摘要目的:研发基于深度卷积神经网络(DCNN)的综合胃镜智能质控系统，并前瞻性评估该系统的临床可行性。方法:针对胃镜检查中胃黏膜观察完整度、胃黏膜可视度、胃镜检查时长、胃癌可疑病变检测等质控目标分别设计胃镜扫查部位识别模型、胃黏膜可视度识别模型、体内外识别模型和胃癌检测模型4个DCNN模型。通过多中心回顾性纳入98 385张白光胃镜图像进行模型训练与测试。计算各模型的准确度、灵敏度、特异度，并绘制ROC曲线。整合模型形成多功能的综合智能质控系统，于山东大学齐鲁医院消化内镜中心前瞻性、连续纳入100例行常规胃镜检查的患者，进一步评估智能质控系统在实际临床应用中的可行性。记录系统各质控功能的运行情况(平均错误提示或正确率)和检查结束后病变检出情况。结果:胃镜扫查部位识别模型识别各部位的准确度、灵敏度、特异度分别为98.40%～99.85%、61.95%～100.00%、98.65%～100.00%，ROC AUC值为0.997 6～1.000 0；胃黏膜可视度识别模型识别黏膜可视度的准确度、灵敏度、特异度分别为97.02%～98.27%、85.14%～99.28%、93.72%～100.00%；体内外识别模型判断体内外的准确度、灵敏度、特异度分别为97.27%、99.85%、94.50%，ROC AUC值为0.961 5；胃癌检测模型检测胃癌的准确度、灵敏度、特异度分别为95.92%、95.64%、96.05%，ROC AUC值为0.975 9。智能质控系统可行性评估结果示，在胃黏膜观察完整度质控中，系统平均错误填充0.32次/例；在胃黏膜可视度质控中，系统平均错误提示0.47次/例；胃镜检查时长智能计时正确率为96.00%；在胃癌可疑病变检测中，系统平均错误提示0.36次/例。最终共检出胃癌3例，胃高级别上皮内瘤变1例，系统均可准确识别、定位。结论:胃镜智能质控系统可在实际检查中对胃黏膜观察完整度、胃黏膜可视度、胃镜检查时长、胃癌可疑病变检测进行准确质控，使精准、高效的胃镜质控成为可能。

abstractsObjective:To develop intelligent quality-control system (IQCS) based on deep convolutional neural networks (DCNN), and to prospectively evaluate the clinical feasibility of this system.Methods:Aimed at quality control objectives during gastroscopy such as the observation integrity of gastric mucosal, gastric mucosa visibility, time spent on gastroendoscopy and suspicious gastric cancer detection, four DCNN models including gastroscopic scanning location recognition model, gastric mucosa visibility recognition model, in vivo and in vitro identification model and gastric cancer detection model were designed. A total of 98 385 white light gastroscopy images were retrospectively collected from multiple centers for training and testing the DCNN models. The accuracy, sensitivity and specificity of each model were calculated and the receiver operating characteristic (ROC) curve was drawn. The models were integrated and formed the multi-function integrated IQCS. At the center of gastroendoscopy, Qilu Hospital of Shandong University, 100 consecutive patients who underwent routine gastroscopy were prospectively enrolled. The feasibility of IQCS in real clinical practice was evaluated. The condition of each quality control function of the system (average error point out or correct rate) and the detection of lesions after the examination were recorded. Results:The accuracy, sensitivity and specificity of the model of gastroscopic scanning location recognition to identify each site were 98.40% to 99.85%, 61.95% to 100.00% and 98.65% to 100.00%, respectively; the area under curve (AUC) of ROC curve ranged from 0.997 6 to 1.000 0. The accuracy, sensitivity and specificity of the model of gastric mucosa visibility recognition to identity the mucosal visibility were 97.02% to 98.27%, 85.14% to 99.28% and 93.72% to 100.00%, respectively. The accuracy, sensitivity and specificity of the model of in vivo and in vitro identification were 97.27%, 99.85% and 94.50%, respectively; the AUC of ROC was 0.961 5. The accuracy, sensitivity and specificity of the model of gastric cancer detection were 95.92%, 95.64% and 96.05%, respectively; the AUC of ROC was 0.975 9. The results of feasibility evaluation of IQCS indicated that in the quality control of gastric mucosa observation integrity, the system average error was 0.32 time/case; in the quality control of mucosal visibility, the system average error was 0.47 time/case; the correct rate of intelligent timing during gastroscopy was 96.00%, in the quality control of suspicious gastric cancer detection, the system average error was 0.36 time/case. A total of 3 cases of gastric cancer and 1 case of high grade gastric intraepithelial neoplasia were detected. The system could accurately identify the location. Conclusions:Gastroscopy IQCS can accurately achieve quality control in the observation integrity of gastric mucosa, gastric mucosa visibility, time spent on gastroendoscopy and suspicious gastric cancer detection in actual examination, which makes accurate and efficient gastroscopy quality control possible.