypical Network for few-shot learning, and the other integrates these CNN backbones with a Vision Transformer (ViT). Prototypical Networks are effective for low-data scenarios by creating representative prototypes that support generalization, while ViT captures global context using self-attention. NASNetMobile and ResNet50V2 are chosen for their efficiency and strong feature representation. Experimental results show that the Prototypical Network-based method achieves outstanding accuracy (99.97%, 99.13%, 98.33%) on MSLD, MSID, and MSCI datasets, respectively. The ViT-based model also performs well, reaching over 90% accuracy on all datasets. These results demonstrate the effectiveness of combining CNNs and Transformers in differentiating Monkeypox from similar diseases. This study’s main contributions include validating the suitability of few-shot learning for real-world diagnostic systems, exploring hybrid architectures in data-limited scenarios, and confirming model generalizability across diverse datasets.</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

Hybrid few-shot learning and vision transformer approaches for the diagnosis of Monkeypox

  • 期刊论文

  • 发表日期：

    2025年10月21日

查看全文
  • Hilal Güven,
  • Ahmet Saygılı

摘要

Accurate diagnosis of skin lesion diseases is challenging, particularly for newly emerging conditions such as Monkeypox, due to their high visual similarity with other skin disorders. This study addresses this issue by leveraging transfer learning techniques to classify Monkeypox using three open-access datasets containing Monkeypox and visually similar lesions. Two hybrid approaches are proposed: one combines NASNetMobile and ResNet50V2 with a Prototypical Network for few-shot learning, and the other integrates these CNN backbones with a Vision Transformer (ViT). Prototypical Networks are effective for low-data scenarios by creating representative prototypes that support generalization, while ViT captures global context using self-attention. NASNetMobile and ResNet50V2 are chosen for their efficiency and strong feature representation. Experimental results show that the Prototypical Network-based method achieves outstanding accuracy (99.97%, 99.13%, 98.33%) on MSLD, MSID, and MSCI datasets, respectively. The ViT-based model also performs well, reaching over 90% accuracy on all datasets. These results demonstrate the effectiveness of combining CNNs and Transformers in differentiating Monkeypox from similar diseases. This study’s main contributions include validating the suitability of few-shot learning for real-world diagnostic systems, exploring hybrid architectures in data-limited scenarios, and confirming model generalizability across diverse datasets.