摘要
目的开发一套基于深度学习的内镜下胃黏膜多病灶辅助识别系统并评估其识别胃黄斑瘤、糜烂、息肉、黏膜下隆起4种病灶的有效性。方法收集武汉大学人民医院消化内镜中心7388张图像作为训练与验证集,另筛选出900张图像作为测试集1;收集连续患者资料,筛选出1240张图像作为测试集2;收集另5家医院连续患者资料,筛选出6536张图像作为测试集3。分别在测试集1~3中评估该系统识别病灶效能和稳定性。结果系统在测试集1中诊断胃黄斑瘤、糜烂、息肉、黏膜下隆起的灵敏度分别为81.33%、88.00%、84.00%、88.67%,特异度分别为99.33%、93.11%、95.11%、93.11%,准确度分别为94.83%、91.83%、92.33%、92.00%;系统在测试集2、测试集3中诊断胃内4类病灶的综合准确度分别为82.39%、87.92%。结论人工智能系统对胃黄斑瘤、糜烂、息肉和黏膜下隆起具有良好的辅助诊断能力,可辅助内镜医师提高胃镜检查的质量。
Objective To develop a deep-learning-based system aimed at identifying multiple types of gastric benign lesions under endoscopy,and to evaluate its diagnostic performance.Methods 7388 images were collected for training and validation,900 images selected as dataset 1,data from consecutive patients were collected,from which 1240 images were selected for dataset 2 from Renmin Hospital of Wuhan University.Data from consecutive patients at five other hospitals were also collected,from which 6536 images were selected for dataset 3.The diagnostic capabilities of the deep learning model were evaluated in dataset 1,2 and 3,respectively.Results The sensitivity for detecting gastric xanthelasma,erosion,polyps,and submucosal lesions in the dataset 1 was 81.33%,88.00%,84.00%,and 88.67%,respectively,with the specificity being99.33%,93.11%,95.11%,and 93.11%,and the accuracy being 94.83%,91.83%,92.33%,and 92.00%,respectively.The comprehensive accuracy in dataset 2 and dataset 3 for identifying four types of lesions in the stomach was 82.39%and 87.92%,respectively.Conclusion The deep-learning-based system has a good detective and diagnostic ability,of the potential for assisting endoscopists in accurately identifying focal lesions such as xanthelasma,erosion,polyps,and submucosal lesions in the stomach.
作者
杜泓柳
董泽华
吴练练
张军
张丽辉
李佳
于红刚
Du Hong-liu;Dong Ze-hua;Wu Lian-lian;Zhang Jun;Zhang Li-hui;Li Jia;Yu Hong-gang(Department of Gastroenterology,Renmin Hospital of Wuhan University,Wuhan 430060,China)
出处
《兰州大学学报(医学版)》
2022年第1期5-8,共4页
Journal of Lanzhou University(Medical Sciences)
基金
湖北省重大科技创新项目(2018-916-000-008)。
关键词
深度学习
内镜
多类别分类
deep learning
endoscopy
multicategory classification