Research On Lesion Recognition Of ME-NBI Gastric Image Based On Deep Learning

The occurrence of gastric cancer goes through the process of inflammation,intestinal metaplasia and low-grade neoplasia,but it is not easy to be found by patients in the early stage.Magnifying endoscopy combined with narrow-band imaging technology(ME-NBI)enables endoscopists to more clearly judge the gastric lesions of patients through the generation mode of gland tubes on the mucosal surface.However,the results of manual endoscopic image diagnosis highly depend on the professional level of doctors.In recent years,the rapid development of deep learning technology has made it possible for computer-aided diagnosis of gastric cancer.Under this background,a computer-aided diagnosis system for early cancer and precancerous lesions is designed in this paper.Firstly,a new ATP-UNet image segmentation model is proposed for the abundant glandular texture and color information in gastric endoscopic images.The model selects VGG16 as the backbone feature extraction network,and combines the output of each layer of the coding layer with the pyramid network model to realize the fusion between channels.At the same time,the attention module is introduced to make the model have stronger ability to extract the detailed features of the region of interest while suppressing the background information.Secondly,according to the characteristics of small medical image data set,the natural image data set is used to transfer the model,and the rationality of the improved model design is compared from three aspects: 1)The performance of six different loss functions is compared,and dice + CE is selected as the combined loss function of this method.2)The UNet,P-UNet,TP-UNet and the final ATP-UNet model in the improvement process are trained respectively,which shows the improvement of the performance of ATP-UNet model.3)The training results of UNet,segnet and pspnet models are compared,which shows that the improved model proposed in this paper has good performance in gastric cancer image segmentation.The improved model is used to segment ME-NBI endoscopy.The reconciliation parameters of precision and recall are96%,and the average intersection and merging ratio can reach 0.9,which shows that the method proposed in this paper has good recognition ability of early cancer and gastric lesions.Finally,combined with the early cancer and gastric lesion images segmented by ME-NBI,the lesion rate of endoscopic image is calculated.The GUI interface of auxiliary diagnosis system is developed by using Labview,which has good man-machine interaction.Endoscopy experts said that the system is very helpful to realize the computer-aided diagnosis of clinical gastric cancer.