Research On Colon Polyp Image Segmentation Algorithm Based On Improved PraNet

Colorectal cancer is one of the most common malignant tumors and can develop from high-risk colon polyps.Screening and detection primarily rely on colonoscopy in clinical practice.Currently,deep learning applied to colon polyp segmentation has achieved high accuracy and can detect small targets well.However,the diversity of polyp shapes and the fuzzy boundaries between them and surrounding mucosa limit the accuracy of segmentation algorithms,resulting in poor segmentation performance in the edge area.To address the above issues,this study designs polyp image segmentation models using five classic image segmentation algorithms.They are trained and tested on publicly available polyp datasets,and model evaluation is conducted through evaluation metrics.The final experiment shows that the segmentation accuracy of Pra Net is higher.Based on the characteristics of polyp image segmentation,this study proposes an improved algorithm based on the Pra Net model:(1)To address the difficulty of the diversity of colon polyp shapes leading to insufficient segmentation accuracy,the input polyp images are first segmented by window and embedded with vector positions,followed by feature extraction using Swin-Transformer V2.After multiple experiments with various backbone networks on the training and test sets,it is finally confirmed that Swin-Transformer V2 has stronger feature extraction ability.(2)To address the difficulty of fuzzy polyp segmentation edges,the deepest three layers of features extracted by the backbone network are enhanced using CFP Blocks.In the upsampling process,the SE channel attention mechanism is used to feature-weight the feature map,and the edge features are implicitly modeled through a reverse attention module.Finally,the feature map is restored to the original image size to obtain the segmentation result.This study selected 1450 images from the publicly available polyp datasets Kvasir and CVC-Clinic DB for training,completed generalization tests on ETIS,CVC-Colon DB,and CVC-300,and verified the effectiveness of the improved module through ablation experiments.After training,the improved model achieved high accuracy,with m Dice and m Io U reaching 92.8% and 87.4%,respectively,on the Kvasir dataset,and improving by 14%and 10.5%,respectively,on the challenging ETIS dataset compared to the previous network.