DentSeg-EDA: A 3D tooth segmentation framework for CBCT images using enhanced dual attention

The accurate three-dimensional (3D) segmentation of individual teeth from cone-beam computed tomography (CBCT) images is essential for enhancing computer-assisted planning in various dental procedures. Despite recent advances, existing methods continue to face difficulties when dealing with complex dental anatomies, particularly in scenarios with limited annotated data. In this study, we propose a novel enhanced dual-attention mechanism integrated within the skip connections of a two-stage V-Net framework to overcome these challenges and address the clinical demand for more robust and efficient solutions. In the initial stage of our pipeline, the network employs 3D convolutions to extract regions of interest, tooth skeletons, and centroids simultaneously, thereby effectively capturing spatial relationships and preserving essential anatomical details. These features are subsequently fed into the encoder of the second-stage V-Net, which performs the final segmentation of individual teeth. The incorporation of the enhanced dual attention block within the skip connections enables the model to focus on both global contextual cues and local anatomical features, leading to superior feature representations and improved boundary delineation accuracy. The proposed method was evaluated on a dataset of CBCT images. Comprehensive experiments demonstrated that our model achieved a Dice score of 92.01 % and an intersection over union of 85.35 %, outperforming the best previous method by 2.00 % and 3.13 % points, respectively, while reducing the Hausdorff distance and average surface distance, which are critical improvements for accurate tooth root and boundary segmentation in clinical practice, by 0.67 mm and 0.11 mm, respectively. This work contributes to the advancement of digital dentistry by offering a more reliable and efficient solution for 3D tooth segmentation, with practical applications in clinical settings.