【摘要】 目的 肺部CT图像中血管的分割有助于疾病的诊断以及为手术导航提供重要参考。针对肺部血管分割任务中标注数据的稀缺、血管的形态复杂、动静脉血管灰度相近等问题，本文提出了一种对比半监督分割模型(semi-supervised contrastive learning and interpolation cross-consistency, Semi-CLIC)用于肺部血管的自动分割，辅助医生进行诊断。方法 设计一种基于非参数的动态记忆库的对比学习策略，通过对记忆库中的特征进行对比，提高同类特征的相似性和不同类特征的差异性。为了更好地利用未标注数据的先验信息和提升模型的泛化性能，本算法对未标注数据进行插值扰动，并结合隐式形状感知和交叉伪监督来构建一致性约束。最后采用来自公开数据集CARVE14的55个CT图像，以Dice系数为主要评价指标，对Semi-CLIC和其他8种算法在肺部动静脉血管上的分割性能进行对比实验。结果 在CARVE14数据集上使用两种标注比例(即10%和20%),该模型的平均Dice得分分别为69.4%和71.4%,与现有的最优半监督算法相比分别提高了1.5%和0.9%。结论 半监督学习可以在只使用少量标注数据的情况下得到与全监督学习相近的泛化性能，是缓解医学图像标注数据稀缺问题的有效方法。更多还原

【Abstract】 Objective The segmentation of vessels in lung CT images helps in the diagnosis of diseases and provides an important reference for surgical navigation. For the current problems of scarcity of labeled data, complex morphology of blood vessels, and similar grayscale of pulmonary artery and pulmonary vein vessels in the task of lung blood vessel segmentation, we propose a Contrast learning based semi-supervised segmentation framework for automatic segmentation of pulmonary vessels. Assist the doctor in diagnosis. Methods The model combines contrast learning and semi-supervised learning, we design a non-parametric dynamic memory-based contrast learning strategy to efficiently achieve improve the similarity between intra-class features and to increase distances between inter-class features by comparing features in the memory bank. This method performs interpolation operations on unlabeled data and combines implicit shape awareness and cross-pseudo-supervision to construct consistency constraints. Finally, 55 CT images from the open data set CARVE14 were used, and Dice coefficient was used as the main evaluation index to perform a comparative experiment on the segmentation performance of pulmonary arteriovenous vessels between Semi-CLIC and other 8 algorithms. Results Using two labeling ratios(10% and 20%) on the CARVE14 dataset, the average Dice scores of the proposed model were 69.4% and 71.4%,respectively, which were 1.5% and 0.9% higher than the best existing semi-supervised algorithm. Conclusions Semi-supervised learning can obtain generalization performance similar to that of fully supervised learning when only a small amount of labeled data is used, which is an effective method to alleviate the scarcity of medical image labeled data.更多还原