摘要
医学图像分割对疾病诊断至关重要,其性能高度依赖于高质量的标注数据。然而,医学图像标注数据稀缺且昂贵,严重制约了深度学习模型的训练效果。半监督医学图像分割可以减少对标注数据的依赖,但伪标签质量不佳会引入噪声,导致模型收敛到次优解。本文提出了一种半监督医学图像分割方法,通过结合伪标签和一致性正则化的策略,显著提升了分割性能。具体而言,通过设计相互学习的解码器框架,摒弃了通过阈值筛选不可靠伪标签的方式,将两个解码器对比学习以获得更为可靠的伪标注,从而提高伪标签的质量,并引入双分支强扰动模块来使两个不同的强扰动分支相互学习从而更充分探索图像级的扰动空间。本文在ACDC数据集上进行了实验验证,结果表明,所提出的方法在仅使用10%标注数据的情况下,Dice系数提升了30.70个百分点,95HD和ASD分别降低了38.90和14.22,分割性能显著提高,验证了所提出方法的有效性和优异性。
Medical image segmentation is essential for disease diagnosis,whose performances are dependent on the high quality of the labeling data.However,medical image labeling data is scarce and expensive,which severely constrains the training effectiveness of deep learning models.Semi-supervised medical image segmentation can reduce the dependence on labeled data,but the poor quality of pseudo-labels will introduce noise,causing the model to converge to the suboptimal solution.In this paper,a semi-supervised medical image segmentation method is proposed,which obviously improves the segmentation performance by combining reliable pseudo-labels and consistent regularization strategies.Specifically,a decoder framework for mutual learning is designed,the method of filtering unreliable pseudo-labels by threshold value is discarded,and the two decoders are contrasted to obtain more reliable pseudo-labels,thereby improving the quality of pseudo-labels.Moreover,a two-branch strong perturbation module is introduced to make two different strong perturbation branches learn from each other to explore the disturbance space of image level more fully.Experimental verification is conducted on the ACDC dataset,and the results show that when only 10%of the labeled data is used in the proposed method,the Dice coefficient is increased by 30.70 percentage points,the 95HD and ASD are reduced by 38.90 and 14.22,respectively,and the segmentation performance is significantly improved,which verifies the effectiveness and superiority of the proposed method.
作者
缪益民
曹立佳
汪毅
MIAO Yimin;CAO Lijia;WANG Yi(School of Computer Science and Engineering,Sichuan University of Science&Engineering,Yibin 644000,China;School of Automation and Information Engineering,Sichuan University of Science&Engineering,Yibin 644000,China;Intelligent Perception and Control Key Laboratory of Sichuan Province,Yibin 644000,China;Zigong Fourth People’s Hospital,Zigong 643000,China)
出处
《四川轻化工大学学报(自然科学版)》
2025年第4期58-67,共10页
Journal of Sichuan University of Science & Engineering(Natural Science Edition)
基金
中国高校产学研创新基金项目(2021ZYA11002)
四川省科技计划项目(2024NSFSC2048)
企业信息化与物联网测控技术四川省高校重点实验室开放基金项目(2023WYY01)
四川轻化工大学科研创新团队计划项目(SUSE652A011)
四川轻化工大学研究生创新基金项目(Y2023110)。
关键词
半监督学习
医学图像分割
伪标签
一致性正则化
semi-supervised learning
medical image segmentation
pseudo-label
consistency regularization
