基于改进3D U-Net模型的肺结节分割方法研究  

Research on Lung Nodule Segmentation Method Based on Improved 3D U-Net Model

作　　者：石征锦[1] 李文慧[1] 高天 SHI Zhengjin;LI Wenhui;GAO Tian(Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]沈阳理工大学,辽宁沈阳110159

出　　处：《现代信息科技》2024年第13期52-55,60,共5页Modern Information Technology

摘　　要：由于肺部CT图像的特征信息复杂度较高,经典3D U-Net网络在肺结节分割方面准确率较低,存在误分割等问题。基于此,提出一种基于改进3D U-Net的网络模型。通过将加入了密集块的3D U-Net网络和双向特征网络(Bi-FPN)融合,提高了模型分割精度。同时采用深度监督训练机制,进一步提高了网络性能。在公开数据集LUNA-16上对模型进行比较实验和评估,结果显示,改进后的3D U-Net网络,Dice相似系数较原模型提高4%,分割精度为93.9%,敏感度为94.3%,证明该模型在肺结节分割精度及准确率方面具有一定的应用价值。Due to the high complexity of feature information in lung CT images,the classic 3D U-Net network exhibits low accuracy in lung nodule segmentation,leading to issues such as miss segmentation.To address this,a network model based on improved 3D U-Net is proposed.This model integrates 3D U-Net network with dense blocks with the Bidirectional Feature Pyramid Network(Bi-FPN)to improve the model's segmentation accuracy.The adoption of deep supervision training mechanism further enhances network performance.Comparative experiments and evaluations are conducted on the public dataset LUNA-16,and the results show that the improved 3D U-Net network has a 4%increase in Dice similarity coefficient,a segmentation accuracy of 93.9%,and a sensitivity of 94.3%compared to the original model.This proves that the model has certain application value in the accuracy and precision of lung nodule segmentation.

关 键 词：肺结节分割 CT 3D U-Net 双向特征网络 深度监督 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]