Trueness of artificial intelligence‐based, manual, and global thresholding segmentation protocols for human mandibles

Abstract

Purpose

To compare the trueness of artificial intelligence (AI)‐based, manual, and global segmentation protocols by superimposing the resulting segmented 3D models onto reference gold standard surface scan models.

Materials and Methods

Twelve dry human mandibles were used. A cone beam computed tomography (CBCT) scanner was used to scan the mandibles, and the acquired digital imaging and communications in medicine (DICOM) files were segmented using three protocols: global thresholding, manual, and AI‐based segmentation (Diagnocat; Diagnocat, San Francisco, CA). The segmented files were exported as study 3D models. A structured light surface scanner (GoSCAN Spark; Creaform 3D, Levis, Canada) was used to scan all mandibles, and the resulting reference 3D models were exported. The study 3D models were compared with the respective reference 3D models by using a mesh comparison software (Geomagic Design X; 3D Systems Inc, Rock Hill, SC). Root mean square (RMS) error values were recorded to measure the magnitude of deviation (trueness), and color maps were obtained to visualize the differences. Comparisons of the trueness of three segmentation methods for differences in RMS were made using repeated measures analysis of variance (ANOVA). A two‐sided 5% significance level was used for all tests in the software program.

Results

AI‐based segmentations had significantly higher RMS values than manual segmentations for the entire mandible (p < 0.001), alveolar process (p < 0.001), and body of the mandible (p < 0.001). AI‐based segmentations had significantly lower RMS values than manual segmentations for the condyles (p = 0.018) and ramus (p = 0.013). No significant differences were found between the AI‐based and manual segmentations for the coronoid process (p = 0.275), symphysis (p = 0.346), and angle of the mandible (p = 0.344). Global thresholding had significantly higher RMS values than manual segmentations for the alveolus (p < 0.001), angle of the mandible (p < 0.001), body of the mandible (p < 0.001), condyles (p < 0.001), coronoid (p = 0.002), entire mandible (p < 0.001), ramus (p < 0.001), and symphysis (p < 0.001). Global thresholding had significantly higher RMS values than AI‐based segmentation for the alveolar process (p = 0.002), angle of the mandible (p < 0.001), body of the mandible (p < 0.001), condyles (p < 0.001), coronoid (p = 0.017), mandible (p < 0.001), ramus (p < 0.001), and symphysis (p < 0.001).

Conclusion

AI‐based segmentations produced lower RMS values, indicating truer 3D models, compared to global thresholding, and showed no significant differences in some areas compared to manual segmentation. Thus, AI‐based segmentation offers a level of segmentation trueness acceptable for use as an alternative to manual or global thresholding segmentation protocols.

Segmentation involves partitioning scanned images to isolate and define boundaries within regions of interest. ¹ , ² Accurate segmentation results in 3D models crucial for diagnosis and treatment planning. ³ , ⁴ , ⁵ Success in segmentation and 3D modeling heavily relies on cone beam computed tomography (CBCT) imaging and data acquisition, with factors such as the type of scanner, imaging technique, angle, and resolution significantly influencing algorithm success. ⁶ , ⁷ CBCT artifacts, often seen in the lingual and posterior borders of the mandible, condyles, and alveolar crest, can introduce streak and ring halation defects, distorting the digital dataset. ⁸ , ⁹ Additionally, image slice spacing and surface geometry changes can impact the accuracy of the 3D model. ¹ Loss of image detail may also occur during post‐processing due to manual editing. ³ , ¹⁰ Finally, user skill plays a crucial role, as segmentations by commercial software companies have been shown to yield more accurate 3D models than those performed by clinicians. ⁹ , ¹¹

Global thresholding uses a threshold value “t” to include voxel gray values greater or equal to “t” in the segmented image. ⁵ The method's simplicity makes it widely available but has drawbacks, leading to the least accurate 3D models. Challenges in delineating borders arise when different tissue voxels are at the borders, and the method does not account for CBCT scan errors like metal artifacts or noise. ⁵ , ¹⁰ , ¹² Global thresholding requires extensive post‐processing and expertise, increasing the cost and time for accurate 3D models. ⁵ , ¹² In manual segmentation, each part of a CBCT image is outlined slice‐by‐slice and then compiled into 2D or 3D images. It can address issues such as missing slices, recognized as ground truth in medical image analysis due to expert input. ¹³ This method produces highly accurate segmentations with volumes matching the original region of interest. ¹⁴ However, due to its precision and the expertise required, manual segmentation is costly, time‐intensive, and subject to variability and subjective bias. ¹³ , ¹⁴

To address the limitations of manual segmentation, semi‐automatic and automatic segmentation methods, including artificial intelligence (AI)‐based auto‐segmentation, have been developed. These use computational tools to delineate regions of interest rapidly with little or no manual input. ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ For example, one AI technology (Diagnocat; Diagnocat, San Francisco, CA) uses a deep multilayer convolutional neural network (CNN) that makes use of a U‐net‐like architecture, which can simulate human brain architecture and processes data through a network of so‐called interconnected neurons. ¹⁹ AI‐based segmented 3D models are instrumental in dental and maxillofacial radiology, assisting in diagnosing cystic lesions, treatment planning, and creating surgical implant guides. ²⁰ However, despite their potential, many advanced segmentation software programs are not yet commercially available. ⁵

Limited research has compared AI‐based segmentation to global thresholding, the most straightforward and prevalent technique, and the highly accurate manual segmentation. Moreover, the accuracy of AI‐based segmentation has been underexplored despite its potential to outperform human and other methods and become a market‐ready software. This study aimed to evaluate the trueness of AI‐based segmentation by comparing it with global thresholding and manual segmentation, using superimposition and comparison against a reference gold standard surface scan model. The null hypothesis was that the segmentation protocol has no effect on the trueness of the resulting 3D models.

MATERIALS AND METHODS

This study was reviewed by the Human Research Protection Program (HRPP) before inception, and it was determined that further review was not required (#19615). The inclusion criteria specified that each mandible must contain at least one tooth, as the AI‐based segmentation software (Diagnocat; Diagnocat) included in this study requires the presence of one remaining tooth. The exclusion criteria ruled out mandibles with fractures or defects that could compromise anonymity or interfere with accurately comparing anatomical landmarks. Twelve anonymized, dry, and macerated mandibles were obtained from the Department of Anatomy, Cell Biology, and Physiology and included in this study (Figure 1a). With a sample size of 12 mandibles, the study had 80% power to detect an effect size of 1.0 between methods, based on a paired t‐test calculation at a two‐sided 5% significance level.

FIGURE 1.

(a) Representative image of a dried mandible included in the study; (b) All mandibles were scanned using a structured light 3D surface scanner. The resulting 3D surface models were exported as reference standard tessellation language (STL) files.

All procedures were standardized among researchers before data experimentation. An experienced radiologist defined each segmentation process and mandibular landmarks on the reference 3D models. Examiners underwent training, and calibration was assessed by repeating intra‐examiner readings at a one‐week interval until intraclass correlation coefficients (ICC) exceeded 0.80. After calibration, the first examiner randomized and numbered the segmented digital imaging and communications in medicine (DICOM) files, maintained a master sheet for data matching, and sent blinded files to the second examiner, who completed segmentation, alignment, measurement, and comparisons.

All mandibles were scanned using a structured light surface scanner (GoSCAN Spark; Creaform 3D, Levis, Canada) (Figure 1b). The resulting 3D surface models were exported as standard tessellation language (STL) files. These models were then imported into computer‐aided design and computer‐aided manufacturing (CAD‐CAM) software (Autodesk Meshmixer; Autodesk Inc, San Francisco, CA) to select and delineate the mandible's anatomical regions for comparative analysis. Seven anatomical landmarks critical for implant‐related planning in the mandible were identified on the 3D surface models. These included the right and left condyles, coronoid processes, rami, angles, body, remaining teeth and alveolar process, and symphysis. Each region and the entire mandible were saved as a separate STL file to serve as a reference 3D model for subsequent comparisons in the study (Figure 2).

FIGURE 2.

Anatomical regions, including the right and left condyles, coronoid processes, rami, angles, the body of the mandible, remaining teeth, alveolar process and remaining teeth, symphysis, and the entire mandible (regions a to g combined), were saved as individual standard tessellation language (STL) files. These files served as references for subsequent comparative analyses in the study. (a) Condyles; (b) Coronoid processes; (c) Rami; (d) Angles of the mandible; (e) Body of the mandible; (f) Symphysis; (g) Remaining teeth and alveolar process.

All mandibles were scanned using a CBCT scanner (Planmeca Viso G7; Planmeca USA Inc, Hoffman Estates, IL) with a field of view (FOV) measuring 100 × 140 mm, voxel size of 0.15 mm, anode voltage of 100 kV, and anode current of 63 mAs per second. The acquired DICOM files were exported and stored. Three different segmentation protocols were employed to create the 3D models: global thresholding segmentation, manual segmentation, and AI‐based segmentation. Initially, the DICOM file was uploaded to the segmentation and implant planning software (BlueSkyPlan; BlueSky Bio, Liberty Drive, IL, USA), where the FOV and panoramic curve were identified. To obtain the global thresholding segmentation 3D model, a single threshold value was selected for inclusion in the segmented model, effectively removing all scatter, noise, and artifacts from the imported CBCT scans. The CBCT scan was then manually segmented slice‐by‐slice using the same software. The segmentation process began in the axial view, using the interactive lasso tool to select the entire slice, followed by the manual brush tool, if necessary, to define the whole outline of the slice. Subsequently, the “fill hole” option was used to segment the remaining areas within the borders. After completing the segmentation of all slices from the axial view, the process was repeated in the cross‐sectional view until all slices were segmented. For the AI‐based segmentations, the CBCT scans were imported into an AI‐based segmentation software (Diagnocat; Diagnocat), where the software fully automated the segmentations. The resulting study STL files of 3D models from all segmentation protocols were saved and exported for comparative analysis.

The STL files generated by the three segmentation protocols were compared with the respective reference files from the surface scanner using point‐based, global, and fine alignment methods. All STL files were imported into CAD software (Geomagic Design X; 3D Systems Inc, Rock Hill, SC). An initial alignment was executed using four consistent anatomical reference points: the right and left mental foramina and the right and left antegonial notches to superimpose the study STL files onto the reference files. A best‐fit algorithm incorporating global and fine registration was applied to automatically align all study and reference files. A 3D compare tool was employed to evaluate mesh deviations (Figure 3). Root mean square (RMS) values were recorded to measure the magnitude of deviation (trueness) for each segmented file compared to the reference surface scans. Color maps were created to visualize the differences, with shades from yellow to red indicating positive deviations, signifying an overestimation of the segmented files relative to the references. Conversely, blue indicated negative deviations, signifying an underestimation of the segmented files compared to the references.

FIGURE 3.

(a) Standard tessellation language (STL) file from the surface scan used as a reference; (b) STL file of a study 3D model from segmentation; (c) A best‐fit algorithm incorporating global and fine registration was applied to align all study and reference files automatically.

Comparisons among the global thresholding, manual segmentation, and AI‐based segmentation methods for differences in RMS were made using repeated measures analysis of variance (ANOVA), which allowed correlations among the measurements from the three methods on the same mandible and different variances for each method. A two‐sided 5% significance level was used for all tests in a software program (SAS version 9.4; SAS Institute Inc, Cary, NC).

RESULTS

The RMS values for global thresholding, manual, and AI‐based segmentation are summarized in Table 1. A boxplot was used to depict the graphic summary of RMS values from each subgroup (Figure 4). Comparisons among the global thresholding, manual segmentation, and AI‐based segmentation methods for differences in RMS were made using repeated measures ANOVA and pair‐wise comparisons (Table 2). Based on pair‐wise comparison data, AI‐based segmentations had significantly higher RMS values than manual segmentations for the entire mandible (p < 0.001), alveolar process and remaining teeth (p < 0.001), and body of the mandible (p < 0.001). AI‐based segmentations had significantly lower RMS values than manual segmentations for the condyles (p = 0.018) and rami (p = 0.013). No significant differences were found between pair‐wise comparisons between RMS values for AI‐based and manual segmentations for the coronoid process (p = 0.275), symphysis (p = 0.346), and angles of the mandible (p = 0.344).

TABLE 1.

The descriptive statistics of each group's root mean square (RMS, mm) and mean (standard deviation).

Location	Segmentation methods	Mean (standard deviation)
Alveolar process and remaining teeth	AI‐based	1.304 (0.093)
	Global thresholding	1.529 (0.202)
	Manual	0.943 (0.168)
Angles of the mandible	AI‐based	1.089 (0.238)
	Global thresholding	1.927 (0.179)
	Manual	0.986 (0.226)
Body of the mandible	AI‐based	1.432 (0.209)
	Global thresholding	2.100 (0.114)
	Manual	0.659 (0.154)
Condyles	AI‐based	0.742 (0.443)
	Global thresholding	1.951 (0.342)
	Manual	1.134 (0.331)
Coronoid processes	AI‐based	0.748 (0.246)
	Global thresholding	1.026 (0.345)
	Manual	0.709 (0.217)
Rami	AI‐based	0.684 (0.117)
	Global thresholding	1.839 (0.278)
	Manual	0.947 (0.257)
Symphysis	AI‐based	0.516 (0.570)
	Global thresholding	2.257 (0.243)
	Manual	0.668 (0.316)
Mandible (entire mandible)	AI‐based	1.148 (0.112)
	Global thresholding	1.871 (0.150)
	Manual	0.794 (0.169)

Abbreviation: AI, artificial intelligence; RMS, root mean square.

FIGURE 4.

Boxplot showing the graphic summary of descriptive statistics on root mean square (RMS) values.

TABLE 2.

Pair‐wise comparisons were made to compare RMS values from all subgroups.

Locations	Comparison made		Difference	SE	p‐value
Alveolar process and remaining teeth	AI‐based	Global thresholding	−0.2247	0.0536	0.002 *
	AI‐based	Manual	0.3607	0.0385	<0.001 *
	Global thresholding	Manual	0.5854	0.0510	<0.001 *
Angles of the mandible	AI‐based	Global thresholding	−0.8385	0.0952	<0.001 *
	AI‐based	Manual	0.1028	0.1039	0.344
	Global thresholding	Manual	0.9413	0.0563	<0.001 *
Body of the mandible	AI‐based	Global thresholding	−0.6679	0.0685	<0.001 *
	AI‐based	Manual	0.7728	0.0626	<0.001 *
	Global thresholding	Manual	1.4407	0.0411	<0.001 *
Condyles	AI‐based	Global thresholding	−1.2087	0.1860	<0.001 *
	AI‐based	Manual	−0.3920	0.1417	0.018 *
	Global thresholding	Manual	0.8167	0.1406	<0.001 *
Coronoid processes	AI‐based	Global thresholding	−0.2781	0.0990	0.017 *
	AI‐based	Manual	0.0390	0.0340	0.275
	Global thresholding	Manual	0.3171	0.0806	0.002 *
Rami	AI‐based	Global thresholding	−1.1552	0.0952	<0.001 *
	AI‐based	Manual	−0.2632	0.0888	0.013 *
	Global thresholding	Manual	0.8921	0.0520	<0.001 *
Symphysis	AI‐based	Global thresholding	−1.7406	0.1762	<0.001 *
	AI‐based	Manual	−0.1522	0.1547	0.346
	Global thresholding	Manual	1.5884	0.1124	<0.001 *
Mandible (entire mandible)	AI‐based	Global thresholding	−0.7234	0.0467	<0.001 *
	AI‐based	Manual	0.3535	0.0474	<0.001 *
	Global thresholding	Manual	1.0769	0.0360	<0.001 *

Abbreviation: AI, artificial intelligence; RMS, root mean square.

^*Denotes the significant finding, α = 0.05.

Global thresholding had significantly higher RMS values than manual segmentations for the alveolar process and remaining teeth (p < 0.001), angles of the mandible (p < 0.001), body of the mandible (p < 0.001), condyles (p < 0.001), coronoid processes (p = 0.002), entire mandible (p < 0.001), rami (p < 0.001), and symphysis (p < 0.001). AI had significantly lower RMS than global thresholding for the alveolar process and remaining teeth (p = 0.002), angles of the mandible (p < 0.001), body of the mandible (p < 0.001), condyles (p < 0.001), coronoid processes (p = 0.017), entire mandible (p < 0.001), rami (p < 0.001), and symphysis (p < 0.001).

Color maps of the surface matching differences for each group were evaluated, where green‐colored areas represented surface matching within a range of ±0.1 mm. Global thresholding predominantly revealed yellow to red areas, indicating greater overestimation and positive deviation compared to the gold standard surface scan (Figure 5a). The color mapping of manual segmentation ranged from green to light blue, suggesting minimal deviation and higher trueness compared to the surface scans (Figure 5b). AI segmentations exhibited a positive deviation, with the majority of areas ranging from green to yellow. It is important to note that some areas in the AI‐based segmentation, although AI could recognize, were shaded in dark blue, thus representing grossly negative deviations in these areas (Figure 5c). It is also important to note that AI‐based segmentation failed to recognize specific areas, such as the four condyles were omitted entirely from the segmentation (Figure 6).

FIGURE 5.

Color maps representing deviation (in mm) from all 12 mandibles. (a) Global thresholding; (b) Manual segmentation; (c) Artificial intelligence (AI)‐based segmentation.

FIGURE 6.

(a) Manual segmentation of a mandible with condyle present; (b) Artificial intelligence (AI)‐based segmentation of the same mandible. The red arrow indicates the missing condyle in the AI‐based segmentation protocol.

DISCUSSIONS

The null hypothesis was partially rejected as AI segmentation, for most areas analyzed, had significantly higher RMS values, more significant deviation than manual segmentation, and significantly lower RMS values than global thresholding.

In prosthodontics, CBCT segmentation has been utilized to create anatomical bone models to facilitate the planning of esthetic crown lengthening and the design of surgical templates. ²¹ When there is a need to duplicate a clinically satisfactory complete denture as a trial prosthesis or custom tray, CBCT segmentation has also been proposed to obtain a digital file of the denture. A study demonstrated that global thresholding in this duplication technique resulted in a digital denture file with an accuracy level of 0.249 ± 0.020 mm. ²² CBCT enables 3D imaging of craniofacial hard tissues but has limited FOV and contrast for soft tissues. Integrating facial scans or photos enhances CBCT segmentation, creating a photorealistic 3D virtual patient for improved diagnosis, treatment planning, and surgical simulation. ²³ Additionally, CBCT segmentation allows custom articulators by replicating patient‐specific jaw movements. ²⁴ It also enables a digital workflow in maxillofacial prosthodontics, using 3D modeling for patients with limited mouth opening when conventional impressions are impractical. ²⁵ , ²⁶ , ²⁷ Additionally, segmentation accurately records maxillofacial defects, integrating with intraoral or facial scans to create precise jaw and defect replicas. ²⁵ , ²⁶ , ²⁷ CBCT segmentation is widely used in implant dentistry as well, enabling the superimposition of segmented dentition onto intraoral and facial scans for designing implant placement templates. ²⁸ , ²⁹ , ³⁰ In severe mandibular atrophy, it facilitates custom 3D‐printed Ti6Al4V subperiosteal implants, reducing the need for complex regenerative procedures. ³¹ It also preserves the natural emergence profile before extraction, aiding in custom healing abutment design. ³² Despite its broad applications, research comparing segmentation protocols is limited. In this study, the trueness of the AI‐based segmentation protocol ranged from 0.516 ± 0.570 to 1.432 ± 0.209 mm, with missing data in 4 out of 24 condyles. Variability and inaccuracies near critical structures pose risks, highlighting the need for optimized, accurate, and user‐friendly segmentation protocols to ensure patient safety.

Global thresholding is a simple and widely used segmentation technique, relying on a single threshold value for segmentation. ⁵ In this study, global thresholding had the greatest deviation in all locations compared to the gold standard surface scan. Choosing a single threshold value makes it difficult to delineate the borders due to the multiple types of voxels in these areas. ⁵ , ¹⁰ The final segmented images of all mandibles were overestimated versions of their surface scan counterparts. A single threshold struggles to delineate borders due to voxel variability, artifacts, noise, and scanner differences, causing inconsistencies. ⁵ , ¹² Pre‐processing methods like background illumination and top‐hat filtering can enhance object‐background distinction, but global thresholding remains limited as it lacks spatial pixel distribution information and is prone to inaccuracies with artifacts. ³³

Manual segmentation was the truest method in this study, consistently achieving lower RMS values than global thresholding and AI segmentation in the alveolus, body of the mandible, and the entire mandible. It involves slice‐by‐slice annotation, ensuring accuracy but making it time‐consuming and labor‐intensive. ³⁴ , ³⁵ , ³⁶ This method also suffers from inter‐ and intra‐observer variability, often requiring radiologists or specialized clinicians. ¹³ , ¹⁴ The results of this study should also be interpreted with caution, as the manual segmentation was conducted by a trained and calibrated research team, which may reduce inter‐ and intra‐observer variability compared to what is observed in clinical outcomes. Radiodense structures like teeth and compact bone are easier to segment, whereas thinner cortical bone, particularly in the condyle and edentulous spaces, showed greater RMS values than AI‐based segmentation. These regions were more challenging to segment, requiring manual adjustments, which contributed to increased variability and segmentation difficulty in this study.

AI‐based segmentation is increasingly favored for its reproducibility and reduced human error. The system in this study employs a CNN with data‐driven algorithms for rapid, automated segmentation. ¹⁷ , ¹⁸ Small errors can be manually refined, and some metal artifacts are automatically removed. ³⁷ AI segmentation had significantly lower RMS values than global thresholding and was truer than manual segmentation in the condyles and ramus, with no significant differences in the coronoid process, symphysis, or mandibular angle. However, it failed to recognize four condyles (Figure 6), likely due to insufficient database representation, leading to underestimation in color mapping. AI segmentation also had higher RMS values than manual segmentation in the entire mandible, alveolar process, and body of the mandible, resulting in overestimation of these structures. These variations are likely due to anatomical differences between subjects. Despite these limitations, AI segmentation has strong potential, with accuracy expected to improve as more data are incorporated. Similar to the findings from this study, past research has shown favorable results in the accuracy of AI segmentations comparable to the manual segmentations done by experienced professionals. Kargilis et al. utilized CBCT scans of mandible and mandibular dentition segmentations that have undergone model training and testing. ³⁸ The AI segmentation achieved high quantitative accuracy, with an average dice similarity coefficient (DSC) from 0.940 to 0.945. In the blind qualitative assessment, oral and maxillofacial surgeons rated the AI‐generated segmentations as comparable in quality to manual segmentations. ³⁸ Lo Giudice et al. compared 20 manual CBCT‐derived segmentations with CNN‐based AI segmentations and found comparable accuracy for delineating the entire mandible. ³⁹ Verhelst et al. utilized layered 3D U‐Net architecture deep learning algorithms to perform AI segmentations with and without user refinement. ⁴⁰ Manual segmentations averaged 1218.4 s while AI segmentations averaged 17 s to complete. Mean RMS values between fully automatic AI and manual segmentations were 0.2624 mm and the intersection of union of 94.6%. Research has shown favorable results for AI segmentations, both in terms of accuracy as well as timing. ⁴⁰ Recent research found that a newly developed SegResNet‐based deep learning model achieved high accuracy and consistency when automatically segmenting the mandible in CBCT scans from patients with stage III to IV periodontitis. Quantitative analysis showed excellent agreement between the model's outputs and semi‐automatic segmentations. ⁴¹ In a recent systematic review, investigators evaluated segmentation algorithms using performance metrics such as DSC. ⁴² For mandibular segmentation, the pooled DSC was 0.94 (95% CI: 0.91–0.98), indicating excellent accuracy. Deep‐learning architectures such as CNNs and U‐Nets consistently outperformed classical machine‐learning techniques. Most published work, however, used expert‐manual segmentation as the reference gold standard. ⁴² By contrast, the present study adopts a surface laser scan as the ground truth, a choice that may yield a more accurate reference and, in turn, more reliable comparative results. The variability in reported metrics highlights the need for uniform evaluation and reporting protocols.

This study has several limitations. A high‐resolution voxel size of 0.15 mm was used for segmentation, but larger voxels and other scanning parameters (mA, kVp, post‐processing) could affect accuracy, requiring further investigation. ⁴³ Only one CBCT scanner was used to limit variables, though scanner type and voxel range may influence segmentation accuracy. AI segmentation was performed with a single software, which may not represent all AI‐based tools. Due to the limitations of the AI software and its version, the remaining teeth and alveolar process were segmented as one entity for comparison in this study. Newer AI software can now segment the remaining teeth as their own entity, allowing for more refined comparisons. The study focused on the mandible, excluding the maxilla due to its complex anatomy, higher artifact incidence, and less dense cortical bone, which could complicate segmentation. ⁴⁴ Additionally, mandibles were scanned without soft tissue, simplifying segmentation compared to real human scans. Future research should assess how soft tissue presence affects segmentation accuracy across different methods.

CONCLUSIONS

AI‐based segmentations produced lower RMS values than global thresholding and showed no notable differences in most areas of the mandible when evaluated against manual segmentation. Nevertheless, there remains a statistically significant difference, with manual segmentation achieving lower RMS values in the alveolus, body of the mandible, and across the entire mandible. Despite these findings, the outcomes of this study highlight the capability of AI‐based segmentation protocols to serve as effective alternatives to both global thresholding and manual segmentation.

Hernandez AKT, Dutra V, Chu T‐MG, Yang C‐C, Lin W‐S. Trueness of artificial intelligence‐based, manual, and global thresholding segmentation protocols for human mandibles. J Prosthodont. 2025;34:939–946. 10.1111/jopr.70008