Scholarlink

<p>Medical image segmentation has significantly progressed owing to the availability of large-scale annotated datasets, yet data annotation is a labor-intensive and time-consuming process. Semi-supervised learning methods reduce this dependency on labeled data by utilizing a substantial amount of unlabeled data along with limited labeled data. Currently, one of the most promising avenues to semi-supervised learning involves enforcing consistency between weakly and strong perturbed versions of the input image. In this work, we propose a semi-supervised learning framework termed Dual Perturbation Consistency Learning (DPCL) for medical image segmentation. Our framework is based on weak-to-strong consistency learning with two modules: the Confidence-guided Adaptive Perturbation (CAP) module and the Mixed Feature Perturbation (MFP) module. Specifically, the CAP module adaptively applies image-level perturbation based on the difficulty of the input image and the learning progress of the model. The MFP module applies feature-level perturbation at multiple layers of the feature pyramid, and facilitates feature mixing from both the weak and the strong views to make the learning process easier and more effective. Extensive experiments on three MRI image segmentation datasets demonstrate that DPCL compares favorably against state-of-the-art semi-supervised medical image segmentation methods, achieving Dice score improvements of 1.24% on the ACDC dataset with 5% labeled data, 3.79% on the PROMISE12 dataset with 20% labeled data, and 0.46% on the LA dataset with 20% labeled data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual perturbation consistency learning for semi-supervised medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年1月12日

查看全文
  • Zuoyong Li,
  • Zhen Zhou,
  • Sien Li,
  • Shenghua Teng,
  • Xiang Wu,
  • Tao Wang

摘要

Medical image segmentation has significantly progressed owing to the availability of large-scale annotated datasets, yet data annotation is a labor-intensive and time-consuming process. Semi-supervised learning methods reduce this dependency on labeled data by utilizing a substantial amount of unlabeled data along with limited labeled data. Currently, one of the most promising avenues to semi-supervised learning involves enforcing consistency between weakly and strong perturbed versions of the input image. In this work, we propose a semi-supervised learning framework termed Dual Perturbation Consistency Learning (DPCL) for medical image segmentation. Our framework is based on weak-to-strong consistency learning with two modules: the Confidence-guided Adaptive Perturbation (CAP) module and the Mixed Feature Perturbation (MFP) module. Specifically, the CAP module adaptively applies image-level perturbation based on the difficulty of the input image and the learning progress of the model. The MFP module applies feature-level perturbation at multiple layers of the feature pyramid, and facilitates feature mixing from both the weak and the strong views to make the learning process easier and more effective. Extensive experiments on three MRI image segmentation datasets demonstrate that DPCL compares favorably against state-of-the-art semi-supervised medical image segmentation methods, achieving Dice score improvements of 1.24% on the ACDC dataset with 5% labeled data, 3.79% on the PROMISE12 dataset with 20% labeled data, and 0.46% on the LA dataset with 20% labeled data.