An innovative AI-based dual segmentation application for head surgery

Accurate anatomical segmentation in computed tomography (CT) imaging is vital for diagnostics and virtual surgical planning in head and neck surgery, yet manual methods remain time-consuming and inconsistent. This study presents a dual-model artificial intelligence (AI) system that automates the segmentation of key craniofacial structures. A total of 388 clinical CT scans were processed using a two-stage nnU-Net-based approach: a coarse global model (3.0 mm resolution) followed by a fine local model (0.5 mm resolution). Ground truth segmentations of the skull, mandible, teeth, sinuses, and soft tissue were created by a single annotator and compared to the AI-based segmentation results using the Dice similarity coefficient (DSC), mean surface distance (MSD), and Hausdorff distance (HD) . High accuracy was achieved: mean DSC of 0.963 for mandible and skull, and 0.986 for soft tissue. Mean MSD values were lowest for maxillary sinus (0.134 mm) and mandible (0.150 mm). While HD remained low overall, soft tissue showed high variability (mean 25.138 mm). The AI system proved robust across varied imaging protocols, including low-resolution scans. By significantly reducing manual segmentation efforts while ensuring clinical precision, this open-access dual-model framework offers a scalable solution for diagnostic, surgical, and educational applications in craniofacial care.