Research On Colon Polyp Segmentation Method Based On Deep Learning

Colorectal cancer caused by colon polyps is a serious disease with a high morbidity and mortality rate.It poses a great threat and harm to people’s health.There is a close relationship between colon polyps and colorectal cancer.How to detect colon polyps quickly,accurately and early becomes an important issue.Traditional segmentation of colon polyps is time consuming,does not have sufficient segmentation accuracy and cannot fuse multiple features based on a single feature.Deep learningbased segmentation methods overcome some of the problems of traditional segmentation methods,but suffer from poor handling of boundary relationships and difficulty in capturing multi-scale features.In conclusion,the difficulties in segmenting colon polyps are:(1)The various shapes and sizes of polyps lead to difficulties in feature extraction.(2)The low contrast between polyps and the tissue surrounding them and the blurred boundaries.To address the problem that the various shapes and sizes of polyps make feature extraction difficult,this paper proposes an APra Net network based on multiscale feature fusion to improve the Pra Net network.The network is embedded with a multilevel multi-scale feature fusion module(ASPP).Feature maps of different sizes in the network are reduced to a uniform size connection and fed together into the ASPP module for multi-scale feature extraction.The ASPP module enhanced the ability to extract multi-scale feature information from colon polyps and improved the accuracy of the model in segmenting colon polyps,as indicated by ablation and comparison experiments.To address the problems of low contrast and blurred boundaries between polyps and their surrounding tissues,this paper further improves on APra Net by proposing a colonic polyp segmentation network APrca Net with a channel space hybrid attention module(CBAM)embedded in the reverse attention module(RA).In the RA module,when unwanted background noise is introduced,the performance of the module suffers,the boundary information is then diminished.The RCA module formed by embedding CBAM in the RA module will attenuate the noise and enhance the boundary information,achieving the goal of differentiating polyps from their surrounding tissue in low contrast situations.To verify the validity of the work,this paper uses five data sets and evaluation metrics mean Dice and mean IOU to test the model.In terms of experiments,this paper uses ablation experiments to verify the effectiveness of the individual module improvements and comparison experiments with other methods to verify the effectiveness of the overall network.The experimental results show that APrca Net has some improvement in the evaluation metrics compared to other methods.