Research Of Helicobacter Pylori Endoscopic Image Classification Based On Lightweight CNN And Transformer Hybrid Model

Helicobacter Pylori(HP)that is one of the important factors leading to gastric diseases is closely related to gastric diseases and gastric cancer.The detection and identification of HP infection may be prone to traumatic injury.Tt also to be interfered by external factors.In recent years,deep learning has been widely used in the field of medical image diagnosis and recognition.However,when the hardware and computing power are limited,The task processing effect of network model tends to decline.In order to solve the above problems,this paper builds lightweight convolutional neural networks and parallel efficient hybrid models of Transformer and CNN based on HP infection recognition tasks.The main work of the paper includes the following three aspects:(1)A gastric endoscopic image dataset of Helicobacter pylori classification was constructed.A lightweight multi-scale Hp classification basic network MoB＿MNext is proposed.The endoscope image data set was collected,and data preprocessing was carried out to construct data set.Through the experiments of several classical network models,it is found that the performance of multi-scale module and lightweight module Sand Glass is superior.A lightweight multi-scale model MoB＿MNext is constructed by integrating multi-scale module and lightweight module Sand Glass.The number of parameters was 4.4M,the accuracy was 91.63%,the sensitivity was87.84%,the specificity was 94.77%,and the AUC was 0.9619.(2)The FD＿F＿ECA＿MoB＿MNext model is constructed by adding F＿ECA attention and adopting fast downsampling strategy based on MoB＿MNext network.Based on the MoB＿MNext network,a F＿ECA attention mechanism combining multifrequency component pooling and cross-channel one-dimensional convolution is proposed,and the FD＿F＿ECA＿MoB＿MNext model is constructed by using fast downsampling strategy to reduce the network size.The number of parameters was2.4M,the accuracy was 92.26%,the sensitivity was 89.23%,the specificity was94.37%,and the AUC was 0.9636.(3)A Hp classification model FECA＿MoB＿seq＿Former is proposed based on multi-branch lightweight CNN module and Transformer.Take advantage of CNN’s advantages in local information processing and Transformer’s advantages in global interaction through two-way bridge parallel connection.CNN subblocks adopt multibranch lightweight F＿ECA modules,and Transformer subblocks add a sequence pooling strategy at the end to alleviate the problem of insufficient data sets in Transformer structure.The number of parameters was 3.83 M,the accuracy was92.74%,the sensitivity was 89.35%,the specificity was 95.09%,and the AUC was0.9688.The proposed network can classify Hp infection in endoscopic images with limited computational resources,and provides a new method for the intelligent diagnosis of HP in endoscopic images.