Semantic Segmentation Of Esophageal Precancerous Lesions Image Based On Deep Learning

Esophageal cancer is one of the seven major cancers in the world.Endoscopy is an effective means to diagnose precancerous lesions of the esophagus.However,due to the limited experience of medical equipment and clinicians,precancerous lesions of the esophagus are usually mild and difficult to be detected.Therefore,it is particularly important to study a technical means to assist doctors in diagnosing precancerous lesions of the esophagus.In this paper,we build a multi-mode esophageal precancerous lesion data set,design different algorithm models based on the deep learning convolutional neural network and Transformer mechanism to conduct image semantic segmentation research,build a simple artificial intelligence-assisted early esophageal cancer diagnosis system,and carry out the following research to solve the problem that the edge of the lesion is difficult to detect and the small target lesion is difficult to locate:(1)A semantic segmentation algorithm for esophageal precancerous lesions based on context feature perception and dual frequency up sampling combined with an attention mechanism is proposed.First,a hybrid domain attention mechanism is designed,which can effectively enhance the feature expression of the lesion edge and enhance the ability to extract detailed features of the algorithm model;Secondly,a context feature awareness module is designed based on the middle transition layer fusion feature and void convolution,which enables the network to collect detailed features at different scales,and at the same time extract semantic information in the large sensing field;Finally,a dual frequency up sampling module is proposed to replace the single upsampling mode,which effectively reduces the sawtooth effect caused by interpolation and the checkerboard effect caused by deconvolution.The Dice similarity coefficient,average cross merge ratio,sensitivity,and specificity indexes on the selfbuilt dataset are 90.29%,83.54%,90.70%,and 92.95% respectively,which are superior over semantic segmentation algorithms such as Trans U-Net and PSPNet.(2)To solve the problem that the details and semantic information in the existing esophageal semantic segmentation algorithm cannot be obtained at the same time,a Transformer-based semantic segmentation network for esophageal precancerous lesions is proposed,which consists of an encoder,a decoder,and an enhancement feature module.The multi-head selfattention method based on divided windows can effectively process esophageal precancerous lesions images of different scales.The block mixing layer and block expansion layer can efficiently complete the resolution transformation of esophageal precancerous lesions,and solve the problem that two-dimensional feature vector expression is limited by online sampling and offline sampling;The proposed enhanced feature module can enhance the expression of detailed features such as bottom lesion edges or small target lesions,and enhance the ability of the model to obtain semantic information of the overall overview of high-level lesions.The accuracy,Dice similarity coefficient,average intersection ratio,sensitivity,and specificity indices on the built multimodal esophageal precancerous lesions dataset are 97.69%,90.98%,83.21%,89.25%,and 91.29%,respectively.The Dice similarity coefficient is superior to the third chapter algorithm.(3)Design and implementation of artificial intelligence-aided diagnosis system for early esophageal cancer.Based on the above algorithm model and the practical application of endoscopy,the auxiliary diagnosis software is compiled.Assisted doctors can find the lesions in endoscopy in time,and accurately locate the lesion area and type,which can improve the efficiency of screening and diagnosis to a certain extent.