The use of artificial intelligence in the planning of orthognathic surgery: prototype of a trained neural network for automated surgical planning
Authors : H. Jr Vercruysse 1,2,3, E. Van de Casteele 1,2,3, J. Bertels 4, W. De Vos 1,2,3, S. Stevens 1, D. Dielen 1, D. Govaerts 1, J. Jonkergouw 1, N. Nadjmi 1,2,3, L. Renier 1,3, J. Van de Perre 1,3, M. Van Genechten 1,3, G. Van Hemelen 1,3, F. Vanhove 1, P. Winderickx 1, S. Willaert 4
1 ZMACK / Associatie MKA, AZ Monica Antwerp, Belgium
2 Department of Cranio-Maxillofacial Surgery, Antwerp University Hospital, Belgium
3 Faculty of Medicine & Health Sciences, University of Antwerp, Belgium
4 Robovision AI, Ghent, Belgium
Objectives: In recent years, artificial intelligence (AI) is increasingly being integrated into surgical planning to enhance precision and outcomes. Machine learning algorithms have shown promise in predicting postoperative skeletal changes, thereby aiding in surgical planning and potentially reducing the workload of surgeons (Ma et al., 2022). As orthognathic surgery planning remains sophisticated and highly experience-dependent, deep learning methods have been explored to predict orthognathic surgery plans from 3D cephalometric analyses, offering a more precise understanding of normal bone structures to assist in surgical planning (Cheng et al., 2023). With the large orthognathic patient turnover in our centre (>1000 cases/year) a large dataset is available to train a neural network for automated 3-dimensional planning of Caucasian patients and a prototype was developed.
Material and Methods: A neural network was trained on 250 samples, retrospectively collected from 2021 to 2022 and operated in the Department of Oral and Craniomaxillofacial Surgery of AZ Monica Hospital in Antwerp or the University Hospital of Antwerp. All patients underwent a Le Fort I osteotomy and a bilateral sagittal split ramus osteotomy (BSSO) or/and genioplasty. Pre-operative CBCT data and smile dynamics were collected for every patient. Based on this neural network, a pipeline software prototype (Robovision AI, Ghent, Belgium) was designed that could generate an automated orthognathic planning. The prediction accuracy of the prototype was validated ‘man versus machine’ for 20 cases of bi- or tri-maxillary surgery.
Results: The pipeline software has proven to produce uniform results, as the introduction of identical samples led to identical results. In the absence of a ground truth for unequivocal orthognathic planning, the software could generate an orthognathic treatment plan, that was comparable with an acceptable standard deviation to a conventional planning made by experienced surgeons with a postoperative prediction ability based on an annual turnover of over 150 cases per year.
Conclusion: The pipeline software prototype, based on a trained neural network, can perform an automated orthognathic surgical planning with high accuracy and good clinically practical effectiveness.
Nos partenaires
