Scholarlink

This study presents a Deformable Inherent Consistent Learning (DICL) network for tooth segmentation in dental panoramic radiographs, addressing the clinical need for accurate diagnostic tools amid the complexity of dental diseases. The DICL network enhances segmentation accuracy by effectively learning from a limited labeled dataset, guiding robust categorical representation of teeth. It employs deformable convolution to capture detailed dental features and a two-stage training strategy, transitioning from semi-supervised to fully supervised learning, optimizing performance with labeled and pseudo-labeled images. Achieving scores of 84.45% (image-level Dice), 73.60% (image-level IoU), 88.57% (image-level NSD), 22.36% (instance-level Dice), 57.39% (instance-level IoU), 69.85% (instance-level NSD) and 65.82% IA on the official test set, our method demonstrates superior segmentation effectiveness, especially in boundary and root segmentation, outperforming algorithms like U-Net. This study’s contributions advance dental image segmentation, improving diagnostic efficiency and reducing errors. The code is available at https://github.com/Dew026/DICL.git .

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

Deformable Inherent Consistent Learning Network for Accurate Tooth Segmentation in Dental Panoramic Radiographs

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xinxu Cai,
  • Yisong Zhang,
  • Zeyuan Guan,
  • Qi Sun,
  • Zhenshen Qu

摘要

This study presents a Deformable Inherent Consistent Learning (DICL) network for tooth segmentation in dental panoramic radiographs, addressing the clinical need for accurate diagnostic tools amid the complexity of dental diseases. The DICL network enhances segmentation accuracy by effectively learning from a limited labeled dataset, guiding robust categorical representation of teeth. It employs deformable convolution to capture detailed dental features and a two-stage training strategy, transitioning from semi-supervised to fully supervised learning, optimizing performance with labeled and pseudo-labeled images. Achieving scores of 84.45% (image-level Dice), 73.60% (image-level IoU), 88.57% (image-level NSD), 22.36% (instance-level Dice), 57.39% (instance-level IoU), 69.85% (instance-level NSD) and 65.82% IA on the official test set, our method demonstrates superior segmentation effectiveness, especially in boundary and root segmentation, outperforming algorithms like U-Net. This study’s contributions advance dental image segmentation, improving diagnostic efficiency and reducing errors. The code is available at https://github.com/Dew026/DICL.git .