基于改进DenseUnet的腹膜后淋巴CT分割方法  

Retroperitoneal Lymphatic CT Segmentation Method Based on Improved DenseUnet

作　　者：肖建英 刘志勤[1] 王庆凤[1] 黄俊[1] 周莹[2] XIAO Jianying;LIU Zhiqin;WANG Qingfeng;HUANG Jun;ZHOU Ying(School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621010,Sichuan,China;Mianyang Central Hospital,Mianyang 621000,Sichuan,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621010 [2]绵阳市中心医院,四川绵阳621000

出　　处：《西南科技大学学报》2022年第3期60-69,共10页Journal of Southwest University of Science and Technology

摘　　要：针对腹部CT图像中淋巴与周围结构相似、对比度低、难识别导致淋巴分割精度较低的问题,提出一种基于空洞卷积和双通道注意力机制的改进密集U型对称语义分割模型D-DenseUnet。将空洞卷积和双通道注意力机制嵌入密集块连接的U型网络结构中,提升模型提取整体特征的能力;为了缓解数据不平衡问题,采用复合损失函数作为改进的密集U型模型D-DenseUnet的损失函数,结合数据扩增和早停法防止过拟合,通过余弦退火衰减学习策略进行优化,最终实现腹膜后淋巴分割。实验结果表明,所提的分割模型在腹膜后淋巴CT图像中能够较好分割淋巴,平均相似系数、交并比和召回率分别为0.796,0.804,0.679,优于传统的Unet网络和密集U型DenseUnet网络。Aiming at the problem of low accuracy of lymph segmentation in abdominal CT image due to the similarity of lymph,surrounding structure and low contrast,so as to result in difficulty in recognition,an improved dense U-symmetric semantic segmentation model D-Denseunet based on dilated convolution and dual attention mechanism was proposed.Dilated convolution and dual attention mechanism were embedded into the U-shaped network structure connected by dense blocks to improve the overall feature extraction capability of the model.To solve the problem of the data imbalance,the compound loss function was used as the loss function of the modified dense U-shaped D-Denseunet,using data amplification and early stop method avoided overfitting,and cosine annealing attenuation learning strategy optimized the process to realize the retroperitoneal lymphatic segmentation.Experimental results show that the proposed segmentation model can better segment lymph nodes in retroperitoneal lymph node CT images,with average DSC,Rcall and IoU reaching 0.796,0.804 and 0.679,respectively,superior to traditional Unet network and dense U-shaped DenseUnet network.

关 键 词：深度学习 图像分割 注意力机制 腹部淋巴 DenseUnet 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]