Scholarlink

<p>Prototypical Networks have proven effective for metric-based few-shot learning, enabling models to generalize from limited labeled examples. However, their application to medical image analysis remains challenging due to the complex, high-variability nature of medical images and the need for robust, domain-specific feature extraction. This work proposes Prototypical Aggregate Network (PANet), an enhanced variant designed specifically for few-shot medical image classification. PANet addresses two key challenges: (a) it incorporates spectral components using Discrete Wavelet Transform (DWT) to explicitly capture texture and frequency information relevant for pathology localization, and (b) it mitigates information loss by aggregating intermediate feature embeddings via depth-wise averaging, allowing downstream layers to benefit from earlier-layer morphological cues. PANet outperforms several state-of-the-art few-shot learning models in low-shot settings, achieving average accuracy gains of 2.8% and 1.55% in 2-way 3-shot and 5-shot classification tasks, respectively, on the BreakHis dataset. Furthermore, it demonstrates statistically significant improvements over comparable architectures and achieves competitive performance against much deeper and more complex models on the PathMNIST and BloodMNIST datasets. Qualitative results using explainability methods further validate PANet’s capability to localize and distinguish subtle morphological patterns, enhancing interpretability and supporting its potential for real-world clinical deployment.</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

Prototypical aggregate network - boosting few-shot learning for medical image classification

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Ranjana Roy Chowdhury,
  • Usma Niyaz,
  • Deepti R. Bathula

摘要

Prototypical Networks have proven effective for metric-based few-shot learning, enabling models to generalize from limited labeled examples. However, their application to medical image analysis remains challenging due to the complex, high-variability nature of medical images and the need for robust, domain-specific feature extraction. This work proposes Prototypical Aggregate Network (PANet), an enhanced variant designed specifically for few-shot medical image classification. PANet addresses two key challenges: (a) it incorporates spectral components using Discrete Wavelet Transform (DWT) to explicitly capture texture and frequency information relevant for pathology localization, and (b) it mitigates information loss by aggregating intermediate feature embeddings via depth-wise averaging, allowing downstream layers to benefit from earlier-layer morphological cues. PANet outperforms several state-of-the-art few-shot learning models in low-shot settings, achieving average accuracy gains of 2.8% and 1.55% in 2-way 3-shot and 5-shot classification tasks, respectively, on the BreakHis dataset. Furthermore, it demonstrates statistically significant improvements over comparable architectures and achieves competitive performance against much deeper and more complex models on the PathMNIST and BloodMNIST datasets. Qualitative results using explainability methods further validate PANet’s capability to localize and distinguish subtle morphological patterns, enhancing interpretability and supporting its potential for real-world clinical deployment.