Scholarlink

In this paper, we propose a semi-supervised approach using a ResUnet50-based deep learning model for tooth instance segmentation in 2D panoramic X-ray images. Unlabeled data was leveraged through pseudo-label generation to enhance model performance. Our method was evaluated on the STS 2024 Challenge dataset, achieving an average Dice score of 79.15% at the image-level and 45.58% at the instance-level. Additionally, it demonstrated competitive performance with IoU scores of 83.90% and 39.29%, respectively. The model efficiently segmented individual teeth, although challenges remain in cases with implants and other high-density artifacts. The average running time for processing each image was 10 s, and the model was trained using an NVIDIA Tesla V100-SXM2 GPU. Despite these promising results, future work will focus on improving instance-level segmentation accuracy, particularly in challenging cases. The code is available at http://github.com/Guo777777 .

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

Efficient Semi-supervised Tooth Instance Segmentation in Panoramic X-Rays Using ResUnet50 and SAM Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xinge Guo,
  • Wenxi Liu,
  • Zihao Cui

摘要

In this paper, we propose a semi-supervised approach using a ResUnet50-based deep learning model for tooth instance segmentation in 2D panoramic X-ray images. Unlabeled data was leveraged through pseudo-label generation to enhance model performance. Our method was evaluated on the STS 2024 Challenge dataset, achieving an average Dice score of 79.15% at the image-level and 45.58% at the instance-level. Additionally, it demonstrated competitive performance with IoU scores of 83.90% and 39.29%, respectively. The model efficiently segmented individual teeth, although challenges remain in cases with implants and other high-density artifacts. The average running time for processing each image was 10 s, and the model was trained using an NVIDIA Tesla V100-SXM2 GPU. Despite these promising results, future work will focus on improving instance-level segmentation accuracy, particularly in challenging cases. The code is available at http://github.com/Guo777777 .