Research On Polyp Segmentation Technology Based On Multi-scale Feature Fusion And Reverse Attention Mechanism

In recent decades,the incidence and mortality of colorectal cancer have been high,and it has become one of the leading causes of death in humans.As we all know,colorectal cancer is a common malignant tumor,which is basically caused by the deterioration of intestinal polyps.Therefore,early screening and removal of colorectal polyps is extremely important for patients,and has great research value in clinical medicine.Currently,colorectal polyp screening is mainly performed by imaging the colon with colonoscopy,but this method may result in missed detection.Therefore,how to use computer-aided diagnosis technology(CAD)to help professional doctors accurately and efficiently detect and segment polyps has become a research topic of great interest to researchers.In recent years,medical image segmentation technology is an important research direction in the field of computer vision.Scholars have studied many methods of medical image segmentation,including traditional methods and methods based on deep learning.The traditional method requires manual acquisition of certain features,which is difficult and the accuracy of segmentation needs to be further improved.However,deep learn-based methods can overcome the problems of traditional methods to a certain extent,but they still have problems such as poor handling of target region and boundary relations,large number of model parameters and high computational complexity.This paper proposes a polyp segmentation network(MAR-UNet)based on attention Ushaped network structure and multi-scale feature fusion.By adding attention gate at the end of jump link in u-shaped network structure,semantic ambiguity caused by jump link is eliminated and salient feature representation is enhanced.The receptive field module is used to fuse the context information of different scales to enhance the network depth representation.Then,the inverse attention module was used to mine the target area and boundary information,and the result of polyp segmentation was obtained.Finally,the refined residual module was used to further refine the segmentation results of polyps,and finally,a more comprehensive and refined segmentation result of intestinal polyps was obtained.On the basis of MAR-UNet,this paper studies how to further reduce the number of model parameters and computational complexity while maintaining the model accuracy,and then proposes a polyp segmentation network(MobileRaNet)based on the lightweight model and reverse attention mechanism.By using coordinated attention module(CA)to improve MobileNetV3,a light network(CaNet)with fewer model parameters and low computational complexity is obtained,which is used as the backbone network.Then a parallel axial receptive field module is proposed on the basis of the receptive field module to carry out additional global refinement of the high-level features of different scales and connect the output.Finally,the inverse attention module was used to establish the relationship between the target area and the boundary cue of the polyp.Finally,efficient and excellent performance of intestinal polyp segmentation results were obtained.By using five challenging data sets,including CVC-ColonDB,CVC-300 and Kvasir,the accuracy,FLOPs and parameter number of MeanDice,MeanIoU and MAE were compared with those of PraNet,SFA and other five typical models.Experimental results show that the performance of MAR-UNet and MobileRaNet proposed in this paper has been improved to varying degrees in the five data sets.In particular,MobileRaNet’s MeanDice and MeanIOU indexes in Kvasir data set have reached 91.2% and 85.6%,respectively.Compared with PraNet,the number of FLOPs and parameters decreased by 83.3% and76.7%,respectively.