摘要
目的建立基于深度学习的人工智能模型,通过对胃良、恶性溃疡及正常黏膜病例图片的学习,使其对良、恶性溃疡病例具有较高的识别能力,从而在一定程度上提高内镜医师的诊断水平,为胃癌的早期诊断提供极大的助力。方法按照纳入和排除标准,共纳入病例3238例,其中恶性胃溃疡、良性胃溃疡及正常胃的病例数分别为:747例、761例及1730例。采用随机原则将原始病例按照大致8∶1∶1的比例分为训练集(正常病例1380例,良性胃溃疡病例614例,恶性胃溃疡病例596例)、验证集(正常病例179例,良性胃溃疡病例75例,恶性胃溃疡病例70例)和测试集(正常病例171例,良性胃溃疡病例72例,恶性胃溃疡病例81例)。建立基于多图片输入的efficientNet-b4深度学习模型对训练集及验证集中的病例进行训练,然后与2名经验丰富的内镜医师一起对测试集的324例病例进行识别,得出相关统计学指标,从而比较人工智能与内镜医师的对胃正常黏膜及良、恶性胃溃疡病例识别的诊断水平。结果经验丰富内镜医师1、经验丰富内镜医师2及人工智能对病例识别的整体准确率分别为:92.09%、91.36%及95.06%,人工智能对病例识别的整体准确率要明显优于经验丰富的内镜医师。对于正常病例的识别,人工智能的灵敏度为98.25%,略低于两位内镜医师的99.42%和100%,但其阳性预测值为100%,优于其中一位医师的97.14%,与另一位医师的100%结果一样;对于良性胃溃疡病例的识别,人工智能的灵敏度和阳性预测值分别为:91.67%和86.84%,均优于两位经验丰富的内镜医师;对恶性胃溃疡病例的识别,人工智能的灵敏度和阳性预测值分别为:91.36%和92.50%,均优于两名经验丰富的内镜医师;人工智能及两名内镜医师对于良、恶性胃溃疡病例识别的灵敏度及阳性预测值都要明显低于其对正常病例的识别。结论通过深度学习的人工智能模型对胃�
Objective The artificial intelligence model based on deep learning is established to have high recognition ability of benign and malignant ulcer cases through learning pictures of benign and malignant ulcer cases and normal mucosa,so as to improve the diagnosis level of endoscopists to a certain extent and provide great help for the early diagnosis of gastric cancer.Methods According to inclusion and exclusion criteria,a total of 3,238 cases were included,including 747 cases of malignant gastric ulcer,761 cases of benign gastric ulcer and 1,730 cases of normal stomach,respectively.The original cases were randomly divided into training set(1,380 normal cases,614 benign gastric ulcer cases,and 596 malignant gastric ulcer cases),validation set(179 normal cases,75 benign gastric ulcer cases,and 70 malignant gastric ulcer cases)and test set(171 normal cases,72 cases were benign and 81 cases were malignant gastric ulcer).Efficientnet-b4 Efficientnet-B4 deep learning model was established with multi-image input to train the cases in the training set and validation set,and then identified the 324 cases in the test set with two experienced endoscopists to obtain relevant statistical indicators.In order to compare the diagnostic level of artificial intelligence and endoscopy in the recognition of normal gastric mucosa and benign and malignant gastric ulcer cases.Results The overall accuracy of case identification by experienced endoscopists 1,experienced endoscopists 2 and artificial intelligence was 92.09%,91.36%and 95.06%,respectively.The overall accuracy of case identification by artificial intelligence was significantly better than that by experienced endoscopists.The sensitivity of artificial intelligence was 98.25%for the recognition of normal cases,slightly lower than 99.42%and 100%of the two endoscopists,but its positive predictive value was 100%,which was better than 97.14%of one physician and the same as 100%of the other physician.For the identification of benign gastric ulcer cases,the sensitivity and positive p
作者
赖春晓
张希钢
白杨
李峰
戴捷
何顺辉
江海洋
LAI Chun-xiao;ZHANG Xi-gang;BAI Yang;LI Feng;DAI Jie;HE Shun-hui;JIANG Hai-yang(Gastrointestinal Cancer Center,Baiyun Branch,Nanfang Hospital,Southern Medical University,Guangzhou 510500;Department of Gastroenterology,Shenzhen Second People′s Hospital,Shenzhen,Guangdong,518000;Department of Gastroenterology,Shenzhen Hospital,Beijing University of Chinese Medicine,Shenzhen,Guangdong 518100;Zidong Information Technology(Suzhou)Co.,LTD.,215123;Department of Gastroenterology,Shunde Hospital,Southern Medical University,Shunde 528000;Department of Gastroenterology,Shayang Hospital of Traditional Chinese Medicine,Shayang 448200)
出处
《现代消化及介入诊疗》
2022年第1期31-35,共5页
Modern Interventional Diagnosis and Treatment in Gastroenterology
基金
广东省基础与应用基础研究基金项目
广东省自然科学基金面上项目(2019A1515012115)。
关键词
人工智能
深度学习
良性胃溃疡
恶性胃溃疡
诊断
Artificial intelligence
Deep learning
Benign gastric ulcer
Malignant gastric ulcer
Diagnosis
