Scholarlink

<p>Diagnosing a whole-slide image is an interactive, multistage process, yet practical agentic systems that navigate fields, adjust magnification and deliver explainable diagnoses remain lacking, largely because the tacit, experience-based viewing behaviour of expert pathologists is absent from model training data. Here we introduce Pathology-CoT, a framework that converts expert viewing chain-of-thought behaviour into scalable agent supervision through three contributions. First, an artificial intelligence session recorder unobtrusively captures routine navigation in standard whole-slide image viewers and converts raw logs into standardized behavioural commands and bounding boxes. Second, a human-in-the-loop review pipeline turns artificial intelligence-drafted rationales into paired ‘where to look’ and ‘why it matters’ supervision, enabling sixfold faster labelling. Third, using these data, we built Pathology-o3, a two-stage agent that proposes regions of interest and performs behaviour-guided reasoning. On gastrointestinal lymph node metastasis detection, Pathology-o3 outperformed state-of-the-art vision–language models, showed consistent gains across multiple vision–language model backbones and maintained strong performance on an independent external validation cohort.</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

Pathology-CoT: learning visual chain-of-thought agents from expert whole-slide image diagnosis behaviour

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Sheng Wang,
  • Ruiming Wu,
  • Charles Herndon,
  • Songhao Li,
  • Yihang Liu,
  • Shunsuke Koga,
  • Xiaowei Xu,
  • David E. Elder,
  • Jonathan Alex Miles,
  • Annie Jin,
  • Ikuko Hirai,
  • Meaghan Dougher,
  • Jeanne Shen,
  • Zhi Huang

摘要

Diagnosing a whole-slide image is an interactive, multistage process, yet practical agentic systems that navigate fields, adjust magnification and deliver explainable diagnoses remain lacking, largely because the tacit, experience-based viewing behaviour of expert pathologists is absent from model training data. Here we introduce Pathology-CoT, a framework that converts expert viewing chain-of-thought behaviour into scalable agent supervision through three contributions. First, an artificial intelligence session recorder unobtrusively captures routine navigation in standard whole-slide image viewers and converts raw logs into standardized behavioural commands and bounding boxes. Second, a human-in-the-loop review pipeline turns artificial intelligence-drafted rationales into paired ‘where to look’ and ‘why it matters’ supervision, enabling sixfold faster labelling. Third, using these data, we built Pathology-o3, a two-stage agent that proposes regions of interest and performs behaviour-guided reasoning. On gastrointestinal lymph node metastasis detection, Pathology-o3 outperformed state-of-the-art vision–language models, showed consistent gains across multiple vision–language model backbones and maintained strong performance on an independent external validation cohort.