Pioneering fully automated bony orbit segmentation: an in silico nnU-Net multicentre approach.
Authors
Affiliations (8)
Affiliations (8)
- Department of Oral and Maxillofacial Surgery, University Hospital RWTH Aachen, Aachen, Germany.
- Department of Oral and Craniomaxillofacial Surgery, University Hospital Basel, Basel, Switzerland; Medical Additive Manufacturing Research Group (Swiss MAM), Department of Biomedical Engineering, University of Basel, Allschwil, Switzerland. Electronic address: [email protected].
- Department of Maxillofacial Surgery and Radiology, Iuliu Hațieganu University of Medicine and Pharmacy, Cluj-Napoca, Romania.
- Department of Oral and Maxillofacial Surgery, University Hospital RWTH Aachen, Aachen, Germany; Institute of Medical Informatics, University Hospital RWTH Aachen, Aachen, Germany.
- Department of Diagnostic and Interventional Radiology, University Hospital RWTH Aachen, Aachen, Germany.
- Department of Oral and Craniomaxillofacial Surgery, University Hospital Basel, Basel, Switzerland; Medical Additive Manufacturing Research Group (Swiss MAM), Department of Biomedical Engineering, University of Basel, Allschwil, Switzerland.
- Department of Oral and Maxillofacial Surgery, Lucerne Cantonal Hospital, Lucerne, Switzerland.
- Institute for AI in Medicine (IKIM), University Hospital Essen, Essen, Germany.
Abstract
In maxillofacial surgery, orbital reconstruction requires precision to address both functional and aesthetic considerations arising from both acute and elective conditions. This study presents a novel, fully automated segmentation software designed specifically for the orbital floor. This software enhances surgical planning through superior accuracy, efficiency, and usability. The authors' transdisciplinary team compiled a dataset of 1004 expert-segmented orbits from computed tomography images across multiple countries, ensuring broad anatomical representation. Developed with the nnU-Net framework, the software achieved segmentation accuracy with a mean Dice similarity coefficient of 0.935 and a mean surface distance of 0.292 mm for the orbit, and a Dice similarity coefficient of 0.917 and mean surface distance of 0.287 mm for the orbital floor-all within approximately 1 s. This performance surpasses traditional manual segmentation, which averages 25 min per orbit. The system delivered consistent results across a range of imaging sources, affirming its reliability for a wide range of clinical applications. By introducing this fully automated, high-precision tool, this study pioneers advancements in AI-driven orbital reconstructions, setting new standards for patient-specific surgical planning. Further development and integration holds the key to transforming the field, ensuring ethical compliance and enhancing informed consent.