Research On Colorectal Polyp Segmentation Method Based On Convolutional Neural Network And Transformer

Colorectal polyp segmentation separates the polyp area in medical images from the background,so that doctors can use the segmentation information to formulate more targeted treatment plans and surgical plans for patients.At present,a large number of studies have combined Convolutional Neural Networks and Transformers to achieve colorectal polyp segmentation,but the segmentation results still have problems such as incomplete morphology of the target area,incoherent boundaries and boundary segmentation errors.This paper conducts in-depth research on colorectal polyp segmentation in view of the problems existing in existing methods,and proposes two new colorectal polyp segmentation networks.The main work and innovations are as follows:(1)Aiming at the insufficient interaction between different feature information and insufficient attention to the similarities and differences between regions and boundaries,this paper proposes a colorectal polyp segmentation network based on dual-branch multi-scale feature fusion.First,parallel convolutional neural networks and Transformer branches extract multi-scale local information and global context information.Second,the feature super-decoder fully fuses multi-scale local information and global context information to generate an initial segmentation map.Then,the multi-scale feature aggregation module is used to aggregate local information,and then the polarized self-attention module is used to reduce the noise information caused by aggregation and enhance the fine-grained segmentation area.Finally,the reverse attention fusion module establishes the relationship between regions and boundary cues,and refines the segmentation boundary layer by layer to obtain the final segmentation result.The performance of the proposed network is verified by five public datasets CVC-ClinicDB,Kvasir,ETIS,CVC-ColonDB,CVC-300,The experimental results show that the network in this paper can effectively solve the problems of incomplete segmentation morphology,wrong segmentation boundaries and incoherent boundaries.(2)Existing methods for colorectal polyp segmentation ignore the interaction within and between scale features,while lacking in exploring the semantic relationship between features across regions.In this paper,we propose a colorectal polyp segmentation network based on multi-scale feature boundary map inference.First,the Transformer block captures local and global information inside the multi-scale features extracted by the backbone branches of the convolutional neural network.Then,the cross-scale feature fusion module performs cross-scale interaction and cascade fusion of local and global information,generate fusion features.Finally,the low-level feature information extracted by the convolutional neural network,the fused features generated by the fusion module or intermediate prediction results are sent to the graph convolutional neural network,and the semantic relationship between the polyp area and the boundary is mined through the information propagation between the graph vertices.Compared with the current popular methods,the network in this paper has achieved better segmentation results on five data sets,among which the optimal values of Dice,Iou,and BAcc reached 94.16%,89.35% and 97.42%.