Scholarlink

<p>Accurate diagnosis of dermatological diseases remains challenging due to fine-grained inter-class similarities, high intra-class variability, and the scarcity of annotated data. While Few-Shot Learning (FSL) offers a promising solution, existing methods often struggle with biased prototype estimation and semantic ambiguity. To address these issues, we propose the Semantic Feature Enhanced Prototypical Transformer (SFEPT), a unified framework for few-shot dermatological image classification. Guided by the philosophy of discriminant-oriented feature refinement, SFEPT employs a Swin Transformer backbone to extract hierarchical features, followed by a sequential refinement pipeline. First, a non-iterative prototype rectification strategy mitigates distribution bias and refines class centroids via confidence-aware pseudo-labeling. Subsequently, a Semantic-Aware Enhancement (SAE) module enriches query representations by projecting them into a latent semantic manifold via a frozen BERT decoder. Notably, our SAE module functions as a latent semantic projector, injecting rich linguistic priors while obviating the dependency on explicit textual inputs during inference, thereby maintaining high efficiency. Evaluations on three dermatological benchmarks (SD-198, Derm7pt, and skinl_cps) demonstrate the superiority and robustness of SFEPT, achieving 89.53% and 81.85% accuracy on SD-198 and Derm7pt, respectively, in 3-way 5-shot tasks. The code is available at <a href="https://github.com/AIYAU/FSL_dermatopathya">https://github.com/AIYAU/FSL_dermatopathya</a>.</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

Semantic-enhanced few-shot learning for precise dermatological diagnosis

  • 期刊论文

  • 发表日期：

    2026年4月8日

查看全文
  • Yuxiang Wang,
  • Jinrong He,
  • Shuai Ma

摘要

Accurate diagnosis of dermatological diseases remains challenging due to fine-grained inter-class similarities, high intra-class variability, and the scarcity of annotated data. While Few-Shot Learning (FSL) offers a promising solution, existing methods often struggle with biased prototype estimation and semantic ambiguity. To address these issues, we propose the Semantic Feature Enhanced Prototypical Transformer (SFEPT), a unified framework for few-shot dermatological image classification. Guided by the philosophy of discriminant-oriented feature refinement, SFEPT employs a Swin Transformer backbone to extract hierarchical features, followed by a sequential refinement pipeline. First, a non-iterative prototype rectification strategy mitigates distribution bias and refines class centroids via confidence-aware pseudo-labeling. Subsequently, a Semantic-Aware Enhancement (SAE) module enriches query representations by projecting them into a latent semantic manifold via a frozen BERT decoder. Notably, our SAE module functions as a latent semantic projector, injecting rich linguistic priors while obviating the dependency on explicit textual inputs during inference, thereby maintaining high efficiency. Evaluations on three dermatological benchmarks (SD-198, Derm7pt, and skinl_cps) demonstrate the superiority and robustness of SFEPT, achieving 89.53% and 81.85% accuracy on SD-198 and Derm7pt, respectively, in 3-way 5-shot tasks. The code is available at https://github.com/AIYAU/FSL_dermatopathya.