Assessment of Efficacy and Accuracy of Segmentation Methods in Dentomaxillofacial Imaging- A Systematic Review

doi:10.21203/rs.3.rs-3958673/v1

Introductions: Radiographic image segmentation is a process that aims to distinguish the voxels and pixels within a defined area of interest from the background, which is vital for facilitating clinical diagnosis, treatment planning, intervention, and follow-up in the field of dentistry and medicine.

Objectives: We aimed to provide an assessment of the efficacy and accuracy of segmentation methods in dentomaxillofacial imaging through a systematic review.

Methods: PubMed and Scopus electronic databases were searched from January 1999 to January 2023. The keywords used for the search were combinations of the following terms for each database: Artificial intelligence, Segmentation, Image interpretation, Deep Learning, Convolutional neural networks, and Head and neck imaging. After the initial search, eligible studies were selected based on the inclusion criteria, and quality assessments were conducted by A Revised Tool for the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2).

Results: Primary electronic database searches resulted in 2763 articles. Finally, a total of 54 records were considered suitable for this systematic review. Twenty-five (46%) used CBCT as a baseline imaging modality, 11 used MDCT (20%), 6 used panoramic (11%), 4 used micro-CT (7%), 3 used periapical (6%), 3 used MRI (6%), and 2 used ultrasonography (4%). Segmentation through automatic algorithms (convolutional neural networks (CNNs) and artificial neural networks (ANNs) mostly including U-net and MS-D net) was used in the majority of the studies.

Conclusion: The systematic review of the current segmentation methods in dentomaxillofacial radiology shows interesting trends, with the rising popularity of deep learning methods over time. However, Continued efforts will be necessary to improve algorithms.

Artificial intelligence

Segmentation

Image interpretation

Deep learning

Convolutional neural networks

Head and neck imaging

Radiographic image segmentation aims to differentiate the voxels and pixels of a specific area of interest (an organ or a lesion) from the image background.(Gillot et al. 2022) Image segmentation is an essential stage in supporting clinical diagnosis, treatment planning, intervention, and follow-up in dentistry and medicine.(Gaviño Orduña et al. 2020)

Segmentation methods can be categorized as the following: manual, semiautomatic, and Automatic. Manual segmentation is performed by the specialist. In this method, the clinician expert annotates the voxels and encircles the region of interest. This is usually regarded as the gold standard for segmentation.(Parmar et al. 2014) This can have some drawbacks; for instance, detailed and accurate manual segmentation of the specific structures requires visual effort, and specialized training, and is a time-consuming, fatiguing, user-dependent task with inter-and intra-observer variability.(Cover et al. 2018)

The semi-automatic method tries to bridge some of the gaps of manual segmentation through the implementation of some algorithms.(Nazem-Zadeh et al. 2012) User-dependent tasks and time needed will be reduced; However, human specialist intervention will still be needed for initialization, finalization, and accuracy checking. (A survey of user interaction and automation in medical image segmentation methods) Moreover, variability between specialists will still be present.(Heye et al. 2013)

Automatic segmentation algorithms are beneficial because they are user-independent.(Pérez de Alejo et al. 2003) The primary commonly used methods were machine learning-based, such as atlas-based registration models or c-means clustering (Chilali et al. 2016). However, in recent years, the most common methods are deep learning-based, primarily with convolutional neural networks (CNN). With the emergence of Fully Convolutional Networks (FCN), some of the problems relating to the CNNs were solved and the segmentation efficiency and accuracy were improved.(Olabarriaga and Smeulders 2001) U-net FCN architecture has been extensively used in medical image segmentation.(Ronneberger et al. 2015)

Many reviews are reporting the performance of image segmentation methods in the medical field.(Cover et al. 2018) However, To the best of our knowledge, no study has previously comprehensively reviewed the performance of segmentation methods in the field of maxillofacial and dental radiology.

This paper aims to provide an assessment of the efficacy and accuracy of segmentation methods in dentomaxillofacial imaging by systematically outlining, analyzing, and categorizing the relevant publications in this field to date, highlighting the current state, and making recommendations for future research in the area.

This systematic review was performed with the guidance of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology. (Page et al. 2021) and registered in “PROSPERO: International prospective register of systematic reviews” (registration number: CRD42022383569).(Fig. 1)

The targeted main question of the study was “What are the efficacy and accuracy of different segmentation methods in dentomaxillofacial radiology?”

The PICO elements are mentioned in the Table 1.

Table 1

The PICO elements used in this study
Population	Oral and maxillofacial radiographs (Periapical, Bitewing, Panoramic, Lateral cephalometry, CBCT, MDCT, MRI, ultrasonography)
Intervention	AI segmentation algorithms trained with dental and maxillofacial radiology imaging
Comparator	Reference protocols (Examiner's judgment, Clinical/radiological /histopathological examination)
Outcome	Performance evaluation of the AI models, including Intersection over union, Dice similarity coefficient, point-to-curve, segmentation quality, recognition quality, accuracy, and mean difference from reference.

Regarding the study design type, original retrospective or prospective English studies were considered for this review. The Inclusion criteria for the articles were: (1) original articles relevant to AI segmentation applications in dental fields, (2) clinical trials/nonclinical trials/observational studies, and (3) English articles; and the exclusion criteria were: systematic reviews, review articles, letters to editors, commentaries, non-English studies, and grey literature. PubMed and Scopus electronic databases were searched from January 1999 to January 2023 with the following search terms:

The keywords used for the search were combinations of Medical Subject Headings (MeSH) terms for each database, including: “Artificial intelligence”, “Image interpretation”, “Image Interpretation, Computer-Assisted”, “Radiographic Image Interpretation, Computer-Assisted neural networks”, “Neural Networks, Computer”, “Dentistry”. All searched articles were added to Rayyan's (Ouzzani et al. 2016) website after the initial search. The articles were then filtered by the titles thoroughly and duplicates were deleted. After the exclusion of unrelated studies, the abstracts of the remaining articles were scanned to include studies for further full-text reading. A manual search was conducted and more relevant articles were added. Furthermore, full texts of the included studies were screened to record the findings.

The data items extracted from the studies included imaging modality, author, year of publication, segmented region, dataset size, segmentation protocol and method, metric findings, and conclusion. Also for the studies with advanced imaging techniques, scanning features were extracted. The fundamental outcome of interest was the current performance of AI algorithms in the segmentation of different structures in dental and maxillofacial diagnostic imaging. The methodological quality was assessed independently with the use of the QUADAS-2 tool, and any conflict was resolved by discussion.(Whiting et al. 2011)

Primary electronic database searches resulted in 2763 articles. Finally, a total of 54 records were considered suitable for this systematic review.(Fig. 1)

Study characteristics are summarized in Tables 2 and 3, which include: the imaging modality, the segmented region, the dataset size, the scanning features, the segmentation protocol and method, the software, the reference test, the architecture, the metric findings, and the conclusion. Table 2 includes the articles that developed segmentation models based on conventional dentomaxillofacial imaging techniques i.e. periapical and panoramic images. Table 3 shows the studies in which the developed algorithm was based on advanced radiographic imaging techniques including CBCT, MDCT, Micro-CT, ultrasonography, and MRI modalities.

Table 2

Characteristics of the models developed on conventional dentomaxillofacial imaging among the included studies
Author-Year	Segmented region	Dataset size		Segmentation Protocol and Method	Reference Standard	Metric Findings	Conclusion
		Training	Testing
PA
Mori et al. 2021[1]	Assessing the technical quality of positioning in periapical radiography of the maxillary canines	training = 350 validation = 70 test = 80		Classification = AlexNet Automatic segmentation = U-net	Experienced radiologist (Manual)	Classification = SE = 0.925 SP = 0.825 ACC = 0.875 AUC = 0.927 Segmentation = R = 0.925 P = 0.961 F = 0.943	The Deep Learning system may have a potential application in assessing the quality of technical positioning in intra-oral radiographs.
Li et al. 2006[2]	Computer Aided Dental X-rays Analysis (automatic pathological segmentation)	-		Automatic segmentation = SVM	Manually chosen representative images are segmented by hierarchical level set region detection	The proposed framework can accelerate the level set segmentation about 10 times.	The Model could effectively reduce the time consumed in clinical evaluations by providing indications of possible problem areas of bone loss and decay to the clinicians.
Li et al. 2005[3]	Semi-automatically detect areas of bone loss and root decay	50		Automatic segmentation = SVM		-	The proposed model could detect the areas of bone loss automatically. Also, it may be able to automatically find the root decay with a seriousness level marked for diagnosis, when given the orientation of the teeth.
Panoramic Jaws
Cha et al. 2021[4]	Maxillary sinus, maxilla, mandible, mandibular canal, normal teeth, treated teeth, and dental implants	training = 30 validation = 11 test = 10		Automatic segmentation= Panoptic DeepLab (DNN)	A dental practitioner and a radiologist	P = 0.658 PQ = 80.47 SQ = 85.73 RQ = 93.90 IOU = 0.795	The model could detect and segment various structures in dental panoramic radiographs. Compared to manual techniques, the proposed model achieved high Dice similarity, specificity, and sensitivity.
Abdi et al. 2015[5]	Mandible	Test: 95		N/A	Manual segmentations	DSC = 93.22%±1.52 SP = 94.68 ± 2.17 SE = 94.44%±2.29	This is an accurate method that can assist in diagnosis. The presented automated model could detect almost 90% of the ground truth pixels of the maxilla, almost 80% of the mandibular canal, and 96% of the maxillary sinus, correctly.
Teeth
Vranckx et al. 2020[6]	Molar segmentation	Total:838 Training:588 Validation:250		Automatic segmentation= CNN-ResNet-101	Human reference measurements	IOU = 0.880 P = 0.940 R = 0.930 Mean HD = 19.2mm	The algorithm is time-efficient, accurate, and consistent and can assist in increasing diagnostic accuracies.
Leite et al. 2020[7]	Tooth detection and segmentation	153		Automatic segmentation= CNN-ResNet-101	Manual segmentations	HD = 13 mm F1 = 96.6%	The algorithm saves time in comparison to the manual method. In this algorithm, the segmentation results for upper and lower molars were lower than other teeth.
Vinayahalingam et al. 2019[8]	Third Molars Mandibular Nerve	81		Automatic segmentation= CNN-UNet	Manual segmentations	DSC: Third molar = 0.947 ± 0.033 Mandibular nerve = 0.847 ± 0.099	The algorithm can help in clinical decision-making.
Lee et al. 2019[9]	Tooth segmentation	Total:50 Training:30 Validation:10 Test:10		Automatic segmentation= R-CNN	Manual segmentations	F1 = 0.875 P = 0.858 R = 0.893 IoU = 0.877	The resemblance between the DL-based method and manual segmentation was high.
ACC, Accuracy; ASSD, Average Symmetric Surface Distance; AUC, Area Under Curve; CFD, Computational fluid dynamics; CNN, Convolutional Neural Network; DNN, Deep Neural Network; DSC, Dice similarity coefficient; F, F Measure; F1, F1-Score; FOV, Field of View; FP, False Positive; HD, Hausdorff distance; IOU, Intersection Over Union; P, Precision; PQ, Panoptic quality; PR, Precision Rate; R, Recall; RQ, recognition quality; RR, Recall Rate; SE, Sensitivity; SP, Specificity; SQ, Segmentation Quality; SVM, Support Vector Machine;
1. Mori M, Ariji Y, Fukuda M, Kitano T, Funakoshi T, Nishiyama W, et al. Performance of deep learning technology for evaluation of positioning quality in periapical radiography of the maxillary canine. Oral Radiol. 2022;38(1):147 − 54. doi: 10.1007/s11282-021-00538-2.
2. Li S, Fevens T, Krzyzak A, Li S. An automatic variational level set segmentation framework for computer aided dental X-rays analysis in clinical environments. Comput Med Imaging Graph. 2006;30(2):65–74. doi: 10.1016/j.compmedimag.2005.10.007.
3. Li S, Fevens T, Krzyzak A, Jin C, Li S. Toward automatic computer aided dental X-ray analysis using level set method. Med Image Comput Comput Assist Interv. 2005;8(Pt 1):670-8. doi: 10.1007/11566465_83.
4. Cha JY, Yoon HI, Yeo IS, Huh KH, Han JS. Panoptic Segmentation on Panoramic Radiographs: Deep Learning-Based Segmentation of Various Structures Including Maxillary Sinus and Mandibular Canal. J Clin Med. 2021;10(12). doi: 10.3390/jcm10122577.
5. Abdi AH, Kasaei S, Mehdizadeh M. Automatic segmentation of mandible in panoramic x-ray. J Med Imaging (Bellingham). 2015;2(4):044003. doi: 10.1117/1.Jmi.2.4.044003.
6. Vranckx M, Van Gerven A, Willems H, Vandemeulebroucke A, Ferreira Leite A, Politis C, Jacobs R. Artificial Intelligence (AI)-Driven Molar Angulation Measurements to Predict Third Molar Eruption on Panoramic Radiographs. Int J Environ Res Public Health. 2020;17(10). doi: 10.3390/ijerph17103716.
7. Leite AF, Gerven AV, Willems H, Beznik T, Lahoud P, Gaêta-Araujo H, et al. Artificial intelligence-driven novel tool for tooth detection and segmentation on panoramic radiographs. Clin Oral Investig. 2021;25(4):2257-67. doi: 10.1007/s00784-020-03544-6.
8. Vinayahalingam S, Xi T, Bergé S, Maal T, de Jong G. Automated detection of third molars and mandibular nerve by deep learning. Sci Rep. 2019;9(1):9007. doi: 10.1038/s41598-019-45487-3.
9. Lee JH, Han SS, Kim YH, Lee C, Kim I. Application of a fully deep convolutional neural network to the automation of tooth segmentation on panoramic radiographs. Oral Surg Oral Med Oral Pathol Oral Radiol. 2020;129(6):635 − 42. doi: 10.1016/j.oooo.2019.11.007.

Among the studies in Table 3, 25 used CBCT, 11 used MDCT, 4 used Micro-CT, 2 used ultrasonography, and 3 used MRI images. Of the studies included in Table 2, 3 used periapical and 6 used panoramic images. The publishing year of the included studies was between 2005 and 2022.

Regarding the segmented region, 18 studies segmented the tooth and pulp chamber and cavity, 12 segmented the jaws (maxilla and mandible), 9 segmented the craniomaxillofacial structures, 4 segmented the lesions, 3 segmented the Airway, 2 segmented the masseter muscle, 1 segmented the TMJ and 1 segmented the parotid gland. additionally, 4 studies had other subjects that did not fit into the mentioned categories.

Regarding the segmentation protocol and method, the majority of studies used convolutional neural networks (CNNs) and artificial neural networks (ANNs) mostly including U-net and MS-D net. (Fig. 2)

The reference tests used by the studies were manual segmentation, semi-segmentation or refined AI segmentation, and other AI-based automatic segmentation methods.

Moreover, regarding the metric findings, studies reported parameters including Precision, Recall, F1-score, F measure, The reliability of correctly detecting, Dice similarity coefficient, Intersection over the union, Hausdorff distance, Average time, Jaccard index, Relative error, Accuracy, Distance from the hand-drawn reference, Correlation coefficient, The Pearson’s correlation coefficient, Average symmetric surface distance (ASSD), Overall success rate, Sensitivity, Specificity, False Positive (FP), Tree length detected, Segmentation error, Positive predictive value, Negative predictive value, Area under the curve, Panoptic quality, Recognition quality, Segmentation quality.

Risk of bias assessment

In all the included studies, the accuracy of segmentation methods was the main focus. The reference standard in 81% of the studies was reported to have a low risk of bias in the current analysis. The AI technology used in the final output was highly standardized, with no impact on the flow or time frame, and was therefore classified as a low-risk category. The index test was reported to have a low risk of bias in 88% of cases in the current systematic review (supplementary materials)

This article aims to have an outlook on the efficacy and accuracy of different segmentation methods in dentomaxillofacial radiology.

Jaw Related Structures

Making virtual 3D models of pertinent anatomic regions of interest, such as the mandible, maxilla, and teeth, from CBCT scans is a crucial first step in the identification of dentofacial anomalies and deformities. 3D representations of a patient at various times can be superimposed to visually and quantitatively analyze orthodontic changes.(Cha et al. 2021) Segmenting craniomaxillofacial and jaw-related structures including the mandible, maxilla, maxillary sinus, mandibular canal, condyle and related structures, and alveolar sockets is a labor-intensive, time-consuming, and highly operator-dependent task. (Wang et al. 2021) Thus there is growing interest in the application of semi-automatic and fully automatic segmentation models for different structures in CBCT, Panoramic, and MDCT imaging. Although panoramic images present a general field of view with a low cost, this radiography technique has limitations in representing 3D structures. (Ghaeminia et al. 2009) CBCT images on the other hand provide full construction of 3D structures with lower dose and cost compared to MDCT. Therefore, CBCT imaging is the major modality used for the segmentation of jaws and associated anatomies.(Janssen et al. 2017)

3D models of craniomaxillofacial structures are crucial for diagnosis, treatment planning, and patient communication, particularly in the field of computer-assisted surgery.(Verhelst et al. 2021) According to our findings, the employment of automatic segmentation techniques with AI-based models in the mandible has shown great promise.

• Mandible

One article proposed a segmentation model for mandible in panoramic radiographs (Abdi et al. 2015) with comparable results to manual segmentation. Another study (Rueda et al. 2006) performed segmentation by an active appearance model on CT images. 2 studies (Verhelst et al. 2021) (Antila et al. 2008) evaluated mandible automatic segmentation in CBCT radiographs and reached a Dice similarity coefficient higher than 0.91. The inability of the U-Net AI technique to differentiate between cortical and medullary bone is one of its limitations. Another study (Wang et al. 2021) proposed a mixed-scale dense (MS-D) CNN that can segment both jaws and teeth simultaneously on CBCT images. This segmentation method was compared to the Binary segmentation (which can either segment the jaw bone or the teeth) and showed advantages such as not generating conflicting labels. The MS-D network showcased impressive performance in segmenting the jaw and teeth, exhibiting the Dice Similarity Coefficients (DSC) of 0.93 for the jaw and 0.95 for the teeth. The edges of the bony structures were where the majority of the segmentation mistakes of the model occurred.

• Maxilla and Maxillary sinus

One of the AI models (Cha et al. 2021) was developed based on Panoptic segmentation which involves a fusion of both semantic segmentation and instance segmentation. This combination makes the segmentation of both countable objects and uncountable regions possible. It was used to segment various structures in dental panoramic radiographs including the maxillary sinus, maxilla, mandible, mandibular canal, normal teeth, treated teeth, and dental implants. Based on the results, the mandibular canal segmentation exhibited the lowest panoptic quality (PQ) and segmentation quality (SQ) scores, aligning with the Intersection over Union (IoU) results. Overall result indicated that this segmentation model was not good enough for teeth segmentation.

• Condyle

Condyles are frequently difficult structures to segment because of their complex and varied shape, poor bone density, and prominently shaded and superimposition-filled glenoid fossa. One article (Xi et al. 2014) proposed a semi-automatic segmentation model for condylar regions.

• Bone

One of the included studies (Minnema et al. 2019) compared AI models including MS-D Net, U-Net, Res-Net, and the snake evolution algorithm (a model-driven segmentation method) in bone segmentation in CBCT scans affected by metal artifacts interactively. Findings suggested that the DSCs of the models were 0.87, 0.87, 0.86, and 0.78 respectively. The CNNs employed in this study demonstrated superior performance compared to the current clinical benchmark (the snake evolution algorithm) in segmenting bony structures and accurately classifying metal artifacts as background in CBCT scans which is attributed to the CNNs' capacity to learn distinctive features that differentiate between bone and metal artifacts. The results suggested that the MS-D network architecture can be applied across a diverse array of applications.

• Mandibular canal

Precise knowledge of the mandibular canal's exact location is crucial for planning appropriate oral-maxillofacial surgeries, including procedures like implant placement and third molar extractions. (Jung and Cho 2014) Anatomical variations in the course, size, and shape of the mandibular canal render it difficult to accurately detect particularly in CBCT due to low contrast resolution.(Valenzuela-Fuenzalida et al. 2021) The segmentation model designed by Jeoun et al.(Jeoun et al. 2022) was Canal-Net, which is a multi-task learning framework with bases of 3D U-Net. The model successfully acquired knowledge of local anatomical variations of the canal by integrating spatiotemporal features besides capturing the global structural continuity information of the canal which resulted in a 0.87 DSC. Also, the comparison of 2D U-Net, SegNet, 3D U-Net, MPL 3D U-Net, and ConvLSTM 3D U-Net was conducted. The lower accuracies in segmentation of mandibular canal were achieved by 2D networks (2D U-Net and SegNet) and the higher accuracies were achieved by the MPL 3D U-Net and ConvLSTM 3D U-Net. The MPL 3D U-Net faced challenges in delineating boundaries, particularly around the mental foramen area. In contrast, the ConvLSTM, by learning anatomical context through spatiotemporal features, achieved smoother boundaries and consistent accuracies in the canal volume. Consequently, the Canal-Net demonstrated the most accurate segmentation of the entire MC volume, learning global structural continuity through MPL and anatomical context through ConvLSTM. Another study (Kwak et al. 2020) also compared 2D and 3D networks in automatic segmentation of the mandibular canal in CBCT images. The models were based on 2D SegNet, 2D and 3D U-Nets. The results indicated that the 2D U-Net had a global accuracy of 0.82, the 2D SegNet achieved a global accuracy of 0.96, and the 3D U-Net outperforms all with the highest global accuracy of 0.99. However, there was a limitation with 3D networks in situations where the cortical layer around the canal was unclear.

Lesions

Segmentation and detection of lesions is a challenging task that can be done using periapical, panoramic, and CBCT imaging. (Kruse et al. 2015)

• Periapical lesions

Periapical diseases are mostly inflammatory lesions that can be seen in CBCT radiographs with high 3D resolution, and better visible canal spaces, without the distortion and superimposition of neighboring structures compared to conventional radiographs. (Davies et al. 2015) Diagnosis of these lesions is challenging due to the variety of diseases with the same symptoms and not having radiographic signs. (Patel et al. 2009) In the study of Orhan et al.(Orhan et al. 2020) a deep learning process with a U-Net-based algorithm measured the volumetric features quite similar to the manual method, however, it did not outperform it. The calculated volume slices by the manual method were 191.41 and by the automatic method was 143.84. Neighboring anatomical structures like the maxillary incisive canal, inferior alveolar canal, mental foramen, maxillary sinus, and nasal fossa may impact AI analysis. In another study (Zheng et al. 2021) the constrained U-Net algorithm outperformed the data-driven method.

• Maxillary sinus lesions

Precise three-dimensional segmentation of the maxillary sinus is vital for various diagnostic and treatment purposes including evaluating sinus changes and lesions, monitoring remodeling over time, conducting volumetric analysis, and generating 3D virtual models.(Janner et al. 2020) (Al Abduwani et al. 2016)

Jung et al.(Jung et al. 2021) compared a 3D nnU-Net algorithm for segmentation of the maxillary sinus into the maxillary bone, air, and lesion performance to expert manual segmentation. Besides being faster than manual segmentation, the model achieved DSCs of 0.93 for air and 0.76 for lesion, respectively. low segmentation performance when the sinus is filled with inflammatory material and severe maxillary sinusitis were limitations of the model. Another study (Hung et al. 2022) proposed an algorithm based on V-Net and Support vector regression to segment mucosal thickening and mucosal retention cysts in both low-dose and full-dose CBCT images. The result was DSCs of 0.66–0.73 for mucosal thickening and 0.68–0.79 for mucosal retention cysts. One of the limitations of this study is that the performance of the model in segmenting unilateral or partial coverage lesions was not evaluated.

Craniofacial structures segmentation:

• Airway

Airway analysis is done based on 3D volumetric images such as MDCT and CBCT. A part of the quantitative airway assessment is the segmentation of the structure.(Weissheimer et al. 2012) This is useful for the diagnosis and treatment planning of pulmonary diseases, assessment of obstructive sleep apnea patients, the prediction of airway changes after orthognathic surgery, and for orthodontic and growth modification treatments.(AlQahtani et al. 2021) However, manual and semi-automatic segmentation techniques are time-consuming.(Alsufyani et al. 2012) Five articles included in this review have developed deep learning-based models for automatic segmentation of the airways. Three of them used CBCT and two used MDCT.

One study (Park et al. 2021) designed a regression neural network-based deep-learning model to conduct airway volume measurements based on CBCT scans. Using the reference plane, the deep learning model of this study splits the airway into the nasopharynx, oropharynx, and hypopharynx segments entirely automatically. The measured volume differences were 48.620 mm3, 37.987 mm3, and 50.010 mm3 in the nasopharynx, oropharynx, and hypopharynx, respectively. Another study (Sin et al. 2021) compared a U-Net-based to human method by ITK-SNAP software. The average volume of the pharyngeal airway was found to be 18.08 cm3 by the human observer and 17.32 cm3 by the artificial intelligence. This study concluded that the developed models performed equally as well as the expert at a reduced time.

• Vocal tracts

Ruthven et al. (Ruthven et al. 2021) developed a U-Net algorithm for the segmentation of vocal tracts and multiple groups of articulators on 2D MR images. The automated model performed segmentation in 6 classes including head, soft palate, jaw, tongue, vocal tract, and tooth space, and reached a 0.92 median Dice coefficient and a 5mm median general Hausdorff distance. The most accurate segmentation belonged to the head class and the least belonged to the soft palate and tooth classes. These results are due to the largest and the smallest number of pixels respectively. The limitation mentioned in this study is that occasionally the method struggled to maintain the integrity of small gaps between the soft palate and the pharyngeal wall.

• Masseter

Masseter is One of the muscles of mastication most affected by Bruxism.(Jiménez-Silva et al. 2017) Ultrasonography (US) imaging enables accurate and exact measurements of the muscle thickness/ width and can identify the changes in the muscle.(Blicharz et al. 2021) Two deep learning-based algorithms were developed and are included in this review.

One study (Orhan et al. 2021) created a D-CNN model, utilizing the U-Net, Pyramid Scene Parsing Network (PSPNet), and Fuzzy Petri Net (FPN) architectures in ultrasonography images of individuals experiencing bruxism. The FPN model achieved an accuracy of 0.985, while the PSPNet demonstrated performance of 0.947, and the U-Net exhibited an accuracy of 0.969.

• Others

Other articles worked on models for segmenting structures including Bone/Vertebrae/Vessels which were developed based on MDCT images. 2 others presented algorithms for segmenting the articular disc of the temporomandibular joint and parotid gland on MRI images.

One (Steybe et al. 2022) developed a multiscale stack of 3D convolutional neural networks based on U-Net architectures for segmentation of head CT structures including: Bones (Viscerocranium/skullbase, Nasal septum, and Mandible), Paranasal sinuses (Frontal sinus, Sphenoid sinus, and Maxillary sinus), Canals (Nasolacrimal duct, Carotid canal, and Jugular foramen), Foramina, and Soft tissue (Ocular globe, Extraocular muscles, and Optic nerve). The mean DSC of the structures achieved by the presented model was 0.81 (Lowest: 0.61 for mental foramen, and Highest: 0.98 for mandible), and the mean Surface DSC was 0.94 (Lowest: 0.87 for mental foramen, and Highest: 0.99 for mandible). Of limitations mentioned in this study is that modified anatomical structures, such as those brought on by tumors, other pathologies, and trauma injuries, would compromise the existing model's segmentation accuracy.

Treatment for head and neck cancer often involves radiation therapy. One of the difficult things about planning radiation therapy is figuring out the precise target volume and the organs that are at risk nearby. Among its many drawbacks are the extreme labor intensity, length of time required, and reliance on radiation oncologists' anatomical expertise for the manual identification of these organs. (Lim and Leech 2016) The automatic model developed by one of the studies (Zhong et al. 2021) is a U-net-based full CNN for the segmentation of organs at risk for head and neck cancer radiotherapy on CT images. All other DSCs generated in this study, except for the optic nerve and chiasm, were bigger than 0.7. Direct clinical consequences of radiotherapy arise from the delineation of the target volume and surrounding organs at risk. The dosimetric impact of the automated segmentation results should be taken into consideration when evaluating them. This is one of the limitations of this study.

One of the studies (Ito et al. 2022) compared the TMJ articular disk segmentation abilities of 3 AI methods including 3DiscNet, U-Net, and SegNet-basic. The DSCs were 0.70, 0.46, and 0.74 respectively. The limitations mentioned in the study were using images from a single institution and training the algorithm by the 'ground truth' manual segmentation images created by a limited number of experts.

Tooth and pulp cavity segmentation

• Pulp cavity segmentation

Recently several models have been developed for pulp segmentation on CBCT and micro-CT radiographs. Zheng et al.(Zheng et al. 2021) and Penaloza et al.(Marroquin Penaloza et al. 2016) developed algorithms for pulp segmentation on CBCT images. With the deposition of secondary dentin over time, the size of the pulp chamber decreases, so this can be used for age estimation based on the pulp chamber volume. Zheng et al. concluded that the model can successfully perform age estimation; however, the algorithm was developed based on the first molars but it is applicable for single and multi-radicular teeth. Penaloza et al. showed that applying the same setting parameters to all teeth for automatic and manual segmentation is impossible. Furthermore, they concluded that manual segmentation is time-consuming. Four other articles developed models using CBCT and micro-CT from extracted teeth. Penaloza et al. used micro-CT of the extracted teeth from which CBCT had been obtained as the ground truth. This article showed promising results and hope for future use in endodontic diagnosis and therapy.

• Tooth segmentation

Tooth identification is quite important in accurate diagnosis, treatment planning, and better clinical decision-making. Manual annotation is a time-consuming process since there is low contrast between cementum, dentin, and bone.13 articles in this review worked on tooth segmentation. 4 used panoramic, 6 used CBCT and 3 used MDCT for algorithm development. During recent years deep learning-based algorithms for segmentation have been proposed.(Minaee et al. 2022) Almost all models reported improved accuracy and a decrease in time for segmentation.

Regarding the assessment of quality in each study included, the most important focus was on the accuracy of segmentation methods. The reference standard in 81% of the studies was reported to have a low risk of bias in the analysis. The AI technology applied for the ultimate output was very standardized, with no impact on flow or time and it has been considered to be in a category of lower risk. The index test was reported to have a low risk of bias in 88% of cases in the current systematic review.

This study has a small number of limitations, although we used a detailed methodology, it is possible that some articles that may have been included were missed, due to the wide range of segmentation methods used in dental imaging. In addition, this systematic review is an all-inclusive analysis of different segmentation methods, thus complete evaluation and in-depth assessment of each of the segmentation methods were not performed. Therefore, further studies that evaluate certain subjects and different clinical applications and their effectiveness are required.

In Summary, AI-driven technologies hold great promise providing a fully automated capability for lesion segmentation, which will result in reduced subjectivity and errors. Automatic AI-driven segmentation methods enable potential future applications in the digital dental diagnostic and treatment planning process while reducing clinical workload. Based on the results of our included studies, automatic AI-based segmentation methods could be a helpful clinical tool able to segment lesions as well as manual methods. Moreover, CBCT imaging is the major modality used for the segmentation of jaws and associated anatomies. However, Continued efforts will be necessary to improve the algorithms and Future work should concentrate on providing larger datasets including various pathologies and structures.

Statements and Declarations: The authors have no conflicts of interest to declare that are relevant to the content of this article.

Financial interests: The authors declare they have no financial interests.

Non-financial interests: none.

Author contributions: M.G.A., M.A., and S.H. had the idea for the article, M.H. and S.H. performed the literature search and data analysis, and drafted the work. G.K. performed the risk of bias evaluation. M.M. and M.G.A. critically revised the work.

Acknowledgments:

All authors gave their final approval and agreed to be accountable for all aspects of the work.

The study was self-funded by the authors and their institution. The authors do not have any financial interest to declare.

Abdi AH, Kasaei S, Mehdizadeh M (2015) Automatic segmentation of mandible in panoramic x-ray. J Med Imaging (Bellingham) 2(4):044003
Al Abduwani J, ZilinSkiene L, Colley S, Ahmed S (2016) Cone beam ct paranasal sinuses versus standard multidetector and low dose multidetector ct studies. Am J Otolaryngol 37(1):59–64
AlQahtani F, George JM, Bishawi K, Kuriadom ST (2021) Comparison of oropharyngeal airway dimensional changes in patients with skeletal class ii and class iii malocclusions after orthognathic surgery and functional appliance treatment: A systematic review. Saudi Dent J 33(8):860–868
Alsufyani NA, Flores-Mir C, Major PW (2012) Three-dimensional segmentation of the upper airway using cone beam ct: A systematic review. Dentomaxillofac Radiol 41(4):276–284
Antila K, Lilja M, Kalke M, Lötjönen J (2008) Automatic extraction of mandibular bone geometry for anatomy-based synthetization of radiographs. Annu Int Conf IEEE Eng Med Biol Soc. 2008:490–493
Blicharz G, Rymarczyk M, Rogulski M, Linek P (2021) Methods of masseter and temporal muscle thickness and elasticity measurements by ultrasound imaging: A literature review. Curr Med Imaging 17(6):707–713
Cha JY, Yoon HI, Yeo IS, Huh KH, Han JS (2021) Panoptic segmentation on panoramic radiographs: Deep learning-based segmentation of various structures including maxillary sinus and mandibular canal. J Clin Med 10(12)
Chilali O, Puech P, Lakroum S, Diaf M, Mordon S, Betrouni N (2016) Gland and zonal segmentation of prostate on t2w mr images. J Digit Imaging 29(6):730–736
Cover GS, Herrera WG, Bento MP, Appenzeller S, Rittner L (2018) Computational methods for corpus callosum segmentation on mri: A systematic literature review. Comput Methods Programs Biomed 154:25–35
Davies A, Mannocci F, Mitchell P, Andiappan M, Patel S (2015) The detection of periapical pathoses in root filled teeth using single and parallax periapical radiographs versus cone beam computed tomography - a clinical study. Int Endod J 48(6):582–592
Gaviño Orduña JF, García García M, Dominguez P, Caviedes Bucheli J, Martin Biedma B, Abella Sans F, Manzanares Céspedes MC (2020) Successful pulp revascularization of an autotransplantated mature premolar with fragile fracture apicoectomy and plasma rich in growth factors: A 3-year follow-up. Int Endod J 53(3):421–433
Ghaeminia H, Meijer GJ, Soehardi A, Borstlap WA, Mulder J, Bergé SJ (2009) Position of the impacted third molar in relation to the mandibular canal. Diagnostic accuracy of cone beam computed tomography compared with panoramic radiography. Int J Oral Maxillofac Surg 38(9):964–971
Gillot M, Baquero B, Le C, Deleat-Besson R, Bianchi J, Ruellas A, Gurgel M, Yatabe M, Al Turkestani N, Najarian K et al (2022) Automatic multi-anatomical skull structure segmentation of cone-beam computed tomography scans using 3d unetr. PLoS ONE 17(10):e0275033
Heye T, Merkle EM, Reiner CS, Davenport MS, Horvath JJ, Feuerlein S, Breault SR, Gall P, Bashir MR, Dale BM et al (2013) Reproducibility of dynamic contrast-enhanced mr imaging. Part ii. Comparison of intra- and interobserver variability with manual region of interest placement versus semiautomatic lesion segmentation and histogram analysis. Radiology 266(3):812–821
Hung KF, Ai QYH, King AD, Bornstein MM, Wong LM, Leung YY (2022) Automatic detection and segmentation of morphological changes of the maxillary sinus mucosa on cone-beam computed tomography images using a three-dimensional convolutional neural network. Clin Oral Investig 26(5):3987–3998
Ito S, Mine Y, Yoshimi Y, Takeda S, Tanaka A, Onishi A, Peng TY, Nakamoto T, Nagasaki T, Kakimoto N et al (2022) Automated segmentation of articular disc of the temporomandibular joint on magnetic resonance images using deep learning. Sci Rep 12(1):221
Janner SFM, Dubach P, Suter VGA, Caversaccio MD, Buser D, Bornstein MM (2020) Sinus floor elevation or referral for further diagnosis and therapy: A comparison of maxillary sinus assessment by ent specialists and dentists using cone beam computed tomography. Clin Oral Implants Res 31(5):463–475
Janssen NG, Schreurs R, Bittermann GKP, Borstlap WA, Koole R, Meijer GJ, Maal TJJ (2017) A novel semi-automatic segmentation protocol for volumetric assessment of alveolar cleft grafting procedures. J Craniomaxillofac Surg 45(5):685–689
Jeoun BS, Yang S, Lee SJ, Kim TI, Kim JM, Kim JE, Huh KH, Lee SS, Heo MS, Yi WJ (2022) Canal-net for automatic and robust 3d segmentation of mandibular canals in cbct images using a continuity-aware contextual network. Sci Rep 12(1):13460
Jiménez-Silva A, Peña-Durán C, Tobar-Reyes J, Frugone-Zambra R (2017) Sleep and awake bruxism in adults and its relationship with temporomandibular disorders: A systematic review from 2003 to 2014. Acta Odontol Scand 75(1):36–58
Jung SK, Lim HK, Lee S, Cho Y, Song IS (2021) Deep active learning for automatic segmentation of maxillary sinus lesions using a convolutional neural network. Diagnostics (Basel). 11(4)
Jung Y-H, Cho B-H (2014) Radiographic evaluation of the course and visibility of the mandibular canal. Imaging Sci Dent 44(4):273–278
Kruse C, Spin-Neto R, Wenzel A, Kirkevang LL (2015) Cone beam computed tomography and periapical lesions: A systematic review analysing studies on diagnostic efficacy by a hierarchical model. Int Endod J 48(9):815–828
Kwak GH, Kwak EJ, Song JM, Park HR, Jung YH, Cho BH, Hui P, Hwang JJ (2020) Automatic mandibular canal detection using a deep convolutional neural network. Sci Rep 10(1):5711
Lim JY, Leech M (2016) Use of auto-segmentation in the delineation of target volumes and organs at risk in head and neck. Acta Oncol 55(7):799–806
Marroquin Penaloza TY, Karkhanis S, Kvaal SI, Vasudavan S, Castelblanco E, Kruger E, Tennant M (2016) Reliability and repeatability of pulp volume reconstruction through three different volume calculations. J Forensic Odontostomatol 34(2):35–46
Minaee S, Boykov Y, Porikli F, Plaza A, Kehtarnavaz N, Terzopoulos D (2022) Image segmentation using deep learning: A survey. IEEE Trans Pattern Anal Mach Intell 44(7):3523–3542
Minnema J, van Eijnatten M, Hendriksen AA, Liberton N, Pelt DM, Batenburg KJ, Forouzanfar T, Wolff J (2019) Segmentation of dental cone-beam ct scans affected by metal artifacts using a mixed-scale dense convolutional neural network. Med Phys 46(11):5027–5035
Nazem-Zadeh MR, Saksena S, Babajani-Fermi A, Jiang Q, Soltanian-Zadeh H, Rosenblum M, Mikkelsen T, Jain R (2012) Segmentation of corpus callosum using diffusion tensor imaging: Validation in patients with glioblastoma. BMC Med Imaging 12:10
Olabarriaga SD, Smeulders AWM (2001) Interaction in the segmentation of medical images: A survey. Med Image Anal 5(2):127–142
Orhan K, Bayrakdar IS, Ezhov M, Kravtsov A, Özyürek T (2020) Evaluation of artificial intelligence for detecting periapical pathosis on cone-beam computed tomography scans. Int Endod J 53(5):680–689
Orhan K, Yazici G, Kolsuz ME, Kafa N, Bayrakdar IS, Çelik Ö (2021) An artificial intelligence hypothetical approach for masseter muscle segmentation on ultrasonography in patients with bruxism. J Adv Oral Res 12(2):206–213
Ouzzani M, Hammady H, Fedorowicz Z, Elmagarmid A (2016) Rayyan—a web and mobile app for systematic reviews. Syst Reviews 5(1):210
Page MJ, Moher D, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, Shamseer L, Tetzlaff JM, Akl EA, Brennan SE et al (2021) Prisma 2020 explanation and elaboration: Updated guidance and exemplars for reporting systematic reviews. BMJ 372:n160
Park J, Hwang J, Ryu J, Nam I, Kim S-A, Cho B-H, Shin S-H, Lee J-Y (2021) Deep learning based airway segmentation using key point prediction. Appl Sci 11(8):3501
Parmar C, Rios Velazquez E, Leijenaar R, Jermoumi M, Carvalho S, Mak RH, Mitra S, Shankar BU, Kikinis R, Haibe-Kains B et al (2014) Robust radiomics feature quantification using semiautomatic volumetric segmentation. PLoS ONE 9(7):e102107
Patel S, Dawood A, Whaites E, Pitt Ford T (2009) New dimensions in endodontic imaging: Part 1. Conventional and alternative radiographic systems. Int Endod J 42(6):447–462
Pérez de Alejo R, Ruiz-Cabello J, Cortijo M, Rodriguez I, Echave I, Regadera J, Arrazola J, Avilés P, Barreiro P, Gargallo D, Graña M (2003) Computer-assisted enhanced volumetric segmentation magnetic resonance imaging data using a mixture of artificial neural networks. Magn Reson Imaging 21(8):901–912
Navab N, Hornegger J, Wells WM, Frangi AF (eds) U-net: Convolutional networks for biomedical image segmentation. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015; 2015// 2015; Cham. Springer International Publishing
Rueda S, Gil JA, Pichery R, Alcañiz M (2006) Automatic segmentation of jaw tissues in ct using active appearance models and semi-automatic landmarking. Med Image Comput Comput Assist Interv 9(Pt 1):167–174
Ruthven M, Miquel ME, King AP (2021) Deep-learning-based segmentation of the vocal tract and articulators in real-time magnetic resonance images of speech. Comput Methods Programs Biomed 198:105814
Sin Ç, Akkaya N, Aksoy S, Orhan K, Öz U (2021) A deep learning algorithm proposal to automatic pharyngeal airway detection and segmentation on cbct images. Orthod Craniofac Res 24 Suppl 2:117–123
Steybe D, Poxleitner P, Metzger MC, Brandenburg LS, Schmelzeisen R, Bamberg F, Tran PH, Kellner E, Reisert M, Russe MF (2022) Automated segmentation of head ct scans for computer-assisted craniomaxillofacial surgery applying a hierarchical patch-based stack of convolutional neural networks. Int J Comput Assist Radiol Surg 17(11):2093–2101
Valenzuela-Fuenzalida JJ, Cariseo C, Gold M, Díaz D, Orellana M, Iwanaga J (2021) Anatomical variations of the mandibular canal and their clinical implications in dental practice: A literature review. Surg Radiol Anat 43(8):1259–1272
Verhelst PJ, Smolders A, Beznik T, Meewis J, Vandemeulebroucke A, Shaheen E, Van Gerven A, Willems H, Politis C, Jacobs R (2021) Layered deep learning for automatic mandibular segmentation in cone-beam computed tomography. J Dent 114:103786
Wang H, Minnema J, Batenburg KJ, Forouzanfar T, Hu FJ, Wu G (2021) Multiclass cbct image segmentation for orthodontics with deep learning. J Dent Res 100(9):943–949
Weissheimer A, Menezes LMd, Sameshima GT, Enciso R, Pham J, Grauer D (2012) Imaging software accuracy for 3-dimensional analysis of the upper airway. Am J Orthod Dentofac Orthop 142(6):801–813
Whiting PF, Rutjes AW, Westwood ME, Mallett S, Deeks JJ, Reitsma JB, Leeflang MM, Sterne JA, Bossuyt PM (2011) Quadas-2: A revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med 155(8):529–536
Xi T, Schreurs R, Heerink WJ, Bergé SJ, Maal TJ (2014) A novel region-growing based semi-automatic segmentation protocol for three-dimensional condylar reconstruction using cone beam computed tomography (cbct). PLoS ONE 9(11):e111126
Zheng Z, Yan H, Setzer FC, Shi KJ, Mupparapu M, Li J (2021) Anatomically constrained deep learning for automating dental cbct segmentation and lesion detection. IEEE Trans Autom Sci Eng 18(2):603–614
Zhong Y, Yang Y, Fang Y, Wang J, Hu W (2021) A preliminary experience of implementing deep-learning based auto-segmentation in head and neck cancer: A study on real-world clinical cases. Front Oncol 11:638197

Table 3 is available in the Supplementary Files section.

No competing interests reported.

Assessment of Efficacy and Accuracy of Segmentation Methods in Dentomaxillofacial Imaging- A Systematic Review

Status:

Version 1

Abstract

Figures

Introduction

Methods and Materials

Results