Scholarlink

To enhance annotation efficiency in 3D dental Cone Beam Computed Tomography (CBCT) image segmentation, this paper explores an active learning (AL) approach that leverages nnU-Net predictions to generate prompts for a specialized 3D Segment Anything Model (SAM). The objective is to minimize the annotation burden without relying on prompts during the inference phase. First, our experiments showed that AL offers similar segmentation performance with less than 20% of the original annotations. Second, random selection offers similar results than more complex sampling method with less more computing demand. Third, the predictions of nnU-Net on unannotated images provided effective prompts for the SAM model specialized in 3D medical images (i.e., SAM-Med3D). Combining these two approaches reduced the required amount of manual annotation by up to 50%. This paper paves the way for more easily obtaining new annotated datasets in the dental domain while simultaneously training a segmentation model, by leveraging SAM-like models.

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

From Prediction to Prompt: Leveraging nnU-Net Outputs to Guide SAM for Active Learning in 3D Dental Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nicolas Martin,
  • Jean-Pierre Chevallet,
  • Philippe Mulhem

摘要

To enhance annotation efficiency in 3D dental Cone Beam Computed Tomography (CBCT) image segmentation, this paper explores an active learning (AL) approach that leverages nnU-Net predictions to generate prompts for a specialized 3D Segment Anything Model (SAM). The objective is to minimize the annotation burden without relying on prompts during the inference phase. First, our experiments showed that AL offers similar segmentation performance with less than 20% of the original annotations. Second, random selection offers similar results than more complex sampling method with less more computing demand. Third, the predictions of nnU-Net on unannotated images provided effective prompts for the SAM model specialized in 3D medical images (i.e., SAM-Med3D). Combining these two approaches reduced the required amount of manual annotation by up to 50%. This paper paves the way for more easily obtaining new annotated datasets in the dental domain while simultaneously training a segmentation model, by leveraging SAM-like models.