Parallel reverse enhance attention network module based on Kolmogorov-Arnold networks for segmenting polyps showed on colonoscopy images
10.13929/j.issn.1003-3289.2025.06.026
- VernacularTitle:基于Kolmogorov-Arnold网络平行反向注意网络模型分割结肠镜图像所示息肉
- Author:
Changzheng XING
1
;
Yuheng HE
1
;
Junfeng LIANG
1
Author Information
1. 辽宁工程技术大学电信学院,辽宁葫芦岛 125105
- Publication Type:Journal Article
- Keywords:
colonic polyps;
deep learning;
image processing,computer-assisted;
colonoscopy
- From:
Chinese Journal of Medical Imaging Technology
2025;41(6):971-975
- CountryChina
- Language:Chinese
-
Abstract:
Objective To observe the value of parallel reverse enhance attention network module based on Kolmogorov-Arnold networks(KAN-PrdaModule)for segmenting polyps showed on colonoscopy images.Methods Nine hundred colonoscopy images in Kvasir-SEG dataset and 550 colonoscopy images in CVC-ClinicDB dataset were selected as training set(n=1 450),while 196 colonoscopy images in ETIS dataset,62 in CVC-ClinicDB dataset,380 in CVC-ColonDB dataset and 100 in Kvasir-SEG dataset were enrolled as test set(n=738).KAN-PrdaModule was proposed through improving U-Net,which improved detection accuracy through multi-scale feature fusion,and the value of KAN-PrdaModule for segmenting polyps showed on colonoscopy images was analyzed according to mean Dice similarity coefficient(mDSC),mean intersection over union(mIoU),weighted metric(Fωβ)and structural metric(Sα),and compared with U-Net,U-Net++,stochastic frontier analysis(SFA)and PraNet models.Results Among the above 5 models,the performance of SFA model for segmenting polyps on colonoscopy images was poor,with blurry edges of polyps and some ones were missed.U-Net and U-Net++models had decent performance,which could roughly identify polyps.PraNet model performed well,and the segmented edges of polyps were clear.KAN-PrdaModule had the best performance,showed high similarity to the true value images,with the best overall mDSC,mIoU,Fωβand Sα.Conclusion KAN-PrdaModule could effectively segment polyps showed on colonoscopy images,with segmenting effect better than U-Net,U-Net++,SFA and PraNet models.
