基于GAN-DAUnet的肝脏CT图像肿瘤分割算法  

Liver CT image tumor segmentation algorithm based on GAN-DAUnet

作　　者：周金治[1,2] 胡震 郭莉莉 龚莉 张翁荣 ZHOU Jinzhi;HU Zhen;GUO Lili;GONG Li;ZHANG Wengrong(School of Information Engineering,Southwest University of Science and Technology,Mianyang 621000,China;Sichuan Provincial Key Laboratory of Robot Technology Used for Special Environment,Mianyang 621000,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621000 [2]特殊环境机器人技术四川省重点实验室,四川绵阳621000

出　　处：《中国医学物理学杂志》2023年第8期971-976,共6页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(61771411)。

摘　　要：针对现有肝脏CT图像肿瘤分割方法中存在欠分割、过分割、边界模糊以及分割精度较低的问题,提出一种基于生成对抗网络(GAN)的肝脏肿瘤自动分割算法。该算法首先对图像进行预分割减少无关信息的影响。其次GAN的生成网络使用DAUnet,该网络在跳跃连接中引入双注意力机制增强肝脏肿瘤的特征。最后通过GAN的生成对抗训练并在训练过程中引入混合损失函数,使预测的肿瘤图像更加精准。在LiTS数据集上的实验结果表明,提出的算法Dice相似系数值(Dice similarity coefficient,DSC)达到了76.15%,相比于Unet提升3.63%。通过对DAUnet进行生成对抗训练能有效提高肝脏图像中肿瘤的分割性能。Aiming at the problems of under-segmentation,over-segmentation,boundary ambiguity,and low segmentation accuracy in the existing liver CT image segmentation methods,an automatic segmentation algorithm for liver tumors based on a generative adversarial network(GAN)is proposed.Firstly,the image is pre-segmented to reduce the effect of irrelevant information.Secondly,the GAN-generative network uses Dual Attention Unet(DAUnet),which introduces dual attention mechanisms in skip connections to enhance the features of liver tumors.Finally,the predicted tumor images become more accurate through the generative adversarial training of GAN and the introduction of the mixed loss function in the training process.The experimental results on the LiTS dataset show that the Dice similarity coefficient of the proposed algorithm reaches 76.15%,which is 3.63%higher than that of Unet.DAUnet for generative adversarial training can effectively improve the performance of tumor segmentation in liver images.

关 键 词：医学图像分割 深度学习 生成对抗网络 注意力机制 肝脏肿瘤 

分 类 号：R318[医药卫生—生物医学工程] R735.7[医药卫生—基础医学]