Research On Colorectal Tumor Segmentation Method Based On CNN And Transformer

Automated segmentation technology based on medical image analysis has become one of the important methods for the diagnosis and treatment of colorectal tumors.It can conveniently and quickly extract tumor regions,providing important information and basis for subsequent treatment.However,due to the significant morphological differences,weak tissue boundary information,and wide distribution range of colorectal tumors,current methods still cannot meet expectations in terms of accuracy and other indicators of colorectal tumor segmentation.In addition,there are few colorectal tumor image data with high-quality annotation,which also affects the segmentation effect of the current conventional segmentation model based on supervised learning to a certain extent.Therefore,designing a fast,accurate,and small sample learning based automated segmentation algorithm for colorectal tumors has important research significance and practical clinical application value.At present,there are methods for colorectal tumor segmentation based on CNN(Convolutional Neural Networks)and Transformer.The CNN based method can enhance feature representation capabilities,but it places more emphasis on local feature extraction,which can easily lead to the loss of global information in the image.While Transformer,based on the self attention mechanism,directly carries out global relationship modeling,which can expand the Receptive field of the image,obtain more context information,and have better feature convergence ability,so it is more global.However,Transformer based methods tend to lose local features.Based on this,this project combines CNN and Transformer,comprehensively utilizing their respective advantages,and mainly completes the following work:(1)A new U-shaped network model based on CNN and Transformer fusion was constructed to address the issues of diverse morphology,varying scales,and weak edge information of colorectal tumors.The proposed model addresses the shortcomings of CNN’s inability to model over long distances and Transformer’s tendency to lose local features.The proposed solution involves alternating learning between U-Net and Transformer,which fully utilizes local and global information in colorectal imaging,Realize multi-scale feature extraction and precise localization of tumor edges.The experimental results show that when this segmentation model is applied to colorectal tumor image segmentation,the average Dice coefficient index is2.67 higher than the current best Trans Unet network,and the average Hausdorff distance index is 13.36 mm higher.(2)Aiming at the problem of less high-quality annotation data at present,a semi supervised image segmentation method based on the above network architecture is designed.Through the cross teaching strategy between CNN and Transformer,cross modal collaborative learning is carried out to improve the performance and robustness of the model.The experimental results show that the network model is superior to other semi supervised models in terms of the average Dice coefficient and the average Hausdorff distance on the colorectal tumor data set.Therefore,this method can achieve more accurate and stable segmentation results on a small number of effective datasets.In summary,the U-shaped network model based on CNN and Transformer fusion constructed in this project and the semi supervised image segmentation method based on this network architecture can effectively improve the automated segmentation performance of colorectal tumor regions.