Scholarlink

<p>Gastrointestinal (GI) diseases pose a major clinical burden, yet conventional histopathology suffers from subjectivity and limited reproducibility. While existing computational pathology foundation models are often validated across many subspecialties in “broad but shallow” benchmarks, they rarely demonstrate deep clinical utility in real-world scenarios. To address this, we develop Digepath—a disease-specialized foundation model focused exclusively on high-impact GI pathology. Our approach employs a two-stage iterative optimization: first, pretraining on over 353 million multi-scale patches from 210,043 H&amp;E-stained slides; second, fine-tuning on 471,443 expert-annotated regions, balancing tumor and non-tumor samples to enhance lesion perception amid sparse pathology in whole-slide images. Digepath achieves state-of-the-art performance on 32 of 33 systematic downstream tasks in GI pathology—including diagnosis, molecular profiling, and survival prognosis—demonstrating robust generalization. Moreover, we integrate its capabilities into an agent-based clinical reasoning framework that supports end-to-end intelligent diagnostic workflows, paving the way for real-world deployment.</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

Subspecialty-specific foundation model for intelligent gastrointestinal pathology

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月4日

查看全文
  • Lianghui Zhu,
  • Xitong Ling,
  • Minxi Ouyang,
  • Xiaoping Liu,
  • Tian Guan,
  • Mingxi Fu,
  • Maomao Zeng,
  • Zhiqiang Cheng,
  • Fanglei Fu,
  • Qiang Huang,
  • Mingxi Zhu,
  • Yibo Jin,
  • Qiming He,
  • Yizhi Wang,
  • Junru Cheng,
  • Xuanyu Wang,
  • Luxi Xie,
  • Houqiang Li,
  • Sufang Tian,
  • Yonghong He

摘要

Gastrointestinal (GI) diseases pose a major clinical burden, yet conventional histopathology suffers from subjectivity and limited reproducibility. While existing computational pathology foundation models are often validated across many subspecialties in “broad but shallow” benchmarks, they rarely demonstrate deep clinical utility in real-world scenarios. To address this, we develop Digepath—a disease-specialized foundation model focused exclusively on high-impact GI pathology. Our approach employs a two-stage iterative optimization: first, pretraining on over 353 million multi-scale patches from 210,043 H&E-stained slides; second, fine-tuning on 471,443 expert-annotated regions, balancing tumor and non-tumor samples to enhance lesion perception amid sparse pathology in whole-slide images. Digepath achieves state-of-the-art performance on 32 of 33 systematic downstream tasks in GI pathology—including diagnosis, molecular profiling, and survival prognosis—demonstrating robust generalization. Moreover, we integrate its capabilities into an agent-based clinical reasoning framework that supports end-to-end intelligent diagnostic workflows, paving the way for real-world deployment.