Research On Colon Polyp Image Segmentation Algorithm Based On Improved HarDNet-MSEG

Colorectal cancer is caused by colorectal adenomatous polyps,and the development of colorectal cancer is closely related to the presence of colorectal polyps,which can be effectively detected by clinical colonoscopy screening.Due to the diverse polyp morphology and the severity of lesions,the inspection process may result in manual misses and misdetections.By using the current segmentation of colon polyp images in deep learning,the segmentation speed and accuracy are greatly improved.However,in view of the similar color of colon polyps and its surrounding tissues and the diversity of size,shape and texture of colon polyps,this paper designs a Transformer encoder and a multi-scale feature fusion module to improve the segmentation algorithm of colon polyp images.The main research works of this paper are as follows.（1）This paper firstly introduces the research background and significance of colon polyp image segmentation.we have reviewed the domestic and international research on colon polyp image segmentation algorithms and analyzed the current research status of colon polyp image segmentation algorithms.Based on these literature and analysis,we have implemented UNet++,Res UNet++,Pra Net and HarDNet-MSEG colon polyp image segmentation algorithms.Comparative experiments of the above four algorithms were conducted using publicly available colon polyp datasets.According to the segmentation results of polyp images and medical image segmentation metrics,the advantages and disadvantages of these algorithms and the difficulties of colon polyp image segmentation are analyzed.（2）Through experimental comparison and analysis,the HarDNet-MSEG image segmentation algorithm of colon polyps has the highest segmentation accuracy among the four algorithms.However,HarDNet-MSEG cannot effectively extract the global semantic information of polyp images due to the difficulty of segmenting colon polyps with similar color to their surrounding tissues and the diversity of size,shape and texture of colon polyps.Therefore we replace the RFB module in the HarDNet-MSEG model with the Transformer encoder,which reduces the number of parameters of the model,speeds up the training and inference of the model,and improves the ability of the model to extract the global semantic information of the polyp image.At the same time,a multi-scale feature fusion module is designed,which can fuse the superficial and low-level position information from the polyp image.The multi-scale feature fusion module is designed to fuse shallow,low-level location information and deep,high-level contextual semantic information from polyp images.In this paper,the improved model is trained and tested by designing comparative experiments.The experimental data shows that the improved model improves the m Dice scores by 1% over HarDNet-MSEG on both Kvasir-SEG and CVC-Clinic DB datasets;the m Dice scores on Colon DB,Endoscene and ETIS datasets increase by 3%,3% and 6%,respectively;the realtime segmentation frame number of the improved model reaches 95,which is an 8%performance improvement compared to HarDNet-MSEG.The segmentation accuracy is higher and faster using the improved algorithm,which has good clinical application value.