Scholarlink

Automatic and precise tooth segmentation is crucial in computer aided dentistry, serving a pivotal role in various applications likes diagnosis and treatment planning. While prominent methods can attain satisfactory segmentation results, disproportion in the proportion of annotated images have an impact on model performance. In this paper, we present a cascade semi-supervised method named C-CPS for tooth segmentation, which designed for imbalanced datasets. C-CPS is built upon the idea that the model can acquire more essential information about uncertainty. It consists of two subtly different decoders and utilizes two distinct strategies for generating pseudo-labels. Particularly, one strategy aims at minimizing entropy and enhancing posterior probabilities, while the other constrains predictions by incorporating prior information. To boost the model training, we integrate the two generation strategies into a cycle supervision module. We evaluate C-CPS in MICCAI 2023 Challenge and respectively achieve 76.70% of Dice, 81.46% of IOU, and 16.42% normalized HD.

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

Prior-Aware Cross Pseudo Supervision for Semi-supervised Tooth Segmentation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Tingyi Lin,
  • Pengju Lyu,
  • Junchen Xiong,
  • Xiaodong Wang,
  • Kehan Song,
  • Qiong Lou

摘要

Automatic and precise tooth segmentation is crucial in computer aided dentistry, serving a pivotal role in various applications likes diagnosis and treatment planning. While prominent methods can attain satisfactory segmentation results, disproportion in the proportion of annotated images have an impact on model performance. In this paper, we present a cascade semi-supervised method named C-CPS for tooth segmentation, which designed for imbalanced datasets. C-CPS is built upon the idea that the model can acquire more essential information about uncertainty. It consists of two subtly different decoders and utilizes two distinct strategies for generating pseudo-labels. Particularly, one strategy aims at minimizing entropy and enhancing posterior probabilities, while the other constrains predictions by incorporating prior information. To boost the model training, we integrate the two generation strategies into a cycle supervision module. We evaluate C-CPS in MICCAI 2023 Challenge and respectively achieve 76.70% of Dice, 81.46% of IOU, and 16.42% normalized HD.