基于改进的 V-Net 模型肺结节分割算法的研究-《计算机技术与发展》

文章信息/Info

Title:: Research on Lung Nodule Segmentation Algorithm Based on Improved V-Net Model

作者:: 李丽¹ ; 林晓明² ; 彭丰平¹ ; 潘家辉¹; 1. 华南师范大学软件学院,广东佛山 528225;
2. 广东佛山联创工程研究生院,广东佛山 528311

Author(s):: LI Li1 ; LIN Xiao-ming2 ; PENG Feng-ping1 ; PAN Jia-hui1; 1. School of Software,South China Normal University,Foshan 528225,China;
2. Guangdong Foshan Lianchuang Engineering Graduate School,Foshan 528311,China

Keywords:: ×pulmonary nodule segmentation; V-Net network; joint loss function; multi-scale convolution; SE modules

摘要:: 由于 CT 图像是三维图像,在原始的 V-Net 模型分割中,易出现结节漏检和边界分割不清晰,以及损失函数 Dice 训练时不稳定等问题。根据这些问题,提出 3D 多尺度 SE V-Net,
简称 MSEV-Net 网络,同时通过联合损失函数来提高训练的稳定性。该网络模型在 V-Net 网络的基础上,使用多尺度卷积模块来替换原有的 5 ×5 ×5 卷积,同时在残差连接后加
入SE 通道注意力模块,通过不同尺度的特征融合和学习不同通道之间的关系,解决肺结节小不易分割的问题。同时在 V-Net 网络残差连接基础上加一条短跳跃连接,使得整个网
络更好利用全局特征。联合损失函数选择 Dice 和交叉熵损失函数进行融合,可以很好地解决训练不稳定问题。提出的 MSEV-Net 网络模型和联合损失函数在平均分割准确率?
PA 达到0. 998,DSC 达到 0. 837。实验结果表明,该方法在提高肺结节分割精度方面具有一定的效果。

Abstract:: Since CT images are three-dimensional images,in the original V-Net model segmentation,it is prone to the problems of nodalomission and unclear boundary segmentation,
as well as the instability during the training of the loss function Dice. According to theseproblems,we propose 3D multi-scale SE V-Net referred to as MSEV-Net network,while improving the stability of training by joint lossfunction. This network model is based on the V-Net network,using the multi-scale convolution module to replace the original 5 ×5 ×5convolution,while adding the SE channel attention module after the residual connection to solve the problem of small lung nodules thatare not easy to segment by fusing features of different scales and learning the relationship between different channels. At the same time,ashort jump connection is added on top of the residual connection of the?
V-Net network,which makes the whole network better utilize theglobal features. The joint loss function selects Dice and cross - entropy loss function for fusion, which can well solve the problem oftraining instability. The MSEV-Net network model and joint loss function proposed reach 0. 998 in the average segmentation accuracyPA and 0. 837 in the DSC. The experimental results show that the proposed method is effective in improving the segmentation accuracyof lung nodules.