Deep Learning Based Automatic 3D Segmentation of The Jaws from CBCT for Patient-Specific Implants Planning in Orthognathic Surgery

Orthognathic surgery, a crucial procedure for correcting jaw deformities, relies on precise jawbone segmentation to ensure effective virtual surgical planning and optimal patient outcomes. This study presents the development of a deep learning-based 3D segmentation model that incorporates customized preprocessing techniques to effectively segment the jawbone from 3D CBCT reconstruction images, an essential component of virtual orthognathic surgery planning. A dataset of 100 CBCT images was retrospectively collected from Dental University Hospital to facilitate this development. A U-Net Transformer architecture was employed to create the automatic deep learning-based 3D segmentation model. The preprocessing approach utilizes CBCT-specific characteristics as normalization parameters, effectively reducing variability in relative Hounsfield Unit (HU) values associated with these images. The model's segmentation accuracy was evaluated using Dice scores and average surface deviation metrics. The model demonstrated impressive segmentation performance, achieving a Dice score of 0.95 for the mandible and 0.85 for the maxilla. Furthermore, it maintained an average surface deviation of less than 0.2 mm across all cases, outperforming traditional methodologies that may experience deviations up to 1 mm. In conclusion, normalizing using CBCT-specific characteristics significantly reduces variability, thereby enhancing segmentation accuracy and surface precision in targeted regions, which is crucial for achieving optimal outcomes in virtual orthognathic surgery planning.