Abstract:

In order to accurately segment lung nodules from CT images, an improved U-Net based method is proposed for lung nodule segmentation. The dense connection introduced into the network can not only strengthen the transmission and utilization of features, but also avoid the vanishing gradient problem. Meanwhile, an improved hybrid loss function is adopted to address class imbalance problem. The experimental results on the public LIDC-IDRI database show that the proposed method can achieve Dice similarity coefficient, precision and recall of 84.48%，85.35% and 83.81%, respectively. Compared with some segmentation methods, the proposed method can accurately segment lung nodules with good performance.

Key words: lung nodules, U-Net, dense connection, semantic segmentation

摘要：

为了对CT图像中的肺结节进行准确地分割，提出了一种基于改进的U-Net网络的肺结节分割方法。该方法通过引入密集连接，加强网络对特征的传递与利用，并且可以避免梯度消失的问题，同时采用改进的混合损失函数以缓解类不平衡问题。在LIDC-IDRI肺结节公开数据库上的实验结果表明，该方法达到的Dice相似系数值、准确率和召回率分别为84.48%、85.35%和83.81%。与其他分割网络相比，该方法能够准确地分割出肺结节区域，具有良好的分割性能。

关键词: 肺结节, U-Net, 密集连接, 语义分割