Scholarlink

Background <p>Artificial intelligence (AI) is a promising tool for pancreatic disease diagnosis using Endoscopic Ultrasound (EUS) images. But, current models often fail to fully account for real-world clinical applicability. To address this limitation, we propose the multi-label, dynamic AI system, designed to mimic real-world physician assessments.</p> Methods <p>We included 2 783 patients (330 706 EUS images) from three cohorts: FUSCC training (<i>n</i> = 2 498), FUSCC internal testing (<i>n</i> = 178), and external testing (<i>n</i> = 107). The AI-Enhanced Pancreatic Multi-Disease Diagnostic System with EUS (AI-Paradise) integrates module 1 (image type classification), module 2 (image quality control), and module 3 (multi-label classification). A computer-assisted diagnostic test (CADT) assessed the diagnostic performance of endoscopists with AI-Paradise assistance.</p> Results <p>Module 1 and Module 2 achieved mean accuracies of 79·0 and 94·7%, respectively. These modules filtered out low-quality images, selecting 81 540 B-mode images for Module 3. In internal cross-validation, the best area under the curve (AUC) for six pancreatic diseases ranged from 71·5 to 87·6%. Module 3 demonstrated strong per-disease diagnostic performance in image-level testing, with accuracies ranging from 73.3 to 85.5% for the six pancreatic diseases (Table&#xa0;<InternalRef RefID="Tab2">2</InternalRef>). The overall patient-level correct diagnosis rates, which are secondary summary metrics, were 66.9% (internal) and 63.6% (external). In CADT, performance of novice endoscopists significantly improved, with the best-performing novice achieving an increase from 39·4 to 57·4% (<i>p</i> &lt; 0·0001).</p> Conclusions <p>AI-Paradise enhances diagnostic performance by assisting endoscopists in filtering out low-quality images and making accurate multiple-disease diagnoses. </p> Graphical abstract <p></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

Multi-label dynamic diagnosis of pancreatic diseases using AI-enhanced endoscopic ultrasound: a multi-cohort real-world study

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Ke Chen,
  • Kai Zhang,
  • Chunbing Zhu,
  • Chao Zhang,
  • Xiangpeng Hu,
  • Zhixi Li,
  • Jing Du,
  • Qianqian Fang,
  • Qijie Rui,
  • Jianwei Qi,
  • Bin Yao,
  • Lingyu Zhang,
  • Liting Zhang,
  • Yuan Liu,
  • Jin Xu,
  • Xianjun Yu,
  • Si Shi

摘要

Background

Artificial intelligence (AI) is a promising tool for pancreatic disease diagnosis using Endoscopic Ultrasound (EUS) images. But, current models often fail to fully account for real-world clinical applicability. To address this limitation, we propose the multi-label, dynamic AI system, designed to mimic real-world physician assessments.

Methods

We included 2 783 patients (330 706 EUS images) from three cohorts: FUSCC training (n = 2 498), FUSCC internal testing (n = 178), and external testing (n = 107). The AI-Enhanced Pancreatic Multi-Disease Diagnostic System with EUS (AI-Paradise) integrates module 1 (image type classification), module 2 (image quality control), and module 3 (multi-label classification). A computer-assisted diagnostic test (CADT) assessed the diagnostic performance of endoscopists with AI-Paradise assistance.

Results

Module 1 and Module 2 achieved mean accuracies of 79·0 and 94·7%, respectively. These modules filtered out low-quality images, selecting 81 540 B-mode images for Module 3. In internal cross-validation, the best area under the curve (AUC) for six pancreatic diseases ranged from 71·5 to 87·6%. Module 3 demonstrated strong per-disease diagnostic performance in image-level testing, with accuracies ranging from 73.3 to 85.5% for the six pancreatic diseases (Table 2). The overall patient-level correct diagnosis rates, which are secondary summary metrics, were 66.9% (internal) and 63.6% (external). In CADT, performance of novice endoscopists significantly improved, with the best-performing novice achieving an increase from 39·4 to 57·4% (p < 0·0001).

Conclusions

AI-Paradise enhances diagnostic performance by assisting endoscopists in filtering out low-quality images and making accurate multiple-disease diagnoses.

Graphical abstract