Scholarlink

<p>Histopathological identification of tumor tissue is one of the routine pathological diagnoses of breast cancer. Recently, pathological examination has been successfully completed by a variety of deep learning-based applications. Nevertheless, due to the huge size of whole slide images (WSI), images are split into small patches that are analyzed independently in most existing approaches, which ignores the spatial correlations among them. In this paper, a novel model based on Transformer framework named AMNet is proposed to obtain breast cancer classification results and malignant tumor location more precisely. The adaptive feature fusion module (AFFM) and the mean value conditional random field (MVCRF) are embedded to train the whole model end-to-end with standard back-propagation algorithm. AFFM combines the self-attention module and the multi-receptive-field convolution module to get multi-scale semantic features and enhance the feature extraction capability from both global and local perspectives. MVCRF considers the spatial correlations between neighboring patches to obtain morphological information between pathological tissues. It’s demonstrated that proposed AMNet outperforms the baseline in breast cancer metastasis detection on BreakHis and Camelyon16 datasets. On patch-level images, it achieves the accuracy up to 94.5%, while on WSIs, the accuracy reaches 97.45% on AUC and 81.02% on FROC.</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

An End-to-end learning for classification and segmentation of breast cancer

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Jinling Chen,
  • Jie Chen,
  • Zhanbo Tan,
  • Zhuowei Tang,
  • Jihong Wei,
  • Qi ke,
  • Yuzhu Ji,
  • Ziqing Gao

摘要

Histopathological identification of tumor tissue is one of the routine pathological diagnoses of breast cancer. Recently, pathological examination has been successfully completed by a variety of deep learning-based applications. Nevertheless, due to the huge size of whole slide images (WSI), images are split into small patches that are analyzed independently in most existing approaches, which ignores the spatial correlations among them. In this paper, a novel model based on Transformer framework named AMNet is proposed to obtain breast cancer classification results and malignant tumor location more precisely. The adaptive feature fusion module (AFFM) and the mean value conditional random field (MVCRF) are embedded to train the whole model end-to-end with standard back-propagation algorithm. AFFM combines the self-attention module and the multi-receptive-field convolution module to get multi-scale semantic features and enhance the feature extraction capability from both global and local perspectives. MVCRF considers the spatial correlations between neighboring patches to obtain morphological information between pathological tissues. It’s demonstrated that proposed AMNet outperforms the baseline in breast cancer metastasis detection on BreakHis and Camelyon16 datasets. On patch-level images, it achieves the accuracy up to 94.5%, while on WSIs, the accuracy reaches 97.45% on AUC and 81.02% on FROC.