Link to paper
Deep convolutional neural network-based automated segmentation of the maxillofacial complex from cone-beam computed tomography: A validation study
Author links open overlay panelFlavia Preda a #, Nermin Morgan a b #, Adriaan Van Gerven c, Fernanda Nogueira-Reis a d, Andreas Smolders c, Xiaotong Wang a, Stefanos Nomidis c, Eman Shaheen a, Holger Willems c, Reinhilde Jacobs a e
a OMFS IMPATH Research Group, Department of Imaging & Pathology, Faculty of Medicine, KU Leuven & Oral and Maxillofacial Surgery, University Hospitals Leuven, Kapucijnenvoer33, BE-3000 Leuven, Belgium
b Department of Oral Medicine, Faculty of Dentistry, Mansoura University, 35516 Mansoura, Dakahlia, Egypt
c Relu BV, Kapeldreef 60, BE-3000 Leuven, Belgium
d Department of Oral Diagnosis, Division of Oral Radiology, Piracicaba Dental School, University of Campinas (UNICAMP), Av. Limeira 901, Piracicaba, São Paulo 13414‑903, Brazil
e Department of Dental Medicine, Karolinska Institutet, Box 4064, 141 04 Huddinge, Stockholm, Sweden
Abstract
Objectives
The present study investigated the accuracy, consistency, and time-efficiency of a novel deep convolutional neural network (CNN) based model for the automated maxillofacial bone segmentation from cone beam computed tomography (CBCT) images.
Method
A dataset of 144 scans was acquired from two CBCT devices and randomly divided into three subsets: training set (n = 110), validation set (n = 10) and testing set (n = 24). A three-dimensional (3D) U-Net (CNN) model was developed, and the achieved automated segmentation was compared with a manual approach.
Results
The average time required for automated segmentation was 39.1 s with a 204-fold decrease in time consumption compared to manual segmentation (132.7 min). The model was highly accurate for identification of the bony structures of the anatomical region of interest with a dice similarity coefficient (DSC) of 92.6%. Additionally, the fully deterministic nature of the CNN model was able to provide 100% consistency without any variability. The inter-observer consistency for expert-based minor correction of the automated segmentation observed an excellent DSC of 99.7%.
Conclusion
The proposed CNN model provided a time-efficient, accurate, and consistent CBCT-based automated segmentation of the maxillofacial complex.
Clinical significance
Automated segmentation of the maxillofacial complex could act as an alternative to the conventional segmentation techniques for improving the efficiency of the digital workflows. This approach could deliver accurate and ready-to-print3D models, essential to patient-specific digital treatment planning for orthodontics, maxillofacial surgery, and implant dentistry.
Graphical abstract
Introduction
The integration of digital technology within each step of a dental workflow has transformed dentistry, where the contemporary two-dimensional (2D) approaches are gradually becoming obsolete and being superseded by advanced three-dimensional (3D) digital tools. In the era of precision dental medicine, digitalization has been widely applied in the majority of dental workflows [1,2]. Although, a consensus exists around the potential benefits of using 3D digital tools for enhancing the quality of care, their universal acceptability is partially hindered by the laborious nature of certain steps requiring considerable expertise in both clinical dentistry and medical image processing tools for various tasks, such as 3D radiographic data segmentation and integration, virtual treatment planning and computer-assisted design/computer assisted manufacturing (CAD/CAM) [3,4].
The first and the most essential step in the digital workflow of the majority of digital dental workflows is known as segmentation, which involves the generation of 3D models of the dentomaxillofacial structures. The most commonly applied methodologies for segmentation are either thresholding - or template-based and semi-automatic in nature [5], [6], [7] (Fig. 1A). These techniques are prone to certain limitations, such as missing thin bony structures, excessive time-consumption [8], steep learning curve, observer variability and need for manual refinement. In the presence of metal artifacts from high-density materials, an extensive amount of manual post-processing is required by a trained operator owing to a high intensity similarity of grey values between bone and artifacts [9]. Furthermore, the currently available dentomaxillofacial segmentation software programs have been optimized based on CT data, which cannot be applied to CBCT scans due to the presence of uncalibrated absolute Hounsfield units (HU), beam hardening artifacts, inhomogeneity, noise, and low-contrast resolution [10], [11], [12]. All these factors negatively affect the quality of the scan and the accuracy of bone segmentation.
Considering the limitations of the conventional segmentation methods, recent application of deep convolutional neural networks (CNNs) has outperformed the previously available algorithms for modelling of the dentomaxillofacial region [13], [14], [15], [16], [17]. These CNNs have been successfully applied with promising results for the CBCT-based automated segmentation of the teeth, upper airway, inferior alveolar nerve canal, mandible, and maxillary sinus [18], [19], [20], [21], [22], [23]. However, a lack of evidence exists considering the CNN-based automated segmentation of the maxillofacial complex.
The maxillofacial complex holds a unique position in the workflows of orthognathic and reconstructive surgery, dental implantology and orthodontics, for ensuring an accurate diagnosis, patient-specific treatment planning (designing and manufacturing patient-specific osteotomy and repositioning guides, orthodontic devices, spacers, occlusal splints, fixation plates/implants and 3D printed models), and follow-up assessment. It is one of the most difficult anatomical regions to segment with conventional approaches owing to the anatomical complexity and reduced bone thickness, which often leads to a clinically significant under- and/or over-estimation of the segmented skeletal structure [24,25] hence requiring a laborious amount of manual corrections. Therefore, it is important to assess whether CNN-based automated segmentation of the maxillofacial complex can simplify the segmentation process by offering an accurate and observer-independent alternative to the present conventional approaches.
The present study aimed to investigate the performance of a novel deep CNN-based model for the automated maxillofacial complex bone segmentation from CBCT images. We hypothesized that a deep CNN approach would offer a more accurate, consistent and time efficient segmentation of the maxillofacial complex compared to the manual segmentation as the reference standard.
Section snippets
Materials and methods
This study was performed in accordance with the Declaration of Helsinki on medical research. Ethical approval was acquired from the Medical Ethics Committee of University Hospitals, KU Leuven, Leuven, Belgium (Reference no.: S57587). Informed consent was not required as patient information was anonymized. The study was carried out in line with the recommendations of Schwendicke et.al for reporting on artificial intelligence in dental research [26].
Timing
The average time required for manual segmentation (12 CBCT scans) was 132.7 min compared to the corresponding time of 39.07 s (203.8-fold time reduction) and 9.13 min (13.5-fold time reduction) for automated and corrected segmentation, respectively (Fig. 4).
The average time needed for the manual correction of the testing set (24 CBCT scans) by the two operators was 9.33 min and 11.11 min, respectively. The corresponding average time for automated segmentation was 43.41seconds.
Accuracy
Table 2 and Fig. 5
Discussion
An accurate segmentation of the maxillofacial complex from CBCT images is a prerequisite in the majority of dentomaxillofacial workflows for the creation of a 3D virtual patient model. The main clinical applications requiring segmented maxillofacial complex include orthognathic and reconstructive surgical treatment planning, designing of patient-specific implants, orthodontic virtual set-up and post-operative follow-up assessment of skeletal tissue [34], [35], [36], [37], [38], [39], [40], [41]
Conclusions
The proposed CNN model is an accurate, consistent, and time-efficient alternative to the conventional manual and semi-automated segmentation methods for the generation of a 3D maxillofacial complex model. An observer-independent 204-fold time reduction for the segmentation task compared to the manual approach and the integration of the model into an online platform can fit the current demands of clinical practice without the need of an experienced operator or a computer with a high
Credit authorship contribution statement
Flavia Preda: Conceptualization, Methodology, Software, Validation, Visualization, Investigation, Formal analysis, Writing – original draft, Writing – review & editing, Project administration. Nermin Morgan: Conceptualization, Methodology, Software, Data curation, Validation, Visualization, Investigation, Formal analysis, Writing – original draft, Writing – review & editing. Adriaan Van Gerven: Conceptualization, Methodology, Software, Data curation, Validation, Investigation, Writing – review
Declaration of Competing Interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:
Adriaan Van Gerven, Stefanos Nomidis and Holger Willems have professional relationship with Relu BV (ownership, development and commercial interests)
References (45)
- J. Minnema et al.Comparison of convolutional neural network training strategies for cone-beam CT image segmentation Comput. Methods Programs Biomed. (2021)
- J. Minnema et al.CT image segmentation of bone for medical additive manufacturing using a convolutional neural network Comput. Biol. Med. (2018)
- P.-.J. Verhelst et al.Layered deep learning for automatic mandibular segmentation in cone-beam computed tomography J. Dent. (2021)
- E. Shaheen et al.A novel deep learning system for multi-class tooth segmentation and classification on cone beam computed tomography. A validation study J. Dent. (2021)
- P. Lahoud et al.Artificial intelligence for fast and accurate 3-dimensional tooth segmentation on cone-beam computed tomography J. Endod. (2021)
- S. Shujaat et al.Automatic segmentation of the pharyngeal airway space with convolutional neural network J. Dent. (2021)
- F. Schwendicke et al.Artificial intelligence in dental research: checklist for authors, reviewers, readers J. Dent. (2021)
- J.M. Plooij et al.Digital three-dimensional image fusion processes for planning and evaluating orthodontics and orthognathic surgery. A systematic review Int. J. Oral Maxillofac. Surg. (2011)
- I. Azarmehr et al.Contemporary techniques in orbital reconstruction: a review of the literature and report of a case combining surgical navigation, computer-aided surgical simulation, and a patient-specific implant J. Oral Maxillofac. Surg. (2020)
- J. Uechi et al.Generation of virtual models for planning orthognathic surgery using a modified multimodal image fusion technique Int. J. Oral Maxillofac. Surg. (2015)
