Scholarlink

The 2D panoramic X-ray image examination is an efficient way for dentists to determine invisible caries, impacted teeth, and supernumerary teeth among children. However, identifying and manually annotating teeth in panoramic X-ray images is time consuming and labor intensive, which limits the availability of labeled cases and hinders the development of deep learning algorithms for teeth segmentation and disease analysis. To address this problem, we propose a self-training pipeline for semi-supervised 2D teeth instance segmentation. Specifically, we propose a pseudo-label generation strategy based on model ensemble and morphological operations to generate more reliable pseudo-labels for self-training. Furthermore, we address the overlapping regions between teeth by expanding the output channels of the segmentation model. Our method achieves average scores on the validation set of 77.55% for instance-level DSC, 89.57% for image-level DSC, 82.05% for instance-level NSD, 93.12% for image-level NSD, 69.02% for instance-level mIoU, 81.19% for image-level mIoU, and 79.36% for identification accuracy. The inference speed of our method is 0.161 s per image on a NVIDIA GeForce RTX 4090 GPU. Ultimately, we ranked second on the test set among all submitted teams. The code is available at https://github.com/Liaaaar/2024-MICCAI-STS-2D .

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

A Self-training Pipeline for Semi-supervised 2D Teeth Instance Segmentation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Kaiwen Fu,
  • Chengyuan Chang,
  • Jiahui Chen,
  • Qinjie Hu

摘要

The 2D panoramic X-ray image examination is an efficient way for dentists to determine invisible caries, impacted teeth, and supernumerary teeth among children. However, identifying and manually annotating teeth in panoramic X-ray images is time consuming and labor intensive, which limits the availability of labeled cases and hinders the development of deep learning algorithms for teeth segmentation and disease analysis. To address this problem, we propose a self-training pipeline for semi-supervised 2D teeth instance segmentation. Specifically, we propose a pseudo-label generation strategy based on model ensemble and morphological operations to generate more reliable pseudo-labels for self-training. Furthermore, we address the overlapping regions between teeth by expanding the output channels of the segmentation model. Our method achieves average scores on the validation set of 77.55% for instance-level DSC, 89.57% for image-level DSC, 82.05% for instance-level NSD, 93.12% for image-level NSD, 69.02% for instance-level mIoU, 81.19% for image-level mIoU, and 79.36% for identification accuracy. The inference speed of our method is 0.161 s per image on a NVIDIA GeForce RTX 4090 GPU. Ultimately, we ranked second on the test set among all submitted teams. The code is available at https://github.com/Liaaaar/2024-MICCAI-STS-2D .