Recognition Of Pathological Sections Of Gastric Cancer And Segmentation Of Cancerous Regions Based On Resnet

Cancer is a difficult disease that humans can hardly overcome.Every day,cancer patients listen to doctors' judgments about fate.In Europe and the United States,the number of people who die from cancer has decreased.However,in other regions such as Asia,the proportion of patients with gastric cancer and liver cancer is increasing.The incidence and mortality of gastric cancer ranks second,second only to lung cancer.Image recognition and segmentation of cancerous pathological sections of gastric cancer is an important means of assisted diagnosis of gastric cancer.Existing Resnetbased recognition models are based on natural image data sets and can be trained using a large number of natural image samples.However,for the recognition of gastric cancer pathological slice images,the basic Resnet model construction does not take into account the way doctors diagnose,and the medical data set is unlikely to reach the same scale as the Imagenet data set.Therefore,based on the different magnifications of gastric cancer pathology and the doctor's diagnosis habits,this paper proposes a Resnetbased image recognition model for gastric cancer pathology based on the small size of the gastric cancer slice image dataset.The supervised learning segmentation model based on medical images,such as U-Net,has a good performance in segmentation of medical images.However,for the gastric cancer pathological slice image dataset used in some of the cancerous regions used in this paper,because there are unlabeled cancerous regions,and the unlabeled region has low signal-to-noise ratio,the partially labeled gastric cancer pathological slice image data set is used to directly train U-Net model,model performance is poor.Existing semi-supervised segmentation models based on natural images are focused on training complex auxiliary branches.In this paper,three kinds of segmentation networks with Resnet as encoder are designed,and a semi-supervised cancerous region segmentation algorithm based on self-training is proposed.It is not necessary to add complex auxiliary branches.In the image recognition task of gastric cancer pathology,this paper proposes a recognition network structure based on Resnet.Based on the pathological image imaging method and the doctor's reading method,this paper designs a multi-scale input Resnet.The data set of gastric cancer pathological slice images used in this paper is small,which is in line with the difficulty of labeling medical data sets in the real world,resulting in smaller scale.Therefore,after the global pooling layer in Resnet,the dropout structure is introduced,and the L2 regular term is introduced in the loss calculation to suppress the over-fitting caused by the training of small-scale data sets.In the segmentation task of cancerous cancerous region,this paper proposes a semi-supervised cancerous region segmentation model based on Resnet.In this paper,Resnet is used as the encoder,and three segmentation models are constructed using UNet decoder,FPN and DFP module proposed in this paper.In the experiment,the label correction algorithm based on color deconvolution algorithm and traditional image processing segmentation algorithm was proposed for professional doctors to label the problem that the cancerous area and the background contact edge were rough.After experiment,the algorithm was modified and can be used in the experiment.Quickly segment the nuclear region of the pathological image.Aiming at the dataset of some labeled cancerous regions,this paper designs a semi-supervised cancerous region segmentation algorithm based on self-training algorithm to improve the performance of the segmentation model.This is the first time that the self-training algorithm has been applied to the segmentation of cancerous pathological sections of gastric cancer.