我们旨在开发一个卷积神经网络(CNN)模型，在实时数字单操作员胆管镜检查(DSOC)期间检测肿瘤病变，并通过与DSOC专家和非专家内镜医生的比较对该模型进行临床验证。在这个两阶段的研究中，我们首先开发和验证了CNN1。然后，我们进行了多中心诊断试验，将四名DSOC专家和非专家与改进的模型(CNN2)进行比较。他们根据Carlos Robles-Medranda (CRM)和Mendoza分解标准对病变进行了分组，肿瘤病变的最终诊断基于组织病理学和十二个月的随访结果。在第一阶段，CNN2实现了0.88的mAP、83.24%的IoU和0.0975的总损失。为了临床验证，我们对新加入的170个患者记录的视频进行了CNN2分析。其中50%的病例有肿瘤病变。该模型在肿瘤诊断病例中取得了显著的准确率值，灵敏度为90.5%，特异度为68.2%，阳、阴性预测值分别为74.0%和87.8%。CNN2模型优于非专家#2([ROC-CRM:0.657 vs ROC-CNN2:0.794,P<0.05 and ROC-Mendoza:0.657 vs ROC-CNN2:0.794,P<0.05)、非专家#4(ROC-CRM:0.683 vs. ROC-CNN2:0.791,P<0.05)和专家#4 ([ROC-CRM:0.755 vs ROC-CNN2:0.848,P<0,05 and ROC-Mendoza:0.753 vs ROC-CNN2:0.848,P<0,05)。该CNN模型可以准确地区分胆道恶性病变，并在两名非专家和一名专家内镜医生中表现出更好的效果。作者：本文是Thieme根据创造共享署名-非衍生-非商业性许可下发布的开放获取文章，允许复制和再生，只要原始工作得到适当的信用。内容不能用于商业目的，也不能进行改编、混合、转换或建立。 (https://creativecommons.org/licenses/by-nc-nd/4.0/)。

We aimed to develop a convolutional neural network (CNN) model for detecting neoplastic lesions during real-time digital single-operator cholangioscopy (DSOC) and clinically validated the model through comparisons with DSOC experts and nonexpert endoscopists.In this two-stage study, we first developed and validated CNN1. Then, we performed a multicenter diagnostic trial to compare four DSOC experts and nonexperts against an improved model (CNN2). They classified the lesions in neoplastic and nonneoplastic in accordance with Carlos Robles-Medranda (CRM) and Mendoza disaggregated criteria. The final diagnosis of neoplasia was based on histopathology and twelve-month follow-up outcomes.In stage I, CNN2 achieved a mAP of 0.88, an IoU of 83.24%, and a total loss of 0.0975. For clinical validation, a total of 170 videos from newly included patients were analyzed with the CNN2. 50% of cases had neoplastic lesions. This model achieved significant accuracy values for neoplastic diagnosis cases, with a 90.5% sensitivity, 68.2% specificity, and 74.0% and 87.8% positive and negative predictive values, respectively. CNN2 model outperformed nonexpert #2 ([ROC-CRM: 0.657 vs ROC-CNN2: 0.794, P<0.05 and ROC-Mendoza: 0.657 vs ROC-CNN2: 0.794, P<0.05]), nonexpert #4 (ROC-CRM: 0.683 vs ROC-CNN2: 0.791, P<0.05), and expert #4 ([ROC-CRM: 0.755 vs ROC-CNN2: 0.848, P<0,05 and ROC-Mendoza: 0.753 vs ROC-CNN2: 0.848, P<0,05]).The proposed CNN models distinguish neoplastic bile duct lesions with good accuracy and outperformed two nonexperts and one expert endoscopists.The Author(s). This is an open access article published by Thieme under the terms of the Creative Commons Attribution-NonDerivative-NonCommercial-License, permitting copying and reproduction so long as the original work is given appropriate credit. Contents may not be used for commercial purposes, or adapted, remixed, transformed or built upon. (https://creativecommons.org/licenses/by-nc-nd/4.0/).