Transformer模型在胃镜视频部位实时识别的研究探索

Research and Exploration of Real-time location recognition of Transformer Model in Gastroscope Video

目的探索深度学习技术(Transformer模型)在胃镜视频中部位实时识别的性能评价。方法使用深圳市第二人民医院2021年5至7月录制的50部胃镜视频为研究资料,抽帧形成“视频图像集合”,其中40部为训练集1,10部为测试集;以含有25525张胃镜图像的“胃镜图像集合”为对照组,形成训练集2。Transformer模型基于训练集1、训练集2、“训练集1+训练集2”学习训练分别形成“基于视频智能系统”、“基于图像智能系统”、“基于视频和图像智能系统”。通过测试集比较视频智能系统、图像智能系统、视频和图像智能系统的准确度、特异度等差异。结果研究组1基于视频智能系统准确度、特异度、总体有效度分别为84.3%、78.9%、81.5%;研究组2基于视频和图像智能系统准确度、特异度、总体有效度分别为82.9%、81.5%、82.2%;对照组基于胃镜图像智能系统准确度、特异度、整体有效度分别为80.0%、76.8%、78.4%。结论结果具有统计学意义,Transformer模型视频学习效果优于胃镜图像,Transformer模型学习视频+图像效果优于单独视频学习。

Objective Performance evaluation of exploring deep learning technology(Transformer model)for real-time location recognition in gastroscope video.Methods Using 50 gastroscope videos recorded by Shenzhen second people′s Hospital from May to July in 2021 as research data,frames were drawn to form a″video image set″,of which 40 were training set 1 and 10 were test sets.The″gastroscope image set″containing 25,525 gastroscope images was used as the control group to form the training set 2.Transformer model is based on training set 1,training set 2 and training set 1+training set 2 to form″video-based intelligent system″,″image-based intelligent system″and″video-based and image-based intelligent system″respectively.Through the test set,the accuracy and specificity of video intelligent system,image intelligent system,video and image intelligent system are compared.Results The accuracy,specificity and overall validity of video-based intelligent system in experimental group 1 were 84.3%,78.9%and 81.5%,respectively,while those in experimental group 2 were 82.9%,81.5%and 82.2%,respectively.The accuracy,specificity and overall validity of the intelligent system based on gastroscope image in the control group were 80.0%,76.8%and 78.4%,respectively.Conclusion The results were statistically significant,the video learning effect of Transformer model was better than that of gastroscopy,and the effect of video+image learning of Transformer model was better than that of video learning alone.