Scholarlink

<p>Distinguishing histological subtypes of early pregnancy loss is crucial for clinical practice. Current gold standards combining pathological and molecular profiling are costly, limiting implementation. We developed a patch-to-slide fusion artificial intelligence (AI) model with adaptive masking to predict histological diagnosis. A total of 1380 haematoxylin and eosin stained, 1057 p57 and 646 Ki-67 immunohistochemistry stained whole-slide images from 1287 patients were used for multicenter development and validation. Cases were classified as complete hydatidiform moles, partial hydatidiform moles, hydropic abortion, and normal control. The model achieved slide-level accuracy of 0.843 and AUROC of 0.959 in the development test, and accuracy of 0.801 with AUROC of 0.930 in independent testing. AI-assisted diagnosis significantly improved pathologist performance (<i>p</i> &lt; 0.05). The multi-stain model integrating haematoxylin and eosin and p57 outperformed single-stain models. This tool may improve diagnostic precision, assist triage for genetic testing, and reduce costs.</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

Patch-to-slide fusion deep learning model for histological diagnosis of early pregnancy loss including hydatidiform mole

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月18日

查看全文
  • Yating Zhao,
  • Xinhui He,
  • Xin Ye,
  • Zhe Wang,
  • Bo Huang,
  • Jiaqi Ning,
  • Limeng Cai,
  • Peilin Han,
  • Huifang Zhang,
  • Hao Chen,
  • Xiangang Yin,
  • Jie Dong,
  • Ping Duan,
  • Huafeng Shou,
  • Qingmei Han,
  • Kai Yan,
  • Jing Zhang,
  • Min Hu,
  • Ruiqi Sun,
  • Qihong Wan,
  • Xiupu Liu,
  • Qiwei Zheng,
  • Wanting Huang,
  • Haixia Luo,
  • Jue Xie,
  • Xiaofei Zhang,
  • Xiang Tian,
  • Jianhua Qian

摘要

Distinguishing histological subtypes of early pregnancy loss is crucial for clinical practice. Current gold standards combining pathological and molecular profiling are costly, limiting implementation. We developed a patch-to-slide fusion artificial intelligence (AI) model with adaptive masking to predict histological diagnosis. A total of 1380 haematoxylin and eosin stained, 1057 p57 and 646 Ki-67 immunohistochemistry stained whole-slide images from 1287 patients were used for multicenter development and validation. Cases were classified as complete hydatidiform moles, partial hydatidiform moles, hydropic abortion, and normal control. The model achieved slide-level accuracy of 0.843 and AUROC of 0.959 in the development test, and accuracy of 0.801 with AUROC of 0.930 in independent testing. AI-assisted diagnosis significantly improved pathologist performance (p < 0.05). The multi-stain model integrating haematoxylin and eosin and p57 outperformed single-stain models. This tool may improve diagnostic precision, assist triage for genetic testing, and reduce costs.