Trans-VNet: Transformer-based tooth semantic segmentation in CBCT images

Three-dimensional dental models hold significant value for advancing digital oral cavity development. Tooth segmentation, based on CBCT image analysis, is a pivotal step in recognizing 3D oral structures due to its comprehensive nature. However, challenges arise from image quality issues such as noise and sparse data distribution, complicating segmentation efforts. In this paper, we present a novel transformer-based framework for tooth semantic segmentation, designed for dental CBCT images. The Trans-VNet architecture integrates V-Net with transformer modules to address the challenges in 3D dental image segmentation. Unlike existing models such as V-Net, nn-UNet and Unetr++, Trans-VNet employs a patch-based region of interest extraction strategy to emphasize the tooth region while mitigating the impact of metal artifacts, which are common in dental CBCT images. Additionally, a cross-attention module is introduced to enhance the model's ability to capture comprehensive contextual information about teeth, overcoming uneven data distribution issues. The experimental results demonstrate that Trans-VNet outperforms existing models in terms of the Dice similarity coefficient (DSC) and IoU. The proposed framework shows promise for advancing tooth segmentation tasks in digital dentistry, with future research focusing on addressing limitations related to datasets with metal artifacts and further optimization for broader applicability in dental practice.