基于坐标信息与多尺度并行网络的气道分割方法-《中国医学物理学杂志》

文章信息/Info

Title:: Airway segmentation method based on coordinate information and multi-scale parallel network

作者:: 刘卫朋1; 2; 3; 李健1; 2; 祁业东2; 3; 任子文2; 3; 王源2; 3; 1.河北工业大学生命科学与健康工程学院，天津 300401； 2.省部共建电工设备可靠性与智能化国家重点实验室，天津 300401； 3.河北工业大学人工智能与数据科学学院，天津 300401

Author(s):: LIU Weipeng1; 2; 3; LI Jian1; 2; QI Yedong2; 3; REN Ziwen2; 3; WANG Yuan2; 3; 1. School of Health Sciences and Biomedical Engineering, Hebei University of Technology, Tianjin 300401, China 2. State Key Lab of Reliability and Intelligence of Electrical Equipment, Tianjin 300401, China 3. School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China

Keywords:: Keywords: airway segmentation coordinate information multi-scale feature aggregation parallel network

摘要:: 为解决手术导航中气道模型精度不足的问题，提出了一种基于坐标信息与多尺度并行网络的气道分割方法。首先通过并行网络分别学习不同尺度的气道特征，以解决不同尺寸气道之间的特征冲突问题。其次提出坐标引导的上采样模块，通过浅层特征中的坐标信息指导深层特征进行特征重建，限制目标的空间位置，提高模型精度。最后提出通道引导的多尺度特征聚合模块，用于在多个尺度上捕获语义信息并探索不同尺度特征之间的通道关系。在公开数据集LIDC-IDRI和EXACT09上对提出的方法和其他模型进行训练和测试。实验表明，该方法的平均骰子系数达到了93.20%，相比于3D U-Net提高了2.61%，而假阳性率只有0.012%。此外，树长检测率和分支检测率分别达到了88.59%和97.42%。该方法可用于肺部疾病诊断或导航支气管检查等领域。

Abstract:: Abstract: An airway segmentation method based on coordinate information and multi-scale parallel network is proposed to solve the problem of insufficient accuracy of airway model in surgical navigation. Airway features at different scales are learned separately by a parallel network to address the feature conflict arising from airways of different sizes. Then, a coordinate guided up-sampling module is designed to utilize coordinate information from shallow features for guiding reconstruction of deeper features, thus restricting the spatial location of the target and improving the model accuracy. Finally, a channel guided multi-scale feature aggregation module is constructed to capture semantic details across multiple scales and investigate channel relationships between features at different scales. The proposed method and other models are trained and tested on two public datasets, namely LIDC-IDRI and EXACT09. Experimental results show that the proposed method achieves an average Dice coefficient of 93.20% which is 2.61% higher than 3D U-Net, a false positive rate of only 0.012%, a tree length detection rate of 88.59%, and a branch detection rate of 97.42%, demonstrating that the method can be applied to lung disease diagnosis or navigation bronchoscopy.