Scholarlink

<p>Multiple Instance Learning (MIL) has become a cornerstone for Whole Slide Image (WSI) analysis, yet existing methods often struggle with feature redundancy and the neglect of slide-level global context. In this study, an enhanced MIL framework is proposed, incorporating a plug-and-play Context-Aware Feature Refinement (CAFR) module and a geometric Orthogonal Regularization mechanism. The CAFR module employs a “Squeeze-Excitation-Refine” strategy to capture bag-level context prototypes and dynamically recalibrate instance features, while the orthogonal regularization enforces inter-channel decorrelation to learn diverse pathological semantics. Extensive experiments on Camelyon16, Camelyon17, and TCGA-NSCLC datasets demonstrate that the proposed approach consistently improves the performance of established backbones, including ABMIL, TransMIL, and CLAM. Notably, when integrated with the proposed module, classic architectures achieve robust performance comparable to recently proposed complex methods, demonstrating superior generalization capabilities. Furthermore, this significant performance boost was achieved with a negligible increase in FLOPs of approximately 0.58%, making the proposed framework a highly efficient solution for precise WSI classification. The source code is available at <a href="https://github.com/zsszc/CAFR-MIL.git">https://github.com/zsszc/CAFR-MIL.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

Context-aware feature refinement with orthogonal regularization for whole slide image classification

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Shicheng Zhou,
  • Zefeng Wang,
  • Jikai Yu,
  • Boyuan Wu,
  • Jiayun Zhu

摘要

Multiple Instance Learning (MIL) has become a cornerstone for Whole Slide Image (WSI) analysis, yet existing methods often struggle with feature redundancy and the neglect of slide-level global context. In this study, an enhanced MIL framework is proposed, incorporating a plug-and-play Context-Aware Feature Refinement (CAFR) module and a geometric Orthogonal Regularization mechanism. The CAFR module employs a “Squeeze-Excitation-Refine” strategy to capture bag-level context prototypes and dynamically recalibrate instance features, while the orthogonal regularization enforces inter-channel decorrelation to learn diverse pathological semantics. Extensive experiments on Camelyon16, Camelyon17, and TCGA-NSCLC datasets demonstrate that the proposed approach consistently improves the performance of established backbones, including ABMIL, TransMIL, and CLAM. Notably, when integrated with the proposed module, classic architectures achieve robust performance comparable to recently proposed complex methods, demonstrating superior generalization capabilities. Furthermore, this significant performance boost was achieved with a negligible increase in FLOPs of approximately 0.58%, making the proposed framework a highly efficient solution for precise WSI classification. The source code is available at https://github.com/zsszc/CAFR-MIL.git.