Scholarlink

<p>Semi-supervised medical image segmentation(SSMIS) aims to efficiently extract and utilize supervisory signals from a small amount of labeled data and a large volume of unlabeled data. While existing methods employ contrastive learning in latent space to explore representation information, they neglect inherent geometric structures and solely rely on logits supervision from model’s output, leading to suboptimal representation utilization. To address these limitations, we propose a novel mutual learning framework to Enhance Representation Supervision(ERS) effectively. Our method has two highlights. <b>i)</b> We propose a Compactness-Aware Contrastive Learning (CACL) strategy that quantifies feature compactness via global-local distance matrices derived from cluster structures, selecting sparse-region features as optimization targets and high-compactness features as cluster proxies to enhance geometric awareness. <b>ii)</b> We propose a Representation-Logits Alignment(RLA) module to align logit and representation spaces through prototype-guided pseudo-label generation and uncertainty-based calibration, enabling cross-space information exchange and collaborative supervision. Results on three public datasets(LA, ACDC and MSD Prostate) demonstrate the superiority of our method compared to state-of-the-art approaches.</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

ERS: Enhancing representation supervision for contrastive-based semi-supervised medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年9月26日

查看全文
  • Shiheng Zhang,
  • Yong Liu,
  • Baoqi Yu,
  • Junjun Li

摘要

Semi-supervised medical image segmentation(SSMIS) aims to efficiently extract and utilize supervisory signals from a small amount of labeled data and a large volume of unlabeled data. While existing methods employ contrastive learning in latent space to explore representation information, they neglect inherent geometric structures and solely rely on logits supervision from model’s output, leading to suboptimal representation utilization. To address these limitations, we propose a novel mutual learning framework to Enhance Representation Supervision(ERS) effectively. Our method has two highlights. i) We propose a Compactness-Aware Contrastive Learning (CACL) strategy that quantifies feature compactness via global-local distance matrices derived from cluster structures, selecting sparse-region features as optimization targets and high-compactness features as cluster proxies to enhance geometric awareness. ii) We propose a Representation-Logits Alignment(RLA) module to align logit and representation spaces through prototype-guided pseudo-label generation and uncertainty-based calibration, enabling cross-space information exchange and collaborative supervision. Results on three public datasets(LA, ACDC and MSD Prostate) demonstrate the superiority of our method compared to state-of-the-art approaches.