摘要
目的:探讨并解决传统U-Net算法在肝脏肿瘤分割中,肝脏及肿瘤上下文信息缺乏、肿瘤形态差异性大导致的分割精度不足问题。方法:提出了一种结合空洞卷积和残差模块的级联肝脏肿瘤分割算法DDR-UNet++。首先,利用LiTS-2017数据集中的CT图像,通过窗位窗宽调整、直方图均衡化和高斯滤波进行预处理,减少噪声并平滑边缘。然后,采用级联肝脏分割模型,增强肝脏区域占比,减轻周围组织干扰,并解决数据不平衡问题。肝脏肿瘤分割模型通过引入可变形空洞卷积和残差网络,扩展感受野,提升特征提取能力。结果:DDR-UNet++在LiTS-2017数据集上的Dice相似系数、相对体积差异和Jaccard指数相比于U-Net模型分别提升了4.7%、1.7%和8.5%,有助于克服传统肿瘤分割中的低效性和低准确性,提高早期肿瘤发现率和患者生存率,减轻医生负担。结论:该方法通过改进模型结构与分割策略,在一定程度上改善了特征提取能力不强的问题,有效提升了肝脏肿瘤分割的精度和鲁棒性,为临床辅助诊断提供可靠的技术参考。
Objective To explore and address the issue of insufficient segmentation accuracy in liver tumor segmentation using the traditional U-Net algorithm,which is caused by the lack of contextual information for both the liver and tumor,as well as the large morphological variability of tumors.Methods A cascaded liver tumor segmentation algorithm,DDR-UNet++,which integrated dilated convolutions and residual modules was proposed.Firstly,CT images from the LiTS-2017 dataset were preprocessed through window width/level adjustment,histogram equalization and Gaussian filtering to reduce noise and smooth edges.Then,a cascaded liver segmentation model was employed to enhance the liver region proportion,mitigate interference from surrounding tissues and address data imbalance issue.For liver tumor segmentation,deformable dilated convolutions and residual networks were introduced to expand the receptive field and improve feature extraction capability.Results DDR-UNet++outperformed the traditional U-Net on the LiTS-2017 dataset,achieving improvements of 4.7%,1.7%,and 8.5%in Dice similarity coefficient,relative volume difference,and Jaccard index,respectively.These enhancements contribute to overcoming the inefficiency and low accuracy issues in conventional tumor segmentation,thereby improving early tumor detection rates,enhancing patient survival outcomes,and alleviating the diagnostic burden on clinicians.Conclusion The proposed method improves the feature extraction capability to some extent by enhancing the model structure and segmentation strategy,effectively increases the accuracy and robustness of liver tumor segmentation,and provides a reliable technical reference for clinical auxiliary diagnosis.
作者
扈蕴琨
王晓艳
王秀娟
HU Yunkun;WANG Xiaoyan;WANG Xiujuan(College of Medical Information and Artificial Intelligence,Shandong First Medical University&Shandong Academy of Medical Sciences,Tai'an 271000,China)
出处
《中国医学物理学杂志》
2025年第7期901-910,共10页
Chinese Journal of Medical Physics
基金
山东省自然科学基金(ZR2020MF156)
国家级大学生创新创业项目(202410439008)
山东第一医科大学校级教育教学改革研究项目(XM2024045)。
作者简介
扈蕴琨,研究方向:医学图像分割、深度学习在医学影像中的应用、医学人工智能,E-mail:2684489216@qq.com;通信作者:王晓艳,博士,教授,研究方向:医学物理学、医学图像处理,E-mail:xywjxc@126.com;通信作者:王秀娟,硕士,副教授,研究方向:医学图像处理,E-mail:22879867@qq.com。
