计算机技术与发展2025，Vol.35Issue(11)：38-45,8.DOI:10.20165/j.cnki.ISSN1673-629X.2025.0167
结合U-Net与KAN的医学图像分割研究
Research on Medical Image Segmentation Combining U-Net and KAN
摘要
Abstract
Due to its simple and efficient network structure,U-Net has become the benchmark model in the current field of medical image segmentation and has achieved good results in various image segmentation tasks.However,the traditional U-Net still has certain limitations in detail feature extraction,cross-scale information fusion and complex structure recognition,and it is difficult to fully adapt to the challenges such as deformation,low contrast and diverse targets existing in medical images.To further improve the segmentation per-formance,we propose an improved model U-KPD(U-Net with Kolmogorov-Arnold Network and ParNet-Deformable Module).This model introduces the Kolmogorov-Arnold Network(KAN)based on U-Net to enhance the network's ability to express the local and global features of the image.Meanwhile,the ParNet-Deformable module(PD)is combined to enhance the model's adaptive modeling of key areas and the ability to capture deformed structures,thereby improving the accuracy and robustness of segmentation.Through thorough experimental verification on two representative datasets,CVC-ClinicDB and BUSI,the results show that U-KPD outperforms the traditional U-Net and other mainstream improved models in terms of multiple evaluation indicators such as IoU,Dice coefficient,and HD95.Especially in terms of the recognition accuracy of complex structures and deformed targets,it performs even more excellently and has good universality and application prospects.关键词
医学图像分割/U-Net/科尔莫哥罗夫-阿诺德网络/特征增强/适应形变特征Key words
medical image segmentation/U-Net/Kolmogorov-Arnold Network(KAN)/feature enhancement/adaptive deformable fea-tures分类
信息技术与安全科学引用本文复制引用
马文旭,李东,刘晓静..结合U-Net与KAN的医学图像分割研究[J].计算机技术与发展,2025,35(11):38-45,8.基金项目
国家自然科学基金(62366043) （62366043）
青海省科技厅2023年十大国家级创新平台项目子课题(ZYYSDPT-2023-02) （ZYYSDPT-2023-02）
