基于逻辑校准的多分类残差网络的肺分割算法

针对图像噪声以及血管、支气管等因素引起的肺分割困难的问题,提出了一种基于逻辑校准的多分类残差网络分割算法.该算法将图像区域划分为肺、背景及边界三类,通过扩大不同类型间的差异来提升分割准确率.算法先将图像分割为固定尺寸区域,然后利用残差网络提取纹理特征进行分类训练与测试,实现粗分割.最后对边界区域阈值处理实现细分割.利用公开数据集对该算法进行了测试,实验结果表明,此分割算法在召回率、精确率以及交并比等方面均优于当下前沿的分割网络之一的U-Net,分别达到99.79％,98.13％和97.83％,可为后续的肺部疾病临床诊断提供参考依据.

An Algorithm of Lung Segmentation Based on Logit Adjustment in Multi-class Residual Network

In order to extract lung precisely, aiming at difficulty in segmentation of lung caused by interfering factors such as image noise, blood vessels and bronchus, an algorithm based on logit adjustment in multi-class residual network was proposed. The algorithm divided the image area into three categories:lung, background and boundary, which improves the segmentation accuracy by expanding the difference between different types of images. Firstly, the image was divided into regions with fixed size, then, a residual network was then trained to extract the texture features for classification and tested to achieve coarse segmentation. Finally, refining segmentation was conducted on regions which were marked as boundary based on threshold method. The segmentation performance of the proposed model was tested and verified by using a public dataset. The recall rate, precision and intersection over union of the algorithm were obtained as 99.79%, 98.13% and 97.83%, respectively, and the overall segmentation performance was higher than that of U-Net, one of the most cutting-edge segmentation networks. According to the experimental results, the proposed algorithm provides a reference basis for subsequent clinical diagnosis of lung diseases.