Introduction
Tooth segmentation plays a fundamental role in digital diagnosis and design. The interproximal surface, a key area for accurate tooth segmentation, remains poorly captured due to intraoral scanner (IOS) limitations. The objective of this study was to evaluate the 3-dimensional accuracy of the interproximal surface in a dentition model segmented using the following: (1) IOS alone; (2) artificial intelligence (AI)–based IOS and cone-beam computed tomography (CBCT) fusion; and (3) IOS-CBCT fusion followed by manual interproximal adjustment.
Methods
Fourteen extracted maxillary teeth were scanned by IOS to be used as a reference, then arranged in a full-arch configuration with moderate crowding. IOS and CBCT segmentation of the model and subsequent fusion of segmented crowns and roots were performed on the Relu AI-based platform (March 2022 version; Relu BV, Leuven, Belgium). The fused models were then manually adjusted at the interproximal surface, as seen from the axial CBCT slices. Three-dimensional deviation analyses were carried out on the 3 segmented models for the interproximal, outer, and total crown surfaces. Statistical analysis was conducted using repeated-measures analysis of variance, Friedman, Tukey honest significant difference, and Wilcoxon signed rank tests ( P <0.05).
Results
Descriptive statistics revealed a tendency for IOS models to overestimate the interproximal surface compared with fused models. The manually adjusted models demonstrated significantly improved interproximal accuracy ( P <0.05). The total surface accuracy of fused models was significantly improved from IOS models ( P <0.05). Outer surface deviations did not vary.
Conclusions
The preliminary findings suggest that CBCT integration significantly enhances the total surface accuracy of AI-segmented tooth models, whereas localized manual refinement further improves interproximal trueness.
Highlights
-
•
Interproximal surface inaccuracies accounted for the significant differences among the 3 models.
-
•
The fusion method produced more accurate crowns overall than intraoral scanner crowns alone.
-
•
Manual interproximal adjustment significantly improved interproximal accuracy in fused models.
-
•
Cone-beam computed tomography integration was introduced to improve segmentation accuracy for clear aligner applications.
Tooth segmentation, which involves separating individual teeth from the digital dentition model, is used in many digital dental applications, such as orthodontics, prosthetic dentistry, and digital implantology, to ensure predictable treatment planning and facilitate accurate simulation of clinical scenarios. ,, Little attention has been given to the adequate reproduction of the interproximal surface in studies evaluating tooth segmentation techniques, primarily because of the failure of the intraoral scanner (IOS) to capture it. ,,, In evaluating multimodal cone-beam computed tomography–IOS (CBCT-IOS) reconstructions, studies, such as that of Qian et al, have successfully addressed the occlusal problem, whereby occlusal surfaces of tooth crowns in CBCT scans were replaced with those of IOS crowns. However, the interproximal problem remains unsolved. IOS does not provide an accurate picture of the interproximal surface between teeth, whereby, as the interproximal distance decreases, the accuracy of the impression decreases. This often produces a bridging effect, whereby teeth appear as 1 block because areas inaccessible by the scanner are connected based on outer surface data. The effect of this complication on tooth segmentation has been interpreted in the literature as variations in dental arch length and inaccurate mesiodistal widths of segmented teeth.
The emergence of deep learning (DL) has allowed fully automatic instance segmentation using software algorithms that are both accurate and efficient. ,, Im et al studied the accuracy of traditional segmentation methods compared with DL-based models, concluding a comparable accuracy between landmark-based and DL-based methods. With insufficient interproximal information, automatic segmentation relies on the reconstruction of the interproximal surface based on training data. For instance, Kim et al used general adversarial networks as a method of image completion to reconstruct the interdental areas of teeth scanned directly with an IOS. Such missing data have later led to numerous studies evaluating the validity of fully automatic tooth segmentation as opposed to manual, time-consuming methods. Many of these studies point to an overestimation of tooth size from IOS segmentation alone. In particular, Yacout et al suggest that both DL-based software packages and manual correction are unable to reproduce proximal anatomy.
In light of the novel concepts of multimodal fusion and artificial intelligence (AI)–based segmentation, the question arises as to whether multimodal reconstructions can accurately restore the interproximal surface. The primary objective of this study was to evaluate the 3-dimensional (3D) interproximal surface deviation among 3 segmented tooth models: IOS-derived, IOS-CBCT fused, and IOS-CBCT fused-adjusted. Therefore, teeth were segmented using IOS alone, then CBCT roots were automatically fused to IOS crowns to produce fused models, which were then adjusted at the interproximal surfaces using the CBCT slices at that level to obtain the fused-adjusted models. As a secondary objective, the total crown surface deviation is analyzed to determine the impact of interproximal inaccuracy on overall tooth crown reconstructions. Although the International Organization for Standardization (ISO) 5727 standard defines accuracy for a series of repeated measurements as the combination of trueness and precision, the objective of this study compels us to use the definition of accuracy applied in engineering and metrology studies, which would correspond to the ISO 5725 standard for trueness. Therefore, for the objectives covered in this study, accuracy is defined as the closeness of a measured value to the true value.
Material and methods
Fourteen dry teeth were obtained from patients whose treatment plan included the extraction of 1 or more teeth. The anterior maxillary teeth (teeth 14 to 24) were acquired from 1 patient undergoing immediate guided implant therapy. A maxillary model of the same patient was obtained after jaw segmentation using CBCT. The remaining teeth were collected from other patients requiring extractions for orthodontic or strategic reasons. Written consent was obtained for the use of the extracted teeth and the maxillary model.
The selection criteria were permanent teeth with intact enamel, without any visible fractures, developmental defects, caries, or physical damage because of extraction.
All teeth were rinsed with normal saline and stored in prelabeled containers. Freshly extracted teeth were stored in distilled water at 37°C until use. Once model preparation began, the teeth were sterilized in an autoclave set to 121°C, 15 psi for 15 minutes to avoid cross-contamination.
Figure 1 illustrates the steps followed for model preparation. The 14 teeth, held in place with pink modeling wax, were individually scanned with the Medit i700 (Medit, Seoul, South Korea) IOS to obtain reference models that would be considered ground truth for subsequent deviation analyses ( Fig 1 , A ). The apices inserted into the wax were then scanned by turning the teeth over and trimming the wax base from the initial scan.
Model preparation: A, Individual tooth and its corresponding surface scan; B, Scanned teeth setup with moderate anterior crowding; C, Digital preparation of maxillary model with teeth sockets; D, Full digital model of scanned teeth aligned on the virtual maxillary bone model; E, Printed model with the teeth inserted in the sockets; F, Surface scan printed model after insertion of teeth and application of wax layer.
The teeth were aligned according to a predetermined configuration ( Fig 1 , B ), and space analysis was conducted using Blenderfordental version 3.6 software (Blenderfordental Pty, Robina, Queensland, Australia) to verify that moderate anterior crowding did not exceed 4 mm. Using the same software, the segmented maxillary model was hollowed out, offset was applied, and a trabecular pattern was inserted inside the model using a Boolean operation to reproduce interior bone architecture ( Fig 1 , C ). The aligned teeth were digitally placed over the model, and sockets were created to allow physical tooth insertion inside the model ( Fig 1 , D ). A 2 mm bone-to-cementoenamel junction distance was adopted.
The alveolar bone model was then 3D-printed using dental model resin (Voco GmbH, Cuxhaven, Germany). A light body impression material (Imflex; Metabiomed, Korea), simulating trabecular bone in radiopacity, was poured into the printed model and the sockets of the teeth that were immediately inserted after ( Fig 1 , E ). Pink modeling wax was added to simulate the gum and papillae.
The study model was scanned with the Medit i700 (Medit) IOS, following the recommended procedure to simulate the clinical scenario ( Fig 1 , F ).
Using a proper support, made of a styrofoam base and placed at the center of the area in which the patient’s head would be positioned, a CBCT scan of the study model was acquired using the NewTom VGi (NewTom, Verona, Italy) CBCT unit, with the following technical specifications: field of view 120 mm × 120 mm, voxel size 0.15 × 0.15 × 0.15 mm 3, tube voltage 85 kV, tube current 10 mA, and frequency 85 kHz.
For the first technique (technique 1) involving IOS AI-based segmentation, the polygon file format file of the IOS model was uploaded to the Relu Virtual Patient Creator platform (March 2022 version; Relu BV, Leuven, Belgium), in which the teeth were automatically segmented using the software’s algorithm ( Fig 2 , A ).
Segmentation techniques: A, IOS scan ( top ) and subsequent segmentation ( bottom ) using the Relu platform (Relu BV); B, CBCT and IOS segmentation fusion ( top ) with corresponding axial view from the CBCT scan ( bottom ); C, Fused models uploaded to Blue Sky Plan (Blue Sky Bio) ( top ) followed by contour modification ( bottom ) using the move outlines option ( yellow ) to obtain the fused-adjusted models.
For the second technique (technique 2), both IOS and CBCT scans were uploaded to the platform. An AI-automated registration, segmentation, and fusion of the IOS and CBCT scans took place, merging the IOS crown segmentation and CBCT root segmentation into individual fused teeth models ( Fig 2 , B ).
For the third technique (technique 3), the obtained fused models from the Relu platform (Relu BV), aligned to the initially uploaded CBCT and IOS models, were imported into Blue Sky Plan software (version 4.12.13; Blue Sky Bio, Ill). The panoramic curve is drawn across the set of teeth to produce axial slices at the level of contact between adjacent teeth. The outlines of the segmented and fused teeth were then manually adjusted by a single operator, using the move outlines option ( Fig 2 , C ). This process involved selecting each tooth individually and adjusting its segmented crown outline at the level of the interproximal surface as it appears in the axial and tangential CBCT slices. The slice-by-slice adjustment is done 1 tooth at a time by selecting the outline of the crown to be adjusted. Any gap or overlap near the area of contact between adjacent teeth is eliminated by adjusting the outline of the segmented crown. The 3 segmentation techniques are illustrated in Figure 2 .
The present step involved selecting the interproximal surfaces of the tooth to be analyzed, as well as the outer and total surfaces. These regions of interest were compared across the 3 models to evaluate segmentation accuracy.
A standardized method for the selection of the interproximal surface was adopted to exclude the outer crown surface not involved in the interdental contact. The Relu software (Relu BV) uses an AI-based completion algorithm to regenerate surfaces deemed inaccessible by the scanner head. These regions are geometrically defined by the software’s algorithm during segmentation and remain open. These open crowns, which exhibit cut-out interproximal surfaces, are closed off using the software’s AI algorithm, yielding closed crowns. The reference tooth model was uploaded to Autodesk Meshmixer (Autodesk Inc, San Rafael, Calif), and superimposition with the open crowns allowed the selection and separation of the interproximal surfaces. The regions of the crown, excluding the interproximal surfaces, were defined as the outer surfaces. The union of both interproximal and outer surfaces is the total crown surface.
Because the interproximal surface is relatively small compared with the total surface, we anticipated a subsequent step to ensure the stability of root mean square (RMS) values: the separation of the interproximal surface from the total surface. The segmented model of each technique is superimposed over the reference model. Using Blenderfordental (B4D, Blenderfordental 2019), a Boolean operation was used to replicate the chosen interproximal surfaces of the reference models from the previous step on each of the 3 segmented models ( Fig 3 , A – C ). This operation yielded 3 interproximal regions, corresponding to techniques 1, 2, and 3, that correspond identically to the interproximal region initially selected in the reference model ( Fig 3 , D ).
Isolation of mesial and distal interproximal surfaces. Example of the maxillary right canine. After a Boolean operation, the mesial and distal interproximal surfaces ( colored ) were isolated after superimposition of the segmented tooth models from techniques 1, 2, and 3 over the reference model. Mesial (A) , buccal (B) , and distal (C) aspects of the superimposed models are shown with the separated interproximal surfaces. Zoomed-in view (D) of the boxed area in part B illustrating the interproximal overlap of the 3 segmentation techniques (color-coded) relative to the reference model. The mesial and distal interproximal surfaces were subsequently cropped from each of the 4 tooth models to generate identical surfaces for superimposition and 3D deviation analysis between the reference interproximal surface and the interproximal surfaces from each of the 3 techniques.
Only 1 interproximal surface exists for the maxillary right and left second molars, yielding 26 interproximal surfaces, 14 total surfaces, and 14 outer surfaces to be compared with the reference scanned teeth by simple superimposition.
The IOS, fused, and fused-adjusted models, along with the total and outer surfaces of the reference model and the individual interproximal surfaces for each of the mentioned 4 models, were imported into the Medit Link (Medit) software for deviation analysis.
For the comparisons, a reference group and a target group were selected. The reference group corresponded to the given surface from the reference model to be compared, and the target group corresponded to 1 of the 3 segmented tooth models. After reference-target superimposition, a heat map illustrating 3D deviation was generated, and RMS error was collected for statistical analysis of the outer, interproximal, and total surfaces. For descriptive statistics of the interproximal area in particular, the average deviation (avg), plus average (avg +), minus average (avg–), and absolute average deviation (abs avg) were collected. These values indicate how much the volume of the segmented crowns is larger (avg +) or smaller (avg–) than the reference crown models for the same selected interproximal surface, as well as the overall negative or positive deviation (avg). The absolute values of the deviations are used to reflect the magnitude of error regardless of direction (abs avg). This was done for the purpose of determining whether teeth were being overestimated or underestimated in mesiodistal width at the level of the interproximal surface.
As illustrated in Figure 4 , the various tooth models were visualized in 2 ways: 3D deviation analysis, yielding an RMS value for each studied surface and curvature display, yielding a topographical map of the tooth surface. To eliminate possible discrepancies at the level of the crown-root boundary and compensate for the presence of modeling wax in the 2 mm separating the cementoenamel junction from the bone model, an area of 2 mm around the crown was selected for orientation purposes ( Fig 4 , A ). Following the determination of a reference interproximal surface ( Fig 4 , B ), the surface deviation as well as the topographical characteristics of this surface for each of the three techniques, IOS alone ( Fig 4 , C ), fused ( Fig 4 , D ), and fused-adjusted ( Fig 4 , E ), is illustrated.
Surface deviation ( top ) and topographical analysis ( bottom ): A, The reference model is divided into regions of interest for subsequent deviation analysis ( pink , interproximal surface; green , outer surface; yellow , the 2-mm bone-to-cementoenamel junction area) and taken as is for topographical analysis; B, Reference control model showing the isolated interproximal surface for comparison among the 3 segmentation techniques ( top ) and reference topography model ( bottom ); C, IOS crowns showing the least accuracy of interproximal segmentation between the groups (color bars are included for reference). The curvature display demonstrates the lack of depressions and grooves for the interproximal surface after IOS segmentation ( red , embossed region; blue , engraved region; green , zero curvature, meaning a flat surface; arrow , the red boundary between the zones with rich topographical information and the rather flat interproximal surface [ green ]); D, Fused group showing lower deviation values and variable texture at the level of the interproximal area; E, Fused-adjusted model demonstrating lower deviation values ( green ) than the previous 2 models. Similar to the fused models, more topographical information is visible at the level of the interproximal area.
Statistical analysis
The collected data were analyzed using SPSS version 30.0 software (IBM Corp., Armonk, NY). For descriptive statistics, mean ± standard deviation (SD) was used. Data distribution was summarized graphically using box plots. Normality of the RMS distribution was verified by the Shapiro-Wilk test and homogeneity of variances using the Levene test. For parametric tests, repeated-measures analysis of variance (ANOVA), followed by the Tukey honest significant difference tests were applied; otherwise, the Friedman test, followed by the Wilcoxon signed rank test were used. The significance level was set at 5%, P value [0-1] ≤0.05. Pairwise comparisons were conducted between IOS (1) and fused groups (2), between IOS (1) and fused-adjusted groups (3), and between fused (2) and fused-adjusted groups (3).
Results
Using the RMS values for the 3 segmented models, the RMS error distribution by technique was illustrated using box plots ( Fig 5 ) for each of the evaluated surfaces: total surface, interproximal surface, and outer surface. For the interproximal surface, technique 1 presented the highest median RMS, followed by technique 2, then technique 3 ( Fig 5 , A ). For the outer surface, techniques 1, 2, and 3 exhibited a comparable median RMS ( Fig 5 , B ). For the total surface, technique 1 presented the highest median RMS, followed by technique 2, then technique 3 ( Fig 5 , C ).
