Fully automated AI-based digital workflow for orbital fracture mirroring: A multicenter approach

Orbital reconstruction after trauma or pathology requires precise restoration of anatomical symmetry to prevent functional and aesthetic impairment. Although virtual surgical planning and computer-aided design/manufacturing have significantly improved workflows, segmentation and mirroring of the orbit remain primarily manual and operator-dependent. This study developed and validated a fully automated, AI-based workflow that integrates deep learning-driven segmentation and algorithmic mirroring for improved efficiency and reproducibility in orbital reconstruction. A multicenter dataset comprising 502 cranial CT scans from Germany, Romania, and the USA was employed to train and evaluate a full-resolution 3D nnU-Net model for automated orbital segmentation. Following segmentation, a mirroring algorithm using Principal Component Analysis and Iterative Closest Point registration computed the sagittal symmetry plane. Automated outcomes were benchmarked against manual segmentation and mirroring by clinical experts, using metrics including Dice Similarity Coefficient, Mean Surface Distance, Hausdorff Distance, and angular deviation between symmetry planes. The AI-driven segmentation achieved high accuracy (mean DSC 0.936), with superior performance in non-fractured orbits (DSC up to 0.941). Angular deviation from manually defined symmetry planes was minimal (fractured: 0.95° ± 0.65°; non-fractured: 0.72° ± 0.53°), consistent across institutions. This automated approach demonstrated robust, clinically applicable segmentation and mirroring accuracy, enabling standardized and efficient orbital reconstruction workflows.