基于深度学习的人工智能技术在肠道准备评估中的应用

Application of deep learning‑based artificial intelligence technology in bowel preparation assessment

摘要:

目的 探究基于深度学习的人工智能肠道准备评估系统（e‑Boston bowel preparation scale system，e‑BBPS系统）评分与腺瘤漏检率之间的关系。方法 回顾性收集2017‑12‑21至2019‑12‑31在武汉大学人民医院内镜中心结肠镜检查的4 373例患者肠镜图像用于模型训练。前瞻性纳入2021‑10‑08至2022‑11‑09在中山大学附属第八医院行结肠镜检查的患者。使用e‑BBPS系统评估患者肠道准备情况，内镜医师根据BBPS评分进行肠道准备评估。如果内镜医师和e‑BBPS都认为肠道准备充分，患者立即接受第2次结肠镜检查，否则患者在第2次肠镜检查前重新进行肠道准备。比较e‑BBPS系统评估肠道准备合格组（e‑BBPS评分≤3分）与肠道准备不合格组（e‑BBPS评分>3分）之间腺瘤和息肉病灶漏诊率的差异。结果 肠道准备合格组患者腺瘤病灶漏检率明显小于肠道准备不合格组［26.72%（62/232）比42.53%（37/87），χ2=7.384，P=0.007，OR=2.029（95%CI：1.212~3.396）］。同时，肠道准备合格组患者息肉漏检率明显小于肠道准备不合格组［27.28%（195/702）比41.24%（113/274），χ2=16.539，P<0.001，OR=1.825（95%CI：1.363~2.443）］。结论 基于深度学习的e‑BBPS系统在肠道准备评估方面具有准确性和可靠性，有望标准化肠道准备评估流程，减少病灶漏检。

Abstract:

Objective To investigate the correlationship between an artificial intelligence‑based e‑Boston bowel preparation scale (e‑BBPS) system score and the adenoma miss rate. Methods Colonoscopy images of 4 373 patients at the Endoscopy Center of Renmin Hospital of Wuhan University from December 21, 2017 to December 31, 2019 were collected for model training. Patients who underwent colonoscopy at the Eighth Affiliated Hospital of Sun Yat‑sen University from October 8, 2021 to November 9, 2022 were prospectively included. Patient''s bowel preparation was evaluated by the e‑BBPS system and endoscopists based on BBPS score. If both the endoscopists and e‑BPPS system believed that the bowel preparation was sufficient, the patient immediately proceeded to a second colonoscopy. Otherwise, the patient underwent bowel preparation again. The differences in adenoma and polyp miss rate between the qualified group (e‑BBPS system score ≤3) and the unqualified group (e‑BBPS system score >3) were compared. Results The adenoma miss rate in the qualified group was significantly lower than that in the unqualified group [26.72% (62/232) VS 42.53% (37/87), χ2=7.384，P=0.007, OR=2.029 (95%CI: 1.212‑3.396)], and the polyp miss rate in the qualified group was significantly lower than that in the unqualified group [27.28% (195/702) VS 41.24% (113/274), χ2=16.539，P<0.001, OR=1.825 (95%CI: 1.363‑2.443)]. Conclusion The deep learning‑based e‑BBPS system demonstrates accuracy and reliability in bowel preparation assessment, offering potential to standardize the process of evaluating bowel preparation and reduce missed lesions.