Scholarlink

<p>Breast cancer classification from histopathological images has been a major focus of recent deep learning research. However, the lack of model interpretability and high computational demands have limited clinical adoption. This paper proposes an explainable deep learning framework that integrates a ConvNeXt backbone, test-time augmentation (TTA), and attention rollout (AR) to achieve high-performance and transparent breast cancer diagnosis. Evaluated on the BreakHis dataset, consistently used across all related works for fair benchmarking. The proposed framework has achieved an accuracy of 98.7%, surpassing numerous related works, while effectively addressing critical challenges in explainability, generalization, and computational efficiency. By providing visual insights into the decision-making process, the framework enhances clinical trust and usability, representing a significant advancement toward next-generation AI-assisted pathology tools.</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

Explainable ConvNeXt framework with attention rollout for breast cancer classification from histopathological images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月9日

查看全文
  • Nattavut Sriwiboon

摘要

Breast cancer classification from histopathological images has been a major focus of recent deep learning research. However, the lack of model interpretability and high computational demands have limited clinical adoption. This paper proposes an explainable deep learning framework that integrates a ConvNeXt backbone, test-time augmentation (TTA), and attention rollout (AR) to achieve high-performance and transparent breast cancer diagnosis. Evaluated on the BreakHis dataset, consistently used across all related works for fair benchmarking. The proposed framework has achieved an accuracy of 98.7%, surpassing numerous related works, while effectively addressing critical challenges in explainability, generalization, and computational efficiency. By providing visual insights into the decision-making process, the framework enhances clinical trust and usability, representing a significant advancement toward next-generation AI-assisted pathology tools.