Scholarlink

<p>Image processing, particularly in the field of computational pathology, increasingly relies on scale information to improve diagnostic accuracy. Whole slide image (WSI) is inherently multi-scale in nature, yet most existing multi-scale approaches adopt static weighting schemes, restricting their ability to adapt to variations across instances. To address this limitation, we propose an Instance-based Self-adaptive Scale Fusion (ISSF) module, which dynamically assigns instance-specific weights to different scales, enabling a more fine-grained and flexible utilization of scale-related information. Unlike traditional strategies, ISSF learns optimal fusion coefficients via gradient-based optimization rather than relying on manually defined parameters, allowing the model to better capture scale diversity within pathology slides. We evaluate ISSF on two widely used public datasets, Camelyon16 and TCGA-NSCLC, where it achieves significant performance gains of up to 11.24% and 4.27% over single-scale baselines, and at least 2.57% and 2.05% over static fusion methods, respectively. These results demonstrate the effectiveness and generalizability of ISSF in computational pathology. The implementation is publicly available at: <a href="https://github.com/kikaiTollenge/instance-based-self-adaptive-scale-fusion.git">https://github.com/kikaiTollenge/instance-based-self-adaptive-scale-fusion.git</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

Enhancing whole slide image classification via fine-grained adaptive scale fusion

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Jikai Yu,
  • Zefeng Wang,
  • Shanshan Sun,
  • Boyuan Wu,
  • Shicheng Zhou,
  • Jiayun Zhu,
  • Lianxin Hu,
  • Yinhang Wu,
  • Shuwen Han

摘要

Image processing, particularly in the field of computational pathology, increasingly relies on scale information to improve diagnostic accuracy. Whole slide image (WSI) is inherently multi-scale in nature, yet most existing multi-scale approaches adopt static weighting schemes, restricting their ability to adapt to variations across instances. To address this limitation, we propose an Instance-based Self-adaptive Scale Fusion (ISSF) module, which dynamically assigns instance-specific weights to different scales, enabling a more fine-grained and flexible utilization of scale-related information. Unlike traditional strategies, ISSF learns optimal fusion coefficients via gradient-based optimization rather than relying on manually defined parameters, allowing the model to better capture scale diversity within pathology slides. We evaluate ISSF on two widely used public datasets, Camelyon16 and TCGA-NSCLC, where it achieves significant performance gains of up to 11.24% and 4.27% over single-scale baselines, and at least 2.57% and 2.05% over static fusion methods, respectively. These results demonstrate the effectiveness and generalizability of ISSF in computational pathology. The implementation is publicly available at: https://github.com/kikaiTollenge/instance-based-self-adaptive-scale-fusion.git