Scholarlink

<p>Whole slide image (WSI) classification in computational pathology faces three critical challenges: (1) computational inefficiency from processing gigapixel images, (2) inadequate integration of multi-scale histopathological features, (3) manual ROI selection and pixel-level annotation are expensive. To address these limitations, we propose Multi-Scale Spatial Attention with Discriminative Instance Selection (MSSA-DIS), a dual-stage framework combining adaptive instance selection with multi-scale binary classification model using H&amp;E strains. The first stage uses a Discriminative Instance Selection (DIS) network with a dynamic threshold and an instance selection constraint to preserve the diagnostic critical area. Through active learning, discriminant instance selection is made based on multi-example pooling instance attention. The second stage introduces a multi-scale attention network and a Cross-Attention Attention-based Multiple Instance Learning (CA-AbMIL) pooling module. CA-AbMIL establishes cross-scale dependencies through weights learned by cross-attention. Multi-scale attention networks further enhance the discernibility of morphological features through spatial channel attention fusion. We also propose MaxEntHinge loss, combining entropy maximization, margin enforcement and other methods to improve robustness to noisy data. We conducted experiments on the public dataset CAMELYON16 and the private pituitary adenoma dataset ZPAD, respectively. This method achieves an accuracy rate of 95.96% on ZPAD, an increase of 9.94% compared with the existing method. At the same time, the accuracy rate of 93.06% and AUC of 98.38% on the CAMELYON16 data set has certain advantages compared with the existing methods. Our method can select 30% to 38% of high-quality instances for prediction through DIS, greatly reducing hardware indicators. Especially suitable for pathological workflows that require accuracy and limited hardware resource throughput.</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

MSSA-DIS: Multi-Scale Spatial Attention with Discriminative Instance Selection for Whole Slide Image Classification

  • 期刊论文

  • 发表日期：

    2025年6月2日

查看全文
  • Yi Lin,
  • Yunjiao Li,
  • Xiongbai Long,
  • Yanyan Ye,
  • Jing Guo,
  • Depei Li

摘要

Whole slide image (WSI) classification in computational pathology faces three critical challenges: (1) computational inefficiency from processing gigapixel images, (2) inadequate integration of multi-scale histopathological features, (3) manual ROI selection and pixel-level annotation are expensive. To address these limitations, we propose Multi-Scale Spatial Attention with Discriminative Instance Selection (MSSA-DIS), a dual-stage framework combining adaptive instance selection with multi-scale binary classification model using H&E strains. The first stage uses a Discriminative Instance Selection (DIS) network with a dynamic threshold and an instance selection constraint to preserve the diagnostic critical area. Through active learning, discriminant instance selection is made based on multi-example pooling instance attention. The second stage introduces a multi-scale attention network and a Cross-Attention Attention-based Multiple Instance Learning (CA-AbMIL) pooling module. CA-AbMIL establishes cross-scale dependencies through weights learned by cross-attention. Multi-scale attention networks further enhance the discernibility of morphological features through spatial channel attention fusion. We also propose MaxEntHinge loss, combining entropy maximization, margin enforcement and other methods to improve robustness to noisy data. We conducted experiments on the public dataset CAMELYON16 and the private pituitary adenoma dataset ZPAD, respectively. This method achieves an accuracy rate of 95.96% on ZPAD, an increase of 9.94% compared with the existing method. At the same time, the accuracy rate of 93.06% and AUC of 98.38% on the CAMELYON16 data set has certain advantages compared with the existing methods. Our method can select 30% to 38% of high-quality instances for prediction through DIS, greatly reducing hardware indicators. Especially suitable for pathological workflows that require accuracy and limited hardware resource throughput.