基于深度学习二次预训练建立Barrett食管内镜图片分类模型的研究

Study on the establishment of Barrett’s esophagus endoscopy image classification model based on secondary pre-trained deep learning

目的:利用深度卷积神经网络二次迁移学习方法构建Barrett食管内镜图片分类模型,并比较经单次与二次预训练的模型效能。方法:选取在ImageNet数据集上进行首次预训练的4个深度卷积神经网络(ResNet、NASNetL、Xception及EfficientNet),经HyperKvasir食管炎数据集二次预训练,而后进行目标训练,得到分类模型,评价其分类能力,并对模型的推理能力进行可视化呈现。结果:在验证集中,经二次预训练所建立的模型准确性均高于单次模型。除Xception模型外,二次预训练模型的精确度和召回率均优于单次模型,其中,EfficientNet模型表现最优。结论:基于深度卷积神经网络二次预训练建立的Barrett食管计算机视觉模型具备良好的内镜图片分类能力;本研究可为临床小数据集进行深度神经网络迁移学习提供思路。

Objective To construct a classification model of Barrett’s esophagus endoscopy images by using the secondary transfer learning method of deep convolution neural network,and to compare the model performance after single and secondary pre training.Methods Four deep convolutional neural networks(ResNet,NASNetL,Xception and Efficient Net)that were pre-trained for the first time on the ImageNet dataset were selected,and then pre-trained for the second time on the HyperKvasir esophagitis dataset,and then target training was carried out to obtain the classification model,evaluate its classification ability,and visualized the reasoning ability of the model.Results In the validation set,the accuracy of the models established by the second pre-training was higher than that of the single pre-training model.Except for Xception model,the accuracy and recall of the secondary pre-training model are better than that of the single model,of which the EfficientNet model performs best.Conclusion The Barrett’s esophagus computer vision model based on the second pre-training of deep convolutional neural network has a good ability to classify endoscopic images.This study may provide ideas for deep neural network transfer learning on small clinical datasets.