EUVIMEDEUVIMED entdecken
Crossref

Learnable Pixel-Level Reward Modeling and Pixel-Wise Probability Mixing for Semi-supervised Medical Image Segmentation

Xianmin Wang, Houren Wei, Jing Li

Journal of Imaging Informatics in Medicine · 12.11.2026

DOI: 10.1007/s10278-026-02472-3 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Imaging Informatics in Medicine, online erschienen am 12.11.2026. Verfasst von Xianmin Wang und 2 weiteren Autor:innen. Beteiligte Einrichtung: Institute of Artificial Intelligence, Guangzhou University, Guangzhou, Guangdong, China. Die Arbeit nennt 37 Literaturangaben. Gefördert von: Natural Science Foundation of Guangdong Province, Scientific Research Project of Guangzhou University.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In semi-supervised medical image segmentation, pseudo-labeling and data mixing are widely adopted to leverage unlabeled images, yet conventional methods suffer from coarse-grained operations. Specifically, global filtering strategies typically rely on heuristic confidence scores that may not accurately reflect local pseudo-label reliability and apply spatially uniform loss weights to all pixels within retained samples, failing to reliably evaluate and accommodate spatial quality variations within pseudo-labels. Meanwhile, coarse-grained region-level mixing violently disrupts the inherent continuity of anatomical structures, frequently introducing severe semantic mismatches and misleading boundary targets during training. To address these limitations, we propose a semi-supervised segmentation framework based on learnable pixel-level reward modeling and pixel-wise probability mixing. A data-driven pixel-level reward map is constructed to provide learnable pixel-level reliability estimation instead of relying solely on heuristic confidence scores. These continuous reward scores are directly converted into fine-grained pixel-wise loss weights, dynamically regulating each pixel's optimization contribution rather than relying on hard confidence-based sample filtering; an empty-foreground safeguard only suppresses degenerate unlabeled samples with negligible predicted foreground. Furthermore, local pseudo-label distributions are adaptively refined via pixel-wise probability mixing within the continuous class probability space. This mechanism softly fuses the probability distributions of low-reward pixels with those of semantically consistent high-reward donor pixels, successfully bypassing the structural constraints of region-level mixing. Extensive experiments on the ACDC, BUSI, and Synapse datasets demonstrate consistent improvements over competitive semi-supervised baselines under limited labeled-data settings.

Schlagwörter

Stichwörter der Autor:innen

Medical image segmentationPixel-wise reward modelingProbability mixingPseudo-label learningSemi-supervised learning

Autor:innen und Institutionen

  • Institute of Artificial Intelligence, Guangzhou University, Guangzhou, Guangdong, China.
  • School of Computer Science and Cyber Engineering, Guangzhou University, Guangzhou, Guangdong, China. lijing@gzhu.edu.cn.

Förderung

  • Natural Science Foundation of Guangdong Province
  • Scientific Research Project of Guangzhou University

Zitieren

APA

Wang, X., Wei, H., & Li, J. (2026). Learnable Pixel-Level Reward Modeling and Pixel-Wise Probability Mixing for Semi-supervised Medical Image Segmentation. Journal of Imaging Informatics in Medicine. https://doi.org/10.1007/s10278-026-02472-3

Vancouver

Wang X, Wei H, Li J. Learnable Pixel-Level Reward Modeling and Pixel-Wise Probability Mixing for Semi-supervised Medical Image Segmentation. Journal of Imaging Informatics in Medicine. 2026. doi:10.1007/s10278-026-02472-3

BibTeX
@article{Wang2026,
  author = {Wang, Xianmin and Wei, Houren and Li, Jing},
  title = {Learnable Pixel-Level Reward Modeling and Pixel-Wise Probability Mixing for Semi-supervised Medical Image Segmentation},
  journal = {Journal of Imaging Informatics in Medicine},
  year = {2026},
  doi = {10.1007/s10278-026-02472-3}
}
Kostenlos starten →