Prompt-based three-dimensional tooth segmentation method based on pre-trained SAM(基于预训练SAM的提示式三维牙齿分割方法)

Abstract

Currently, most studies employ supervised learning techniques to train networks on three-dimensional tooth data to perform annotation tasks. However, these methods often perform poorly on cases of missing teeth, severe misalignments, or incomplete jaw structures, exhibiting weak generalization capabilities. To address these issues, a new approach based on a pre-trained SAM large model and prompt-based segmentation techniques was introduced. Initially, the model was fine-tuned on the MICCAI 3DTeethSeg'22 public dataset. Then, the three-dimensional dental models were projected onto multiple two-dimensional views, and the SAM network was utilized for image segmentation. After segmentation, each pixel's label was mapped back to the original three-dimensional triangular facets to complete the three-dimensional tooth segmentation. This approach was tested on 900 relatively ideal 3D upper and lower teeth data in the dataset, achieving results comparable to mainstream technologies. For complex cases such as missing teeth, misaligned teeth, and incomplete upper and lower jaws, it exhibited significantly better performance than existing technologies, showcasing enhanced generalizability and stability.(目前，大多研究采用有监督学习方法在牙齿的三维数据上训练网络，完成分割任务，但在处理缺牙、严重错位或颌部不完整的牙齿时效果不佳，泛化能力较弱。为此，提出了一种基于预训练分割一切模型（segment anything model，SAM）和提示分割技术的方法。首先，在2022年国际医学图像计算和计算机辅助干预会议（MICCAI 2022）的三维牙齿公开数据集上微调模型。然后，将三维牙齿模型投影至多个二维视图，利用SAM网络进行图像分割。再将每个像素的标签映射回原始的三维三角形面片，完成三维牙齿分割。在该数据集中，测试了900个较理想的三维上下牙数据，取得了与主流技术相当的结果。对于缺牙、牙齿错位以及上下颚不完整的复杂情况，本文方法表现出显著优于现有技术的效果，展示了更强的泛化能力和稳定性。)

Keywords

• orthodontics(口腔正畸)
• three-dimensional tooth segmentation(三维牙齿分割)
• sam(sam)
• image segmentation(图像分割)