Generation of Tooth Replicas by Virtual Segmentation Using

Artificial Intelligence
Ignacio Pedrinaci

Harvard School of Dental Medicine

Anita Nasseri
Harvard School of Dental Medicine
Javier Calatrava
Complutense University of Madrid
Emilio Couso-Queiruga
University of Bern
William V. Giannobile
Harvard School of Dental Medicine
German O. Gallucci
Harvard School of Dental Medicine
Mariano Sanz
Complutense University of Madrid

Research Article

Keywords: Autotransplantation, Computer Aided Manufacturing, Digital Dentistry Artificial Intelligence, Three-dimensional Printing,
Stereolithography.

Posted Date: July 11th, 2024

DOI: https://doi.org/10.21203/rs.3.rs-4576625/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License

Additional Declarations: No competing interests reported.

Page 1/16
Abstract
Objectives: The primary aim of this investigation was to validate a method for generating 3D replicas through virtual
segmentation, utilizing artificial intelligence (AI) or manual-driven methods, assessing accuracy in terms of volumetric and linear
discrepancies. The secondary aims were the assessment of time efficiency with both segmentation methods and the effect of
post-processing 3D replicas.

Methods: Thirty teeth were scanned through Cone Beam Computed Tomography (CBCT), capturing the region of interest from
human subjects. DICOM files underwent segmentation through both AI and manual-driven methods. Replicas were fabricated with
a stereolithography 3D printer. After surface scanning of pre-processed replicas and extracted teeth, STL files were superimposed
to evaluate linear and volumetric differences using the extracted teeth as the reference. Post-processed replicas were scanned to
assess the effect of post-processing on linear and volumetric changes.

Results: AI-driven segmentation resulted in statistically significant mean linear and volumetric differences of -0.709mm and
-4.70%, respectively. Manual segmentation showed no statistically significant differences in mean linear (-0.463mm) and
volumetric (-1.20%) measures. Comparing manual and AI-driven segmentations, showed that AI-driven segmentation displayed
mean linear and volumetric differences of -0.329mm and -2.23%, respectively. Additionally, AI segmentation reduced mean time by
21.8 minutes. When comparing post-processed to pre-processed replicas, there was a volumetric reduction of -4.53% and a mean
linear difference of -0.151mm.

Conclusion: Both segmentation methods achieved acceptable accuracy, with manual segmentation slightly more accurate and AI-
driven segmentation more time-efficient. Continuous improvement in AI offers the potential for increased accuracy, efficiency, and
broader application in the future.

Clinical Significance: Tooth replica generation in the context of tooth autotransplantation therapy may contribute to enhanced
success and survival rates. Accurate CBCT-based virtual segmentation and 3D printing technologies are particularly important in
the fabrication of 3D replicas. Therefore, it is crucial to assess the accuracy of available techniques and alternatives to
demonstrate their reliability and accuracy in the fabrication of tooth replicas.

1. INTRODUCTION
Digital dentistry is a dynamic and rapidly evolving discipline that has revolutionized dentistry. One of its fundamental principles is
the process of acquiring and accurately segmenting three-dimensional (3D) images [1]. Virtual segmentation of DICOM data files
obtained from Cone Beam Computed Tomography (CBCT) scans provides useful diagnostic tools for enhancing the accuracy and
efficiency in dental implant placement, orthodontic planning, disease detection, and computer-aided rapid prototyping (CARP) for
creating 3D tooth replicas [2–6]. These CARP replicas may play an important role in therapeutic interventions such as tooth
autotransplantation [7].

Tooth autotransplantation is a viable surgical-restorative alternative to replace a missing or hopeless tooth by repositioning an
autologous tooth within the same individual [7, 8]. This treatment option is particularly well suited for patients under active
alveolar process growth or those with malocclusions where the orthodontic movement of the transplanted tooth is indicated since
a successfully transplanted tooth generally maintains a vital periodontium [9, 10]. In the conventional autotransplantation
technique, the extracted donor tooth is used to prepare the new recipient site [11]. This method requires multiple fitting attempts
and adjustments, increasing the risk of potential damage to the periodontal ligament of the future transplant [7, 8, 12].
Additionally, these procedures prolong the time that the tooth remains outside the oral environment, risking the intervention’s
success [7].

In the digital autotransplantation technique, a 3D replica of the donor’s tooth is fabricated based on the virtual segmentation of
DICOM files from a previous CBCT scan. The 3D replica is then used as the template to create and adjust the recipient site, thus
avoiding any damage to the donor’s tooth and, at the same time, reducing the time the extracted tooth is out of the socket [8, 12,
13]. The use of 3D replicas has demonstrated increased success and survival rates [7, 14, 15]. Nonetheless, to fabricate 3D
replicas, it is crucial to establish an accurate method for CBCT virtual segmentation, which may be performed either manually or

Page 2/16
using artificial intelligence tools (AI) [1, 16]. Manual segmentation relies on the operator's skill and experience, leading to longer
segmentation time. AI-driven segmentation is fully automatic based on algorithms, offering a time-efficient alternative [1, 17, 18].
However, it is yet unknown whether AI-driven CBCT segmentation provides the same accuracy as manual segmentation.

Therefore, the primary aim of this study was to validate a method for generating 3-D tooth replicas through virtual segmentation,
utilizing either artificial intelligence (AI) or manual-driven methods. This validation was accomplished by assessing the method's
accuracy regarding volumetric and linear discrepancies. As secondary outcomes, the time spent during the tooth segmentation
procedures with either method and the effect of post-processing 3D tooth replicas were assessed. Previous studies have shown a
statistically significant difference (SSD) in linear and volumetric measurements when comparing replicas made with manual
segmentation versus the original teeth [19]. Therefore, the null hypothesis of this investigation will be that there are no differences
in the replica’s linear and volumetric measurements between AI and manual-driven methods.

2. MATERIALS AND METHODS

2.1. Experimental design, setting, and timeframe.
This clinical investigation is in compliance with the Checking for Reporting In-Vitro studies (CRIS) [20]. All the therapeutic and
follow-up interventions in this study were carried out in the Specialization Postgraduate Implant Clinic of the Harvard School of
Dental Medicine between September 2021 and December 2023.

2.2. Ethical approval, patient recruitment, and extraction protocol

After obtaining approval from the Institutional Review Board (IRB21-1687), selected patients fulfilling the defined inclusion criteria
were informed on the purpose of the study on their extracted teeth by one of the researchers (I.P., A.N.). Eligible patients included
in this study signed the informed consent.

The inclusion criteria were: 1) The patient’s ability to sign an informed consent form for enrollment in the study and 2) Any tooth
suitable for extraction with a pre-operative CBCT scan taken no more than 60 days before the intervention. The exclusion criteria
included teeth with carious lesions, cracked or fractured, presence of fixed dental prosthesis, endodontic treatment, or any
restorative material that could cause scattering or might interfere with the CBCT virtual segmentation procedure.

All extractions were conducted as minimally traumatic as possible to avoid any damage to the tooth or adjacent anatomical
structures. Following extractions, teeth were labeled and gently cleansed with water to remove any attached soft tissue and
subsequently immersed for 30 min in a 1:10 solution of bleach for decontamination.

2.3. Sample preparation

2.3.1. CBCT acquisition and segmentation
Preoperative CBCT scans were obtained at Harvard School of Dental Medicine using the Veraview X800 MORITA device (MORITA
Inc, Kyoto, Japan) for orthodontic, surgical or implant planning diagnosis. Standardized settings were utilized for all CBCT scans,
with an 80x40 field of view (FOV), 100 kV, 7 mA, and a resolution of 1.0 mm.

Manual segmentations were conducted by a single investigator (AN) following a standardized protocol. The CBCT scans were
exported to the Blue-Sky Bio software (Blue Sky Bio, LLC, Libertyville, Illinois), and using the "Advanced Tooth Segmentation" tool
of this software, segmentations were carried out in the area of interest. Teeth were manually outlined layer by layer using the lasso
tool, and subsequent refinements were achieved utilizing the brush tool. Fifteen slices, with a minimum density grey values
threshold of 900, and the “Smooth” function were applied to all manual segmentations. Then, the resulting 3D replicas were saved
in standard tessellation language (STL) files (Fig. 1). AI-driven segmentations were carried out using the Diagnocat® software
(Diagnocat, San Francisco, California) by a single investigator (I.P.). This software uses a Convolutional Neural Networks (CNN)
algorithm following a progressive coarse-to-fine framework for resolution analysis. Then, the resulting 3D replicas were exported
from Diagnocat® to STL files.

Page 3/16
2.3.2 3D printing
All 3D replicas were printed with a 3D printer (Formlabs Form 3B+, Formlabs, Somerville, Massachusetts), using Low Force
Stereolithography (LFS) technology. Temporary crown-bridge (CB) resin (Formlabs, Somerville, Massachusetts) was used as the
printing material. The pre-processed replicas underwent a thorough washing procedure in isopropyl alcohol for 3 minutes
following manufacturer recommendations (Fig. 2).

2.3.3 Post-processing
Following the manufacturer’s recommendation, replicas with the supports still attached were first cured in the Form Cure
(Formlabs, Somerville, Massachusetts) at 60°C (140°F) for 20 minutes. After the first curing, supports and rafts (3–5 supports,
0.70mm diameter, only on the occlusal surface) were manually removed, and replicas were carefully sandblasted to refine their
surface quality. Finally, the replicas underwent a second curing process at 60°C (140°F) for 20 minutes.

2.3.4. Surface scanning

The extracted teeth and their corresponding 3D pre- and post-processed replicas were digitally scanned using a laboratory
scanner (3Shape E Series Lab Scanner, 3Shape, Copenhagen, Denmark). The teeth were secured on the scanner's platform by the
root, and the coronal portion of each tooth was scanned first. Next, the teeth were inverted to scan the apical portion. The coronal
and apical scans were then superimposed and aligned to create a complete 3D surface model of each tooth, which was then
saved as an STL file.

An overview of the methodology can be found in Fig. 3.

2.4. Outcome measurements

2.4.1. Volumetric and linear assessment
For the volumetric measurements, STL files were analyzed by a single examiner (E.C.Q) using a previously published methodology
with a specialized software package (Geomagic Control X, 3D Systems, Rock Hill, SC, USA) [21]. The entire projected volume was
measured in mm3 (Fig. 4 and Fig. 5). The examiner was trained and calibrated by conducting a series of 10 separate volumetric
assessments in duplicate.

To assess linear differences, STL files were exported to a software package (Autodesk Meshmixer, San Francisco, California),
superimposed, aligned, and compared with the STL files from the extracted tooth. These measurements along the X, Y, and Z axes
were carried out by a single examiner (A.N.) using the software's "Unit/Dimensions" analysis tool (Fig. 4 and Fig. 5).

2.5. Statistical analysis

Sample size was calculated based on data from the study of Lee et al. 2022 [19], forecasting a mean difference of 0.38mm in
linear measurements using 3-D printed resin replicas (Formlabs®) [18] with a standard deviation of ± 0.22 for the mean linear
measurement. Based on an alpha error of 5%, a power of 85%, and a two-sided (equivalence) test, 12 specimens were deemed
necessary in this study.

Statistical analyses were done using each extracted tooth as the statistical unit. Outcome variables are presented through
descriptive statistics, expressing continuous variables as means, standard deviations (SD), and confidence intervals of 95%, while
categorical variables are expressed as percentages (%). Data normality was calculated using a Shapiro-Wilk test.

The primary outcome variable was the volumetric and linear changes between the final post-processed 3-D printed replicas with
the original extracted tooth, with either segmentation method used (i.e., manual or AI segmentation). Differences were evaluated
using the 2-sided paired sample Student’s T-test, with a p-value of p ≤ 0.05 as statistically significant. When data did not meet
normality criteria, a Wilcoxon signed-rank test was used. Binary categorical data were evaluated with a Chi-squared test. Intraclass
correlation coefficients (ICCs) were also calculated for each of these comparisons to evaluate the correlation between the
volumetric measurements of different protocols.

Page 4/16
Secondary outcomes include time efficiency between different segmentation methods, as well as volumetric and linear changes
due to post-processing 3D replicas. Continuous variables used paired Student t-test or Wilcoxon signed-rank test depending on
the normality of data, as well as ICCs. All data analyses were performed with SPSS version 21.0 software (Chicago, IL, USA).

3. RESULTS
3.1. Sample Characteristics
The final sample consisted of 30 extracted teeth from 8 patients (5 males and 3 females) with ages ranging between 13 to 55
years (mean age 32.25, SD 14.69). The extracted teeth comprised multi-rooted teeth (1 mandibular and 3 maxillary molars), and
26 were single-rooted teeth (15 premolars (7 mandibular and 8 maxillary), 4 canines (2 mandibular and 2 maxillary), 5 lateral
incisors (2 mandibular and 3 maxillary), and two central incisors (1 mandibular and 1 maxillary)). (Table 1).

Table 1
Sample characteristics
Total (Patients) 8 (%)

Male 5 (62.5)

Female 3 (37.5)

Age

≤ 20 2 (25)

≤ 35 3 (37.5)

35–55 3 (37.5)

Total (Teeth) 30

Molar 4 (13.3)

Premolar 15 (50)

Canine 4 (13.3)

Lateral Incisor 5 (16.7)

Central Incisor 2 (6.7)

Number of Roots

Multi (Three) 1 (3.3)

Multi (Two) 3 (10)

Single 26 (86.7)

3.2. Intra-examiner Reliability

The calibration exercise provided a high intra-examiner agreement, with a strong ICC ranging from 0.97–0.99 for the volumetric
and linear analysis (Table 2).

Page 5/16
 Table 2
Intraclass correlation analysis of volumetric and linear measurements.
Volumetric Linear (Y)

ICC p-value ICC p-value

Pair 1 Extracted 0.997 < .001 0.987 0.001

Vs. Scan_Manual_Postprocessed

Pair 2 Extracted 0.994 < .001 0.971 0.003

Vs. Scan_AI_Postprocessed

Pair 3 Extracted 0.997 < .001 0.995 < .001

Vs. Segmentation_Manual

Pair 4 Extracted 0.998 < .001 0.975 < .001

Vs. Segmentation_AI

Pair 5 Segmentation_AI 0.997 < .001 0.983 < .001

Vs. Segmentation_Manual

Pair 6 Preprocessed_pooled 0.996 0.004 0.987 < .001

Vs. Postprocessed_pooled

Page 6/16
 Table 3
Comparative analysis of volumetric and linear measurements of STL files from virtual segmentation replicas
(A)

Volumetric Linear
(Y)

Mean SD 95% p- Mean SD 95% p-

CI value CI value

Pair Extracted 5.651 19.469 [-1.62, 0.123 0.463 0.335 [0.34, < .001
1 12.92] 0.59]

Vs.
Scan_Manual_Postprocessed

Pair Extracted 22.128 14.917 [16.56, < .001 0.709 0.491 [0. 53, < .001
2 27.70] 0.89]

Vs. Scan_AI_Postprocessed

Pair Extracted 1.766 0.566 [-5.84, 0.638 0.221 0.281 [0.12, < .001
3 9.37] 0.33]

Vs. Segmentation_Manual

Pair Extracted 12.232 11.334 [7.99, < .001 0.550 0.574 [0. 34, < .001
4 16.46] 0.76]

Vs. Segmentation_AI

Pair Segmentation_AI 10.466 17.354 [3.99, 0.003 0.329 0.566 [0. 12, 0.003
5 16.95] .054]

Vs. Segmentation_Manual

(B)

Volumetric Linear
(Y)

Mean SD 95% p- Mean SD 95% p-

CI value CI value

Pair Scan_AI_Preprocessed 21.419 6.523 [18.98, < .001 0.210 0.777 [-.008, 0.150
1 23.85] 0.50]

Vs. Scan_AI_Postprocessed

Pair Scan_Manual_Preprocessed 21.975 7.232 [19.27, < .001 0.093 0.190 [0.02, 0.012
2 24.68] 0.16]

Vs.
Scan_Manual_Postprocessed

Pair Preprocessed_pooled 21.697 6.833 [19.93, < .001 0.151 0.564 [0.01, 0.042
3 23.46] 0.30]

Vs. Postprocessed_pooled

(A) AI-driven and manually driven analysis.

(B) Post-processed effect analysis

3.3. Accuracy of the comprehensive Computer-Aided Rapid Prototyping (CARP) process using either manual segmentation or AI-
driven segmentation

3.3.1 Linear measurements

Page 7/16
A mean linear difference of 0.463mm (SD 0.335) was observed when comparing post-processed 3-D replicas obtained by manual
segmentation with the corresponding extracted teeth. These differences were statistically significant (p < 0.001) (Table 3A).

Similarly, the mean linear difference when comparing post-processed 3-D replicas obtained by AI segmentation with the
corresponding extracted teeth was 0.709mm (SD 0.491), with these differences being statistically significant (p < 0.001)
(Table 3A).

Direct comparison between the 3D replicas obtained from manual and AI segmentation resulted in a mean linear difference of
0.221mm (SD 0.281). These differences were statistically significant (p < 0.001) (Table 3A).

3.3.2. Volumetric measurements

A mean volumetric difference of 5.651mm3 (SD 19.469) was obtained between the manually segmented 3D-printed replicas and
the extracted teeth, corresponding to a 1.20% volume reduction. These differences, however, were not statistically significant (p =
0.123) (Table 3A and Table 4).

Conversely, the mean volumetric difference when comparing replicas obtained by AI segmentation with the corresponding
extracted teeth was 22.128mm³ (SD 14.917), corresponding to a -4.70% volume reduction. These differences were statistically
significant (p < 0.001) (Table 3A and Table 4).

Direct comparison between STL files (3D surfaces) generated from AI-driven and manual segmentation found an overall mean
volumetric difference of 10.466mm³ (SD 17.354, p = 0.003), equivalent to a -2.23% change in volume (Table 3A and Table 4).
Replicas from AI-driven segmentation were smaller than those originating from manual segmentation

Table.4. Descriptive volumetric comparison.

Page 8/16
 Volumetric

Mean Mean diff % diff in Vol

Pair 1 Extracted tooth 470.782 5.651 -1.20

Vs. Scan_Manual_Postprocessed 3D printed replica 465.132

Pair 2 Extracted tooth 470.782 22.128 -4.70

Vs. Scan_AI_Postprocessed 3D printed replica 448.654

Pair 3 Extracted tooth 470.782 1.766 -0.38

Vs. Segmentation_Manual 469.016

Pair 4 Extracted tooth 470.782 12.232 -2.60

Vs. Segmentation_AI 458.550

Pair 5 Segmentation_AI 458.550 10.466 -2.23

Vs. Segmentation_Manual 469.016

Volumetric

Mean Mean diff % diff in Vol

Pair 1 Scan_AI_Preprocessed 470.073 21.419 -4.56

Vs. Scan_AI_Postprocessed 448.654

Pair 2 Scan_Manual_Preprocessed 487.107 21.975 -4.51

Vs. Scan_Manual_Postprocessed 465.132

Pair 3 Preprocessed_pooled 478.590 21.697 -4.53

Vs. Postprocessed_pooled 456.893

Comparisons for pairs 1, 2, 4, and 5 were conducted by establishing the extracted teeth as a reference for comparing the size of
the replicas. In the case of pair 3, AI-segmented 3D replicas were compared against manually segmented replicas, with the
manual group serving as the reference. For all post-processing groups, pre-processed replicas were used as the reference for
comparison.

** “Segmentation manual” and “Segmentation IA” are digital files (STL) that have yet to be printed.

** “Pre-processed” and “Post-processed” represent digital files (STL) obtained after scanning the 3D-printed replicas.

3.4. Accuracy of post-processed replicas in comparison to preprocessed

replicas
Comparative analyses were conducted by separately comparing all the teeth replicas generated through AI-driven segmentation
and manual segmentation, pre- and post-processing. Subsequently, data from all replicas were pooled together to assess the
overall impact of post-processing on both volumetric and linear measurements.

3.4.1. Linear measurements

Analyzing the pre- and post-processed replicas from AI-driven segmentation showed a mean linear difference of 0.210mm (SD
0.777), demonstrating no statistical significance (p = 0.15). In contrast, the same comparison for the manual segmentation
counterpart group indicated a statistically significant mean linear difference of 0.093mm (SD 0.190, p = 0.012). Combining both

Page 9/16
groups, an overall mean linear difference of 0.151mm (SD 0.564) was observed, and this difference was statistically significant (p
= 0.042) (Table 3B).

3.4.2. Volumetric measurements

When comparing the pre and post-processed replicas obtained from AI-driven segmentation, there was a mean volumetric
difference of 21.419mm³ (SD 6.523),-4.56% volume reduction. The same comparison for the replicas obtained from manual
segmentation showed a mean volumetric difference of 21.975mm³ (SD 7.232), -4.51% volume reduction. When both groups (AI
and manually segmented) were pooled together, an overall mean volumetric difference of 21.697mm³ (SD 6.833), -4.53% volume
reduction, was found between pre-processed and post-processed replicas (Table.4). Importantly, all groups demonstrated
statistically significant results (p < 0.001) (Table.3B). Post-processed replicas were generally smaller than the pre-processed ones.
3.5. Comparison of segmentation times between manual and AI-driven
methods
The average time required for manual segmentation of 30 teeth was 23.97 minutes, and the average time required for AI-driven
segmentation of the same teeth was reported to be 2.1 minutes.

4. DISCUSSION
The application of CARP in tooth autotransplantation involves multiple steps, which include CBCT acquisition and segmentation,
3D printing, and subsequent post-processing of 3D-printed replicas. The accuracy of each step may have a substantial impact on
the overall accuracy of the final 3D replica. This study validates the entirety of the CARP process, assessing the accuracy of
manual and AI tooth segmentation and its resulting 3D-printed replicas compared with the reference extracted teeth. We observed
that both methods were reliable and suitable for the fabrication of 3D tooth replicas, as the observed statistically significant
differences between methods can be considered non-clinically significant. However, time-efficiency analysis demonstrates a
reduction in time of 21 minutes for the AI-driven method.

Manual segmentation is a well-established method to obtain tooth replicas, and previous studies have validated its accuracy,
considering it to be the gold standard [2]. However, it is a time-consuming process that demands training and experience and
relies on the interpretation skills of the operator. The average time for manual segmentation of each tooth in this investigation was
23.97 minutes. Nonetheless, other studies, such as Lee et al. [22], reported an average time of 15 minutes per tooth. Another
study on manual segmentations of single and double-rooted teeth reported an average time of 6.6 minutes. Interestingly, AI-driven
segmentation resulted in a 12.5-fold reduced time compared to manual segmentation [23]. Considering that manual segmentation
involves the investigator selecting and individually outlining multiple image slices, the reported time in different studies can vary
significantly based on the type of teeth, the number of slices selected, and the precision of the outlining process [1, 22, 24]. This
fact was noted in this study considering the higher standard deviations in the manual vs. the AI method.

Several AI algorithms and deep learning models have recently been developed to carry out a fully automatic tooth segmentation
more efficiently within a few minutes [1, 23, 25–27]. One of the most effective models is the CNN, which has been integrated into
the software used for AI-driven segmentation in this study [1, 28]. Comprising multilayer neural networks, CNN algorithms excel in
identifying visual patterns quickly and with minimal pre-processing requirements [1, 28]. However, these models have certain
limitations, and recent review studies have underscored the necessity for validating their accuracy and reliability [1]. Several
challenges noted in other studies involve the segmentation of intricate root anatomy and apices, supernumerary and impacted
teeth, especially third molars, and cases of crowding [1, 17, 23, 25, 29]. These factors may reasonably account for our findings
regarding the lower accuracy of AI-driven segmentation versus the manual segmentation group. This finding may also be
explained by the fact that manual segmentation was performed by the same experienced operator under ideal and controlled
circumstances.

This study demonstrated a reduction in volume (-0.38 to -2.6% for manual and AI segmentation, respectively) when comparing the
virtual files obtained after segmentation and the scanned tooth. Interestingly, volumetric and linear analyses of post-processed
replicas showed a smaller trend compared to the extracted teeth (-4.53% volume reduction). These findings could be attributed to
resin shrinkage during post-curing. Similarly, a study by Lee and Kim also reported that 3D replicas from CT images were generally

Page 10/16
smaller than the actual teeth [30]. Their results revealed that, on average, the 3D images of donor teeth were − 0.149 mm smaller
than the actual teeth, and the 3D replicas were, on average, -0.067 mm smaller than their corresponding 3D images. Despite the
observed size discrepancy, it is noteworthy that this error may be clinically acceptable for the application of these replicas in the
context of tooth autotransplantation therapy. Recognizing the benefits of utilizing 3D replicas to reduce extraoral time and
minimize damage to the periodontal ligament, the size discrepancy can be clinically manageable [7]. This factor, coupled together
with the volume reduction after virtual segmentation reported in this study, can be taken into consideration during the planning
phase. Therefore, clinicians should be aware that the surgical area should be minimally overprepared based on the 3D replica and
to allow some physical space for blood clot formation and establishment around the roots of the autotransplanted tooth.

Another important aspect to consider is that, while this study reports AI-driven segmentation as less accurate than the manual
method, it was notably more time-efficient and less dependent on operator input, demonstrated by a higher SD on the manual-
segmentation group. Continuous advancements in AI algorithms and deep learning models hold the promise of significant
improvements in tooth segmentation software. By harnessing the potential for ongoing training and improvement of AI systems,
there is a clear path towards achieving higher accuracy and efficiency in digital segmentation processes. Efficiency is a significant
factor in treatment planning and the practice of modern dentistry. In the context of autotransplantation procedures, earlier studies
indicated that using 3D replicas can significantly enhance the success and efficiency of surgery. Shahbazian et al. and Verweij et
al., reported extra-oral times of less than 1 minute when 3D replicas were employed and an overall significant reduction in
procedural time [7, 8, 31, 32]. If AI can streamline the treatment planning phase by reducing the time and effort required for tooth
segmentation while maintaining an acceptable level of accuracy, it holds the potential to be a promising tool for enhancing the
overall efficiency of surgical treatment planning.

The relevance of this investigation lies in a direct evaluation of the accuracy of AI-driven tooth segmentation, both in terms of
volumetric and linear data. However, this study also presents some limitations that should be acknowledged. Firstly, the strict
inclusion criteria and consistent use of the same CBCT machine, parameters, and standardized operator enhance the study's
internal validity but also make it challenging to extrapolate these findings to other protocols. Secondly, only one software for AI-
driven segmentation has been tested, as well as 3D printing workflow, and the reported accuracy may not apply to other
approaches utilizing different technologies. Lastly, despite conducting a sample size calculation and achieving high power, further
studies with larger sample sizes are recommended to investigate the influence of multi-radicular teeth, furcation areas, and
complex anatomy on the segmentation process, as well as the relationship between the time dedicated to manual segmentation
and its final accuracy.

5. CONCLUSIONS
Within the limitations of this study, the following conclusions can be inferred:

1. AI-driven and manual virtual tooth segmentation methods are reliable and suitable for obtaining 3D tooth replicas.
2. AI-driven tooth segmentation proved to be more time-efficient and independent of the operator's experience.
3. Post-processing 3D-printed tooth replicas showed consistently reduced dimensions.

Declarations
Authors’ contributions: IP (Concept/Design, Data analysis/interpretation, Critical revision of article, Data collection, Approval of
article); I.P, A.N, J.C, E.C.Q, M.S (Data analysis/interpretation, Critical revision of article, Data collection, Writing, Approval of
article), W.V.G, G.G (Critical revision of article, Approval of article). All authors critically revised the manuscript, gave final approval,
and agreed to be accountable for all aspects of the scientific work.

Conflict of interest: The authors have no conflicts of interest to report pertaining to the conduction of this study.

Data availability statement: The data that support the findings of this study are available from the corresponding author upon
reasonable request.

Ethics approval statement: This study was approved by the Harvard of Dental Medicine Institutional Review Board (IRB21-1687).

Page 11/16
Funding statement: No financial support or sponsorship was received for the conduction of this study.

No datasets were generated or analysed during the current study.

No funding has been obtained for the development of this investigation.

References
1. Polizzi A, Quinzi V, Ronsivalle V, Venezia P, Santonocito S, Lo Giudice A, Leonardi R and Isola G (2023) Tooth automatic
segmentation from CBCT images: a systematic review. Clin Oral Investig 27:3363-3378. doi: 10.1007/s00784-023-05048-5
2. Shahbazian M, Jacobs R, Wyatt J, Willems G, Pattijn V, Dhoore E, C VANL and Vinckier F (2010) Accuracy and surgical
feasibility of a CBCT-based stereolithographic surgical guide aiding autotransplantation of teeth: in vitro validation. J Oral
Rehabil 37:854-9. doi: 10.1111/j.1365-2842.2010.02107.x
3. Moin DA, Hassan B, Parsa A, Mercelis P and Wismeijer D (2014) Accuracy of preemptively constructed, cone beam CT-, and
CAD/CAM technology-based, individual Root Analogue Implant technique: an in vitro pilot investigation. Clin Oral Implants
Res 25:598-602. doi: 10.1111/clr.12104
4. Liaw CY and Guvendiren M (2017) Current and emerging applications of 3D printing in medicine. Biofabrication 9:024102. doi:
10.1088/1758-5090/aa7279
5. Lee S, Woo S, Yu J, Seo J, Lee J and Lee C (2020) Automated CNN-Based Tooth Segmentation in Cone-Beam CT for Dental
Implant Planning. IEEE access 8:50507-50518. doi: 10.1109/ACCESS.2020.2975826
6. Wang H, Minnema J, Batenburg KJ, Forouzanfar T, Hu FJ and Wu G (2021) Multiclass CBCT Image Segmentation for
Orthodontics with Deep Learning. Journal of dental research 100:943-949. doi: 10.1177/00220345211005338
7. Verweij JP, Jongkees FA, Anssari Moin D, Wismeijer D and van Merkesteyn JPR (2017) Autotransplantation of teeth using
computer-aided rapid prototyping of a three-dimensional replica of the donor tooth: a systematic literature review. Int J Oral
Maxillofac Surg 46:1466-1474. doi: 10.1016/j.ijom.2017.04.008
8. Verweij JP, van Westerveld KJH, Anssari Moin D, Mensink G and van Merkesteyn JPR (2020) Autotransplantation With a 3-
Dimensionally Printed Replica of the Donor Tooth Minimizes Extra-Alveolar Time and Intraoperative Fitting Attempts: A
Multicenter Prospective Study of 100 Transplanted Teeth. J Oral Maxillofac Surg 78:35-43. doi: 10.1016/j.joms.2019.08.005
9. Dhillon IK, Khor MMY, Tan BL, Wong RCW, Duggal MS, Soh SH and Lu WW (2023) Tooth autotransplantation with 3D‐printed
replicas as part of interdisciplinary management of children and adolescents: Two case reports. Dental Traumatology 39:81-
89. doi: 10.1111/edt.12837
10. Czochrowska EM, Stenvik A, Album B and Zachrisson BU (2000) Autotransplantation of premolars to replace maxillary
incisors: A comparison with natural incisors. American journal of orthodontics and dentofacial orthopedics 118:592-600. doi:
10.1067/mod.2000.110521
11. Andreasen JO, Paulsen HU, Yu Z, Ahlquist R, Bayer T and Schwartz O (1990) A long-term study of 370 autotransplanted
premolars. Part I. Surgical procedures and standardized techniques for monitoring healing. European journal of orthodontics
12:3-13. doi: 10.1093/ejo/12.1.3
12. Han S, Wang H, Chen J, Zhao J and Zhong H (2022) Application effect of computer-aided design combined with three-
dimensional printing technology in autologous tooth transplantation: a retrospective cohort study. BMC Oral Health 22:5. doi:
10.1186/s12903-021-02030-z
13. Lee SJ, Jung IY, Lee CY, Choi SY and Kum KY (2001) Clinical application of computer-aided rapid prototyping for tooth
transplantation. Dent Traumatol 17:114-9. doi: 10.1034/j.1600-9657.2001.017003114.x
14. Cousley RRJ, Gibbons A and Nayler J (2017) A 3D printed surgical analogue to reduce donor tooth trauma during
autotransplantation. J Orthod 44:287-293. doi: 10.1080/14653125.2017.1371960
15. Lucas‐Taulé E, Llaquet M, Muñoz‐Peñalver J, Nart J, Hernández‐Alfaro F and Gargallo‐Albiol J (2021) Mid‐term outcomes and
periodontal prognostic factors of autotransplanted third molars: A retrospective cohort study. Journal of periodontology
(1970) 92:1776-1787. doi: 10.1002/JPER.21-0074

Page 12/16
16. Zanjani FG, Pourtaherian A, Zinger S, Moin DA, Claessen F, Cherici T, Parinussa S and de With PHN (2021) Mask-MCNet: Tooth
instance segmentation in 3D point clouds of intra-oral scans. Neurocomputing (Amsterdam) 453:286-298. doi:
10.1016/j.neucom.2020.06.145
17. Gardiyanoğlu E, Ünsal G, Akkaya N, Aksoy S and Orhan K (2023) Automatic Segmentation of Teeth, Crown-Bridge
Restorations, Dental Implants, Restorative Fillings, Dental Caries, Residual Roots, and Root Canal Fillings on
Orthopantomographs: Convenience and Pitfalls. Diagnostics (Basel) 13:1487. doi: 10.3390/diagnostics13081487
18. Vinayahalingam S, Kempers S, Schoep J, Hsu T-MH, Moin DA, van Ginneken B, Flügge T, Hanisch M and Xi T (2023) Intra-oral
scan segmentation using deep learning. BMC oral health 23:1-643. doi: 10.1186/s12903-023-03362-8
19. Lee CKJ, Foong KWC, Sim YF and Chew MT (2022) Evaluation of the accuracy of cone beam computed tomography (CBCT)
generated tooth replicas with application in autotransplantation. J Dent 117:103908. doi: 10.1016/j.jdent.2021.103908
20. Krithikadatta J, Gopikrishna V and Datta M (2014) CRIS Guidelines (Checklist for Reporting In-vitro Studies): A concept note
on the need for standardized guidelines for improving quality and transparency in reporting in-vitro studies in experimental
dental research. Journal of conservative dentistry 17:301-304. doi: 10.4103/0972-0707.136338
21. Couso-Queiruga E, Ahmad U, Elgendy H, Barwacz C, Gonzalez-Martin O and Avila-Ortiz G (2021) Characterization of Extraction
Sockets by Indirect Digital Root Analysis. The International journal of periodontics & restorative dentistry 41:141-148. doi:
10.11607/prd.4969
22. Lee S-C, Hwang H-S and Lee KC (2022) Accuracy of deep learning-based integrated tooth models by merging intraoral scans
and CBCT scans for 3D evaluation of root position during orthodontic treatment. Progress in orthodontics 23:15-15. doi:
10.1186/s40510-022-00410-x
23. Lahoud P, EzEldeen M, Beznik T, Willems H, Leite A, Van Gerven A and Jacobs R (2021) Artificial Intelligence for Fast and
Accurate 3-Dimensional Tooth Segmentation on Cone-beam Computed Tomography. Journal of endodontics 47:827-835. doi:
10.1016/j.joen.2020.12.020
24. Al-Ubaydi AS and Al-Groosh D (2023) The Validity and Reliability of Automatic Tooth Segmentation Generated Using Artificial
Intelligence. TheScientificWorld 2023:5933003-11. doi: 10.1155/2023/5933003
25. Gan Y, Xia Z, Xiong J, Zhao Q, Hu Y and Zhang J (2015) Toward accurate tooth segmentation from computed tomography
images using a hybrid level set model. Medical physics (Lancaster) 42:14-27. doi: 10.1118/1.4901521
26. Cui Z, Fang Y, Mei L, Zhang B, Yu B, Liu J, Jiang C, Sun Y, Ma L, Huang J, Liu Y, Zhao Y, Lian C, Ding Z, Zhu M and Shen D (2022)
A fully automatic AI system for tooth and alveolar bone segmentation from cone-beam CT images. Nature communications
13:2096-2096. doi: 10.1038/s41467-022-29637-2
27. Jang TJ, Kim KC, Cho HC and Seo JK (2022) A fully automated method for 3D individual tooth identification and segmentation
in dental CBCT. IEEE transactions on pattern analysis and machine intelligence 44:1-1. doi: 10.1109/TPAMI.2021.3086072
28. Ezhov M, Zakirov A and Gusarev M Coarse-to-fine volumetric segmentation of teeth in cone-beam ct. Book title. IEEE,
29. Hao J, Liao W, Zhang YL, Peng J, Zhao Z, Chen Z, Zhou BW, Feng Y, Fang B, Liu ZZ and Zhao ZH (2022) Toward Clinically
Applicable 3-Dimensional Tooth Segmentation via Deep Learning. Journal of dental research 101:304-311. doi:
10.1177/00220345211040459
30. Lee S-J and Kim E-S (2012) Minimizing the extra-oral time in autogeneous tooth transplantation: use of computer-aided rapid
prototyping (CARP) as a duplicate model tooth. Restorative dentistry & endodontics 37:136-141.
31. Verweij JP, Moin DA, Mensink G, Nijkamp P, Wismeijer D and Merkesteyn JPRv (2016) Autotransplantation of Premolars With
a 3-Dimensional Printed Titanium Replica of the Donor Tooth Functioning as a Surgical Guide: Proof of Concept. Journal of
Oral and Maxillofacial Surgery 74:1114-1119. doi: 10.1016/j.joms.2016.01.030
32. Shahbazian MDDSP, Jacobs RDDSP, Wyatt JDDSM, Denys DDDSM, Lambrichts IDDSP, Vinckier FDDSP and Willems GDDSP
(2013) Validation of the cone beam computed tomography–based stereolithographic surgical guide aiding
autotransplantation of teeth: clinical case–control study. ORAL SURGERY ORAL MEDICINE ORAL PATHOLOGY ORAL
RADIOLOGY 115:667-675. doi: 10.1016/j.oooo.2013.01.025

Figures

Page 13/16
Figure 1

Manual segmentation of a representative tooth from CBCT data.

Figure 2

Side-by-side views of the extracted tooth next (A) to its corresponding 3D printed-replica from virtual segmentation (B).

Page 14/16
Figure 3

Summary of methodology. Using the obtained STL files, comparisons were made between E vs M, E vs A, E vs M2, E vs A2, and M
vs A groups to assess segmentation accuracy; and effect of post-processing was assessed with comparisons between M1 vs M2
and A1 vs A2 groups. Time required for both manual and AI segmentation was recorded.

Figure 4

(A) Paired superimposition of the STL files showing discrepancies between files as represented by 3D comparison color-map. (B)
Paired superimposition of the STL files for linear analysis.

Page 15/16
Figure 5

Example of the methodology followed. Left: AI segmentation analysis and Right: manual segmentation analysis. Each row, from
left to right: Geomagic superimposition for volumetric analysis (3D color map comparison); Meshmixer superimposition for linear
analysis (dark grey); SLT file of extraorally scanned tooth; STL file of (IA or Manual) segmented tooth.

Page 16/16