基于对比学习的半监督肝脏血管分割方法  

Semi-supervised liver vessel segmentation method based on contrastive learning

作　　者：刘哲[1] 胡芮 宋余庆[1] 刘毅[1] LIU Zhe;HU Rui;SONG Yuqing;LIU Yi(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang 212013,Jiangsu China)

机构地区：[1]江苏大学计算机科学与通信工程学院,江苏镇江212013

出　　处：《华中科技大学学报（自然科学版）》2024年第5期70-75,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61976106,62276116);江苏省六大人才高峰计划资助项目(DZXX$-$122).

摘　　要：针对肝脏血管分割方法通常依赖高质量有标签数据训练模型及现存对比学习方法过度关注局部信息而忽略全局信息的问题,提出一种基于全局和局部对比学习的半监督肝脏血管分割方法.本方法首先以Mean Teacher模型作为框架,对编码器输出高维特征进行全局对比,捕捉特征全局一致性,获取更丰富全局上下文信息;然后对解码器输出进行局部对比,获取语义分割中的局部像素级特征;最后引入分布差异最小化方法,通过使用判别器进行对抗训练,减少有标签和无标签数据之间的分布差异,提升模型的泛化性能.实验结果表明:本方法在肝脏血管分割任务上取得了显著的效果,在3Dircadb数据集上的Dice系数、Jaccard系数、平均表面距离(ASD)和95%豪斯多夫距离(95HD)分别为74.36%,59.73%,2.65 mm和13.57 mm,优于其他半监督方法.Aiming at the problem that existing methods often depend on high-quality labeled data for model training,with a tendency to focus excessively on local information while neglecting global information,a semi-supervised liver vessel segmentation method was proposed based on global and local contrast learning.First,the Mean Teacher model was used as a framework to globally contrast the high-dimensional features output by the encoder,the global consistency of the features was captured,and richer global context information was obtained.Then,local contrast on the decoder output was performed to obtain the local pixel level in semantic segmentation features.Finally,the distribution difference minimization method was introduced to reduce the distribution difference between labeled and unlabeled data using the discriminator for adversarial training,which improved the generalization performance of the model.Experiment results show that the proposed method achieves remarkable results in the liver vessel segmentation task,with Dice of 74.36%,Jaccard of 59.73%,average surface distance(ASD)of 2.65 mm,and 95%Hausdorff distance(95HD)of 13.57 mm on the 3Dircadb dataset,which is superior to other semi-supervised methods.

关 键 词：肝脏血管分割 半监督学习 深度学习 对比学习 对抗学习 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]