中国全科医学 ›› 2026, Vol. 29 ›› Issue (14): 1873-1877.DOI: 10.12114/j.issn.1007-9572.2025.0305

• 论著 • 上一篇    

基于深度学习的结直肠息肉实时测量系统测量结直肠息肉尺寸的验证研究

许庆洪1, 王静2, 黄元1, 祝辞1, 操文兵1, 李盈1,*()   

  1. 1.430012 湖北省武汉市第八医院消化内镜中心
    2.430060 湖北省武汉市,武汉大学人民医院消化内科
  • 收稿日期:2025-08-07 修回日期:2025-12-21 出版日期:2026-05-15 发布日期:2026-04-14
  • 通讯作者: 李盈

  • 作者贡献:

    许庆洪提出主要研究目标,负责研究的构思与设计,研究的实施,撰写论文;王静负责研究资料以及各项数据的收集整理及统计学分析;黄元、祝辞负责数据的校对、表格及图片的编辑与整理;操文兵负责论文修订、文章的质量控制及审校;李盈负责文章的构思与设计,最终版本修订,对文章整体负责。

  • 基金资助:
    武汉市第八医院医学科学研究项目(YX24Y08)

A Validation Study on Measuring Colorectal Polyp Size Using a Deep Learning-based Real-time Colorectal Polyp Measurement System

XU Qinghong1, WANG Jing2, HUANG Yuan1, ZHU Ci1, CAO Wenbing1, LI Ying1,*()   

  1. 1. Endoscopy Center, the Eighth Hospital of Wuhan, Wuhan 430012, China
    2. Department of Gastroenterology, Renmin Hospital of Wuhan University, Wuhan 430060, China
  • Received:2025-08-07 Revised:2025-12-21 Published:2026-05-15 Online:2026-04-14
  • Contact: LI Ying

摘要: 背景 结肠镜下息肉切除术可显著降低结直肠癌的患病风险,适宜的息肉切除手术方式和术后的监测间隔主要取决于结肠镜检查中对结肠息肉大小测量的准确率。 目的 评估基于深度学习的结直肠息肉实时测量系统(ENDOANGEL-CPS系统)在白光内镜下测量结直肠息肉尺寸的效能。 方法 利用ENDOANGEL-CPS系统评估息肉的深度,结合结肠镜的透镜参数自动计算息肉的直径,该系统通过虚拟内镜视频训练深度评估模型完成开发训练,并通过7个结肠视频的730张图像进行了测试;在9个粘贴模型息肉的模拟结肠视频中评估了该系统的性能。收集武汉市第八医院2022年6—8月已知结直肠息肉且行内镜下治疗或切除潜在的结直肠息肉患者的36个内镜视频影像,在收集到的真实内镜检查视频中完成ENDOANGEL-CPS系统验证,并同时评估5名内镜医师(1名高级医师、2名中级医师、2名初级医师)测量息肉大小的准确性,以游标卡尺测量内镜切除后息肉的尺寸为金标准。 结果 ENDOANGEL-CPS系统测量值与真实值之间的一致性相关系数(CCC)为0.84(95%CI=0.67~0.89),ENDOANGEL-CPS系统测量的平均相对误差为17.94%(0,25.00%)。所有医师测量的中位息肉尺寸与真实值之间的CCC为0.45(95%CI=0.25~0.56),平均相对误差为35.19%(16.67%,50.00%);初级医师测量的中位息肉尺寸与真实值之间的CCC为0.84(95%CI=0.67~0.89),平均相对误差为40.00%(25.00%,50.00%);中级医师测量的中位息肉尺寸与真实值之间的CCC为0.32(95%CI=0.11~0.45),平均相对误差为26.79%(16.07%,50.00%);高级医师测量的中位息肉尺寸与真实值之间的CCC为0.74(95%CI=0.40~0.82),平均相对误差为25.66%(8.33%,35.00%)。 结论 ENDOANGEL-CPS系统有望提高结直肠息肉尺寸测量的准确性,为基于息肉尺寸制定的术后监测方案提供重要参考依据。

关键词: 结直肠息肉, 深度学习, 人工智能

Abstract:

Background

Colonoscopic polypectomy significantly reduces the incidence of colorectal cancer. The choice of optimal resection technique and subsequent surveillance intervals is largely dependent on the accurate in-vivo sizing of colorectal polyps during colonoscopy.

Objective

To evaluate the performance of a deep learning-based real-time colorectal polyp measurement system (ENDOANGEL-CPS) in measuring colorectal polyp size under white-light endoscopy.

Methods

The ENDOANGEL-CPS system automatically calculated polyp diameter by estimating polyp depth in conjunction with colonoscope lens parameters. The depth estimation model was developed and trained on virtual endoscopic videos, and initially tested on 730 images extracted from 7 colon videos. Further evaluation was conducted on 9 simulated colon videos with affixed model polyps. For clinical validation, 36 endoscopic video sequences were prospectively collected from patients undergoing endoscopic resection of known or suspected colorectal polyps at the Eighth Hospital of Wuhan between June and August 2022. The measurement accuracy of the ENDOANGEL-CPS system was validated against the gold standard, defined as the measurement of resected polyps using a vernier caliper. Concurrently, the sizing accuracy of five endoscopists (one senior physician, two intermediate physicians, and two primary physicians) was assessed.

Results

The measurements from the ENDOANGEL-CPS system demonstrated a concordance correlation coefficient (CCC) of 0.84 (95%CI=0.67-0.89) with the gold standard, and a mean relative error of 17.94% (0, 25.00%). In contrast, the CCC between the median size estimated by all endoscopists and the true value was 0.45 (95%CI=0.25-0.56), with a median relative error of 35.19% (16.67%, 50.00%). When stratified by experience, junior endoscopists achieved a CCC of 0.84 (95%CI=0.67-0.89) with a median relative error of 40.00% (25.00%, 50.00%). Intermediate-level endoscopists had a CCC of 0.32 (95%CI=0.11-0.45) and a median relative error of 26.79% (16.07%, 50.00%). The expert endoscopist showed a CCC of 0.74 (95%CI=0.40-0.82) and a median relative error of 25.66% (8.33%, 35.00%).

Conclusion

The ENDOANGEL-CPS system shows promise for improving the accuracy of colorectal polyp size measurement during colonoscopy. This could provide a more reliable basis for guiding post-polypectomy surveillance strategies tailored to polyp dimensions.

Key words: Colorectal polyp, Deep learning, Artificial intelligence