AI-based three-dimensional identification of safe harvesting area of perpendicular plate of ethmoid in rhinoplasty

Background: In rhinoplasty, using the perpendicular plate of the ethmoid (PPE) as a graft offers advantages such as strong support and avoiding costal cartilage use, but identifying a safe harvesting area (inferior one-third portion) is challenging for surgeons. This study aimed to develop an AI-based 3D segmentation model for precise identification of PPE and safe harvesting area from computed tomography (CT) images. Methods: CT data from 200 patients were manually segmented in Mimics 20.0 to define PPE and its safe harvesting area, with boundaries based on anatomical landmarks. The robust medical image segmentation model, nnU-Net framework, was trained on 121 samples, validated on 31 samples, and tested on 39 samples. Performance was evaluated using the Dice coefficient, Jaccard index, Hausdorff distance, mean surface distance, and segmentation time compared to manual methods. Results: The model achieved mean Dice coefficients of 0.770 (PPE) and 0.713 (safe area), with Jaccard indices of 0.635 and 0.576, respectively. AI-based segmentation reduced time by 38.7% compared to manual segmentation. Visual overlaps between AI and manual segmentations were satisfactory. Conclusions: The nnU-Net model enables efficient and accurate 3D segmentation of PPE and its safe harvesting area, and holds the potential for use in preoperative planning to reduce surgical risks. Future integration with augmented reality navigation systems may enhance intraoperative precision.