Research On Segmentation Algorithm Of Pathological Image Of Gastric Cancer Based On TransUNet

Gastric cancer is one of the most common malignant tumors and has become a public health problem worldwide.As the "gold standard" of cancer diagnosis and treatment,pathological examination plays an irreplaceable and critical role in the treatment of patients.However,with the increase in the number of new gastric cancer patients and the rise in health awareness,the surge in the demand for pathology diagnosis has placed a heavy burden on pathologists,and the shortage of pathologists has made it difficult for patients to receive timely diagnosis and treatment.There have been many studies on semantic segmentation algorithms for pathology images in recent years,but these methods are difficult to achieve accurate and robust segmentation for pathology images of gastric cancer.To address the difficulties in the segmentation of gastric cancer pathology images,this paper proposes an improved Trans UNet-based segmentation algorithm for gastric cancer pathology images to solve the above problems.In this paper,we analyze the characteristics of gastric cancer pathology images,introduce the Trans UNet semantic segmentation model,take advantage of the Transformer to establish global dependencies to encode the features,and improve the network structure by introducing a multi-scale fusion supervised structure in the decoding part,and use the feature maps of different scales output from each decoder in the upsampling stage to supervise the training of the network,and introduce The network structure is improved by introducing a multiscale fusion supervision structure in the decoding part,and using the feature maps of different scales output from each decoder in the upsampling stage to supervise the training of the network,so as to make the whole network approximate a multi-network parallel structure with fewer parameters,so that the network can fully utilize the feature maps of each scale.The experimental results show that the improved network model has better performance,and the average Dice similarity coefficient on the dataset reaches 93.15%.The staggered overlap prediction mechanism is proposed to address the problems of missing features and stitching gaps caused by block slicing of pathological images.The staggered tessellation step is introduced in the block tessellation process to expand the "observation field" of the model,and the stitching gaps are eliminated by overlapping stitching.To address the problem of model generalization caused by large differences in staining effects of HE pathology images,a generalization preprocessing mechanism based on HE staining normalization is proposed.The mechanism obtains the HE staining components of the model training data through singular value decomposition,and uses the staining components to normalize the staining of the images to be segmented so that their color distribution approximates the image data familiar to the model,which indirectly improves the generalization ability of the model and reduces the data dependence of the pathology image segmentation task.The experimental results on the test set show that this mechanism can effectively deal with the generalization problem brought about by the coloring differences,which is of high research significance for the medical pathology image segmentation task.