人工智能胃镜在盲区监测和自主图像采集中的应用研究

Application of artificial intelligence gastroscope in blind area monitoring and independent image acquisition

目的分析实时胃镜监控系统(即胃镜精灵)在胃镜检查中的盲区监测功能与自主图像采集功能。方法收集武汉大学人民医院消化内镜中心数据库2017年5月至2018年5月间的全部胃镜图像,根据入选排除标准,共选取38 522张胃镜图像训练和验证胃镜精灵。利用计算机产生随机数的随机方法,选取91个胃镜检查视频资料评估胃镜精灵部位识别准确率,选取45个胃镜检查视频资料及内镜医师采集的与之相匹配的胃镜图像资料,比较机械采图与人工采图胃镜检查部位覆盖个数及覆盖率。邀请2位水平相当的内镜医师,收集使用胃镜精灵的医师1使用胃镜精灵前后分别完成的45个胃镜检查图像资料,收集不使用胃镜精灵的医师2同期分别完成的20、22个胃镜检查图像资料,比较两者胃镜检查部位覆盖率。结果胃镜精灵的部位识别总体准确率为85.125%(1 156/1 358)。使用胃镜精灵的医师1使用胃镜精灵前后胃镜检查部位覆盖率分别为(76.790±8.848)%和(87.325±7.065)%,未使用胃镜精灵的医师2相应时间段内胃镜检查部位覆盖率分别为(75.926±11.565)%和(75.253±14.662)%。使用胃镜精灵前,医师1和医师2水平相当(t=0.324,P=0.747);使用胃镜精灵后,医师1胃镜检查部位覆盖率高于使用前(t=6.222,P=0.001),亦高于同期医师2(t′=3.588,P=0.002)。机器采图的胃镜检查部位覆盖个数为(20.956±3.406)个,部位覆盖率为(77.613±12.613)%,人工采图分别为(15.467±2.296)个、(57.284±8.503)%。机器采图部位覆盖个数(t=11.523,P=0.000)与覆盖率(t=11,523,P=0.000)均高于人工采图。结论胃镜精灵可提高胃镜检查部位覆盖率及覆盖个数,改善传统胃镜检查时检查部位覆盖不全及采图不全面的情况。

Objective To analyze the blind area monitoring and independent image acquisition function of gastroscopic elves (a real-time gastroscopic monitoring system) in gastroscopy. Methods A total of 38 522 gastroscopic images from the database of Digestive Endoscopy Center of Renmin Hospital of Wuhan University were collected to train and validate the gastroscopic elves.Using computer to generate random numbers, 91 gastroscopic videos were selected to assess the position recognition accuracy of the gastroscopic elves, and 45 gastroscopic videos and matching gastroscopic images collected by endoscopists were selected to compare the coverage number and rate of gastroscopy sites between gastroscopic elves and endoscopists image acquisition. Two endoscopists entered the study to perform gastroscopies with or without gastroscopic elves. Forty-five gastroscopies respectively performed by the endoscopist A before and after usage of gastroscopic elves were collected, and 42 gastroscopies divided into 20 and 22 performed by the endoscopist B without use of gastroscopic elves in the same period were also collected. The coverage rate of gastroscopy sites was compared between the two endoscopists. Results The total position recognition accuracy of gastroscopic elves was 85.125%(1 156/1 358). The coverage rate of gastroscopic sites for the endoscopist A was (76.790±8.848)% and (87.325±7.065)%, respectively, before and after using gastroscopic elves, and the coverage rate in the same period for the endoscopist B was (75.926 ±11.565)% and (75.253 ±14.662)%, respectively. The coverage rate before using gastroscopic elves had no statistical difference between the two endoscopists (t=0.324, P=0.747). The coverage rate for the endoscopist A after using gastroscopic elves was higher than that before using gastroscopic elves (t=6.222, P=0.001), and that of the endoscopist B in the same period (t′=3.588, P=0.002). The coverage number and rate of gastroscopy sites for gastroscopic elves and endoscopists image acquisition were 20.956 ±3.406 and (77.613±12.613)%, and 15.467±2.296 and (57.284±8.503)%, respectively, with statistical differences (t=11.523, P=0.000;t=11.523, P=0.000). Conclusion Gastroscopic elves can improve the coverage number and rate of gastroscopy sites, and is worthy of promotion in clinics.