摘要
目的构建一个基于深度学习的人工智能辅助诊断模型,用于实时动态识别胃镜下的胃部病变及位置,并评估其对胃部病变检出和位置识别的能力。方法回顾性分析我院104例患者的胃镜检查视频,对视频帧进行人工标注,将已标注的病变类别图片帧按8:2的比例划分成训练集与验证集,将已标注的位置类别图片帧根据患者来源按8:2的比例划分成训练集与验证集,分别用于模型的训练及验证。病变识别部分的训练采用YoloV4模型,位置识别部分的训练采用ResNet152模型。评估辅助诊断模型对病变识别的准确率、敏感度、特异度、阳性预测值、阴性预测值,及对位置识别的准确率。结果共标注图片68351帧,其中训练集图片54872帧,包括病变类别41692帧,位置类别13180帧;验证集图片13479帧,包括病变类别10422帧,位置类别3057帧。在验证集中,病变识别模型的总体准确率为98.8%,敏感度为96.6%,特异度为99.3%,阳性预测值为96.3%,阴性预测值为99.3%;位置识别模型top5总体准确率为87.1%。结论基于深度学习的人工智能辅助诊断模型用于实时动态识别胃镜下胃部病变及位置有较好的病变检出能力和位置识别能力,临床应用前景巨大。
Objective To construct an artificial intelligence assisted diagnosis model based on deep learning for dynamically recognizing gastric lesions and their locations under gastroscopy in real time,and to evaluate its ability to detect and recognize gastric lesions and their locations.Methods The gastroscopy videos of 104 patients in our hospital was retrospectively analyzed,and the video frames were manually annotated.The annotated picture frames of lesion category were divided into the training set and the validation set according to the ratio of 8∶2,and the annotated picture frames of location category were divided into the training set and the validation set according to the patient sources at the ratio of 8∶2.These sets were utilized for training and validating the respective models.YoloV4 model was used for the training of lesion recognition,and ResNet152 model was used for the training of location recognition.The accuracy,sensitivity,specificity,positive predictive value,negative predictive value and location recognition accuracy of the auxiliary diagnostic model were evaluated.Results A total of 68351 image frames were annotated,with 54872 frames used as the training set,including 41692 frames for lesion categories and 13180 frames for location categories.The validation set consisted of 13479 frames,comprising 10422 frames for lesion categories and 3057 frames for location categories.The lesion recognition model achieved an overall accuracy of 98.8%,with a sensitivity of 96.6%,specificity of 99.3%,positive predictive value of 96.3%,and negative predictive value of 99.3%in validation set.Meanwhile,the location recognition model demonstrated an top-5 accuracy of 87.1%.Conclusion The artificial intelligence assisted diagnosis model based on deep learning for real-time dynamic recognition of gastric lesions and their locations under gastroscopy has good ability in lesion detection and location recognition,and has great clinical application prospects.
作者
郭宪
吴应洋
江艾芮
樊超强
彭学
聂绪彪
林辉
柏健鹰
GUO Xian;WU Ying-yang;JIANG Ai-rui;FAN Chao-qiang;PENG Xue;NIE Xu-biao;LIN Hui;BAI Jian-ying(Department of Gastroenterology,the Second Affiliated Hospital of Army Medical University,Chongqing 400037,China)
出处
《局解手术学杂志》
2024年第10期849-854,共6页
Journal of Regional Anatomy and Operative Surgery
基金
重庆市技术创新与应用发展专项重点项目(cstc2019jscx-gksbX0038)。
关键词
人工智能
胃部疾病
早期胃癌
深度学习
胃镜检查
artificial intelligence
gastric diseases
early gastric cancer
deep learning
gastroscopy examination
