A novel artificial intelligence‐powered tool for automated root canal segmentation in single‐rooted teeth on cone‐beam computed tomography

Abstract

Aim

To develop and validate an artificial intelligence (AI)‐powered tool based on convolutional neural network (CNN) for automatic segmentation of root canals in single‐rooted teeth using cone‐beam computed tomography (CBCT).

Methodology

A total of 69 CBCT scans were retrospectively recruited from a hospital database and acquired from two devices with varying protocols. These scans were randomly assigned to the training (n = 31, 88 teeth), validation (n = 8, 15 teeth) and testing (n = 30, 120 teeth) sets. For the training and validation data sets, each CBCT scan was imported to the Virtual Patient Creator platform, where manual segmentation of root canals was performed by two operators, establishing the ground truth. Subsequently, the AI model was tested on 30 CBCT scans (120 teeth), and the AI‐generated three‐dimensional (3D) virtual models were exported in standard triangle language (STL) format. Importantly, the testing data set encompassed different types of single‐rooted teeth. An experienced operator evaluated the automated segmentation, and manual refinements were made to create refined 3D models (R‐AI). The AI and R‐AI models were compared for performance evaluation. Additionally, 30% of the testing sample was manually segmented at two different times to compare AI‐based and human segmentation methods. The time taken by each segmentation method to obtain 3D models was recorded in seconds(s) for further comparison.

Results

The AI‐driven tool demonstrated highly accurate segmentation of single‐rooted teeth (Dice similarity coefficient [DSC] ranging from 89% to 93%; 95% Hausdorff distance [HD] ranging from 0.10 to 0.13 mm), with no significant impact of tooth type on accuracy metrics (p > .05). The AI approach outperformed the manual method (p < .05), showing higher DSC and lower 95% HD values. In terms of time efficiency, manual segmentation required significantly more time (2262.4 ± 679.1 s) compared to R‐AI (94 ± 64.7 s) and AI (41.8 ± 12.2 s) methods (p < .05), representing a 54‐fold decrease.

Conclusions

The novel AI‐based tool exhibited highly accurate and time‐efficient performance in the automatic root canal segmentation on CBCT, surpassing the human performance.

INTRODUCTION

The accurate diagnosis and establishment of an appropriate treatment plan are crucial steps for the success of endodontic treatment (Pinto et al., 2023). Thus, the clinician must have a thorough understanding of the complex anatomical configuration of the root canals system to provide satisfactory clinical results for the patient (Mazzi‐Chaves et al., 2020). Cone‐beam computed tomography (CBCT) is an essential imaging method for evaluating the internal anatomy of teeth, significantly contributing to accurate diagnosis and treatment planning in endodontics. This technology provides detailed insights that enhance the understanding of the root canal system by clinicians, thereby supporting improved treatment outcomes.

In particular, CBCT scans enable the segmentation of anatomical structures to obtain three‐dimensional (3D) models, which are essential steps in establishing an effective digital workflow in dentistry (Fontenele et al., 2022, 2023; Lahoud et al., 2021; Shujaat et al., 2021). However, these processes are negatively impacted by poor image quality, partial volume effect and artefacts generated by high‐density materials, such as metallic objects and endodontic sealers with high radiopacity. Additionally, conventional segmentation techniques are time‐consuming and can be greatly influenced by human subjectivity (Elgarba et al., 2023; Fontenele et al., 2022, 2023; Lahoud et al., 2021; Shujaat et al., 2021).

Recent advances in computational technology have facilitated the integration of innovative tools guided by artificial intelligence (AI) to overcome limitations associated with classical approaches for segmenting dentomaxillofacial structures on CBCT scans (Elgarba et al., 2023; Fontenele et al., 2022, 2023; Lahoud et al., 2021). In this context, convolutional neural network (CNN), a robust type of deep learning algorithm, have been employed for digital image analysis (Aminoshariae et al., 2021; Fontenele et al., 2022, 2023). Essentially, CNNs consist of multiple layers of neural connections, inspired by the connectivity pattern of neurons in the human brain, and have proven to be an outstanding tool for automatically segmenting various anatomic structures on CBCT scans with high accuracy and time efficiency (Fontenele et al., 2023; Jindanil et al., 2023; Lahoud et al., 2022; Morgan et al., 2022; Nogueira‐Reis et al., 2023; Oliveira‐Santos et al., 2023; Preda et al., 2022; Shaheen et al., 2021; Verhelst et al., 2021).

CNN‐based automated root canal segmentation promises to transform digital workflows in endodontics. This method offers several benefits for clinical practice. It enables precise identification and segmentation of complex root canal anatomy, such as lateral canals, isthmuses and apical deltas, thereby improving the accuracy and success of treatment (Duan et al., 2021; Lin et al., 2021; Sherwood et al., 2021; Wang et al., 2023; Zhang et al., 2021). Moreover, it facilitates measurement of root canal volume changes over time, which is particularly advantageous in managing vital pulp treatments for traumatized teeth. By locating narrow root canals while minimizing unnecessary tooth structure removal, this approach reduces the risk of root perforation (Khanagar et al., 2023). Additionally, the use of accurate 3D models for preclinical training can significantly benefit dental students, enhancing their learning experience in root canal access (Aminoshariae et al., 2024; Khanagar et al., 2023). Finally, the application of CNN‐based segmentation in epidemiological research allows for the analysis of large data sets, facilitating in the study of anatomical variations in root canals across diverse populations (Khanagar et al., 2023). However, to the best of our knowledge, there is still no study in the literature on CNN‐based automatic segmentation of root canals including all groups of single‐rooted teeth.

Therefore, the aim of this study was to develop and validate an AI‐driven tool for automatic root canal segmentation of single‐rooted teeth on CBCT scans. The adopted null hypothesis was that there would be no difference in terms of accuracy and time efficiency between the AI‐based method and human performance.

MATERIALS AND METHODS

Ethical criteria

This study underwent review and approval by the local ethics committee before being conducted (reference number: S67798). Additionally, the investigation adhered strictly to the World Medical Association Declaration of Helsinki on medical research and followed the recommendations outlined in the STROBE guidelines. As all patient data were anonymized before analysis, obtaining informed consent was deemed unnecessary.

CBCT data collection

A total of 69 CBCT scans were retrieved from the Dentomaxillofacial Imaging Center database. These scans were initially acquired for different purposes not related to the present research, including endodontic treatment planning, orthognathic surgery and third molar extraction. The CBCT data set was acquired using NewTom VGi evo (Cefla, Imola, Italy) and 3D Accuitomo 170 (J Morita, Kyoto, Japan) devices, employing different acquisition parameters to ensure heterogeneous sampling, as described in Table 1.

TABLE 1.

Acquisition parameters of the cone‐beam computed tomography data set.

CBCT device	kVp	mA	FOV	Voxel size (mm)
NewTom VGI evo	110	3–20	8 × 8, 10 × 10, 12 × 8, 16 × 16, 24 × 19 cm	0.125–0.300
3D Accuitomo 170	90	5	8 × 8, 10 × 10, 14 × 10, 17 × 12 cm	0.125–0.250

Abbreviations: CBCT, cone‐beam computed tomography; FOV, field of view; kVp, kilovoltage peak; mA, tube current; mm, millimetres.

Inclusion criteria consisted of patients aged 18 years or older, as well as CBCT scans with a full permanent dentition and satisfactory image quality, characterized by adequate sharpness, contrast and noise levels, enabling accurate delineation of pulp chambers and root canals. Additionally, the CBCT scans needed to have a field of view (FOV) covering the maxillary and/or mandibular teeth.

The selected CBCT data set included a wide range of anatomical diversity within single‐rooted teeth, representing different clinical scenarios. The data set encompassed teeth with one or two canals, canals of varying widths from wide to narrow, and lengths ranging from short to long (18–26 mm). Both straight and curved root canal configurations were represented, reflecting the natural variation commonly seen in clinical practice. Scans with significant expression of high‐density material artefacts, such as multiple dental implants within the same arch or motion artefacts, were excluded from the study. This anatomical diversity is summarized in Table 2, demonstrating the sample heterogeneity of the sample in the current investigation.

TABLE 2.

Sample descriptions in the AI tool's test dataset by root canal type.

Types of root canals	Number of teeth (n)
Teeth with wide and straight root canals	66
Teeth with wide and curved root canals	25
Teeth with straight root canals and narrowing in at least one of the root thirds	18
Teeth with curved root canals and narrowing in at least one of the root thirds	6
Single‐rooted teeth with two root canals	5

Abbreviation: n, number of teeth.

The selected CBCT scans were exported in Digital Imaging and Communication in Medicine (DICOM) format and randomly distributed into three steps:

1. Training step (n = 31, 88 teeth): This involved training the AI‐driven tool using manual segmentation of root canals;

2. Validation step (n = 8, 15 teeth): Internal validation of the CNN model was performed, optimizing parameters to achieve an ideal model;

3. Test step (n = 30, 120 teeth): This step was dedicated to evaluating the performance of the AI‐driven tool by comparing the models obtained from automatic segmentation of the root canals with those from refined automated segmentation (R‐AI) performed by an expert.

The study design, along with the data set distribution, is presented in Figure 1. The data set is comprised of all types of single‐rooted teeth, encompassing upper central incisors, upper lateral incisors, upper canines, lower incisors, lower canines and lower premolars. Initially, the CBCT scans were individually imported to the online cloud‐based platform ‘Virtual Patient Creator’ (Relu, Leuven, Belgium). This software provides a range of tools for manual segmentation of dentomaxillofacial structures on CBCT scans. These tools, including brush, contour and interpolation, facilitated delineating the limits of the root canals and pulp chambers in the axial, coronal and sagittal reconstructions.

FIGURE 1.

Flowchart showing the overview of the study design including the data set used for the training, validation and testing of the CNN model for automated root canal segmentation in single‐root teeth. AI, artificial intelligence; CBCT, cone beam computed tomography; LC, lower canine; LI, lower incisor; LPSR, single‐rooted lower premolar; R‐AI, refined artificial intelligence; UCI, upper central incisor; ULI, upper lateral incisor; UC, upper canine.

Two operators (AOSJ and FSN), who had undergone prior training and calibration, independently performed the manual segmentation of the root canals and pulp chamber from the teeth included in the training and validation data sets. Before employing this segmentation data set for further AI‐driven tool development, the operators held a consensus meeting to review all manual segmentations and made refinements as needed. In cases of divergences, a third operator (RCF) was consulted to reach a consensus. Subsequently, the 3D models obtained after manual segmentation were exported in Standard Triangle Language (STL) format, serving as input for the training and validation of the AI networks.

CNN architecture

The AI segmentation pipeline was constructed with two sequential neural networks, utilizing the 3D U‐Net architecture, comprising four encoder and three decoder blocks. Each block featured two convolutions with a kernel size of 3 × 3 × 3, followed by rectified linear unit (ReLU) activation and group normalization with eight feature maps. Max pooling was applied using a 2 × 2 × 2 kernel with a stride of two, effectively reducing the resolution by half in all dimensions. Both networks operated as binary classifiers, assigning a label of 0 or 1. To ensure consistency across all scans, the CBCT images were standardized to a uniform voxel size, and downsampling was performed to a fixed size to overcome GPU memory limitations. These interconnected networks were essential for the automatic segmentation of endodontic structures, specifically the pulp chamber and root canals. The decision to adopt a two‐step approach was strategic, acknowledging the inherent limitations of CNNs when applied to CBCT scans with a large FOV (Fontenele et al., 2022, 2023; Preda et al., 2022).

The first neural network performed an initial detection of the structures of interest and propose 3D patches serving as the preliminary segmentation model. Subsequently, these patches were then transferred to the second network to refine the segmentation obtained in the first step, enabling automatic segmentation of endodontic structures at full resolution. These CNN models were implemented using PyTorch, and data augmentation techniques were employed to enrich the training data set, thereby enhancing the robustness of the developed CNN model. Techniques such as random cropping, scaling, rotation, mirroring and elastic deformation were utilized for this purpose.

Additionally, the parameters of the CNN model underwent optimization through the ADAM optimization algorithm, initializing with a learning rate of 1.25e⁻⁴, which was halved seven times over 300 epochs. This optimization process involved reducing the learning rate and implementing early stopping based on the validation set to prevent overfitting. A saturation point was identified, indicating that additional training data would lead to overfitting without improving the ability of the model to generalize. Notably, this study employed a heterogeneous data set throughout the AI‐driven tool development stages, as previously described. This diversity enabled the CNN model to effectively account for the anatomical complexities commonly encountered in root canal systems. Finally, the CNN model was implemented on the online cloud‐based AI platform ‘Virtual Patient Creator’. This platform is specialized in the automated segmentation of dentomaxillofacial structures on CBCT scans, including teeth (Fontenele et al., 2022; Shaheen et al., 2021), mandibular canal (Lahoud et al., 2022; Oliveira‐Santos et al., 2023), mandibular incisive canal (Jindanil et al., 2023) and maxillary alveolar bone (Fontenele et al., 2023).

Automated root canal segmentation—Testing step

Each CBCT scan, formatted in DICOM file format, was uploaded to the ‘Virtual Patient Creator’ platform. The platform automatically segmented the endodontic structures using a multi‐class approach, segmenting all root canals from the same scan simultaneously. Subsequently, the platform generated the individual 3D models of the pulp chamber and root canal for each single‐rooted tooth, exporting in STL format. Additionally, the platform automatically recorded the time required (in seconds) to generate the 3D models.

Refinement of the automated root canal segmentation—Testing step

All automated segmentations were examined by an experienced operator (RCF) to correct potential errors in AI segmentation, which could involve either undersegmentation and/or oversegmentation. This process was carried out using the aforementioned cloud‐based platform after a thorough examination of the CBCT scans in the axial, sagittal and coronal multiplanar reconstructions. It was determined that all segmentation maps required some level of correction, and these corrections were done using the brush tool (removing and adding voxels in the segmentation map) after utilizing the resliceable axes tool. The use of the resliceable axes tool allowed the operator to adjust the axial, sagittal and coronal planes of CBCT scans to become parallel and perpendicular to each other, considering the areas of interest of the root canal that required corrections. This alignment was critical for accurate visualization and assessment of the root canal morphology. It improved visualization of the full length and shape of the root canal, facilitating manual segmentation and refining the automated AI‐driven segmentation, particularly in cases with significant root curvature.

Following these adjustments, the platform generated a new R‐AI model for the root canal in STL format. Figure 2 shows the great diversity of anatomical variations observed in the testing data set. The time taken for performing the manual refinements was recorded using a digital stopwatch and added to the time required for the initial automatic segmentation provided by the platform.

FIGURE 2.

Three‐dimensional transparent models of teeth showcasing their internal anatomy and the variability within the testing data set. (a) Wide and straight root canal, (b) wide and curved root canal, (c) straight root canal with narrowing in at least one root third; (d) curved root canal with narrowing in at least one root third, (e) single‐rooted tooth with two root canals.

Validation metrics

The performance of the developed CNN model was assessed using a confusion matrix at the voxel level, comparing AI and R‐AI segmentation maps. From this comparison, the following variables were established:

1. False positive (FP): Voxels initially included in the automated segmentation but subsequently removed by the operator during the refinement of the root canal segmentation.

2. False negative (FN): Voxels initially not included in the automated segmentation but later added by the operator during the refinement of the root canal segmentation.

3. True positive (TP): Voxels that accurately belonged to root canals and were correctly segmented.

4. True negative (TN): Voxels that did not pertain to root canals and were not correctly excluded from the final segmentation map.

The values of these variables were used to calculate the following accuracy metrics, which were employed to assess the performance of the AI‐driven tool:

1. 95% Hausdorff Distance (HD): Signifies the 95th percentile of the maximum distance between a point on the AI segmentation map and its nearest point on the segmentation map obtained after refinement. This metric provides a measure of spatial accuracy by assessing the maximum distance between the predicted AI segmentation and the refined AI segmentation (R‐AI segmentation map).

   $$P_{95}\left(\min |\mathrm{ǀ}p–g\mathrm{ǀ}{|}_{2}u\min |\mathrm{ǀ}g–p\mathrm{ǀ}{|}_2\right)$$

   $$g\in Gp\in P$$
2. Intersection over union (IoU): Quantifies the extent of overlap between the AI segmentation map and the R‐AI segmentation map (R‐AI).

   $$\mathrm{IoU}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}$$
3. Dice similarity coefficient (DSC): Measures the level of intersection between the AI and R‐AI segmentation maps. This metric quantifies the overlap between the AI‐generated segmentation map and the segmentation map generated after refinements (R‐AI segmentation map), providing an indication of how similar they are.

   $$\mathrm{DSC}=\frac{2\times \mathrm{IoU}}{1+\mathrm{IoU}}$$
4. Precision: Indicates the ratio of accurately identified voxels to all voxels classified as belonging to the root canal by the AI tool.

   $$\text{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$
5. Recall: Signifies the ratio of voxels that belonged to the root canal and were correctly identified by the AI tool.

   $$\text{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$
6. Accuracy: Indicates the ratio of voxels that were correctly detected among all the observed voxels.

   $$\text{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}$$

A 95% HD value of 0 mm and values of 100% for the other metrics signify a perfect segmentation. Figure 3 illustrates two important accuracy metrics, 95% HD and DSC, by overlapping the contours of the AI‐predicted segmentation map (in red) and the refined AI segmentation map (R‐AI segmentation map, in green) on the CBCT sagittal reconstruction.

FIGURE 3.

Visual representation of the 95% Hausdorff distance (HD) and dice similarity coefficient (DSC). The contours of the AI‐predicted segmentation map (in red) and the refined AI segmentation map (R‐AI segmentation map, in green) are overlapped on the CBCT reconstruction to illustrate these metrics. The 95% HD measures spatial accuracy by evaluating the distance between the segmentation boundaries, whereas the DSC quantifies the overlap between the segmentation maps.

Comparison between manual and AI‐driven segmentation

Manual segmentation performed by a human served as the reference to evaluate the performance of automatic segmentation conducted by the developed CNN model. Thirty per cent of the teeth included in the testing step (n = 36) were randomly selected, ensuring the inclusion of different single‐rooted tooth types. An experienced operator (RCF) conducted manual segmentation of the endodontic structures using the ‘Virtual Patient Creator’ platform.

First, the specialist outlined the pulp chamber and root canal manually using the contour tool on the axial reconstructions of the CBCT scans. Next, the operator activated the resliceable axes tool on that platform, navigating through the sagittal and coronal reconstructions by adding or removing voxels from the segmentation map until achieving a suitable 3D model. This procedure was repeated at two‐time points, with a 30‐day interval between the segmentation sections, to assess the accuracy of manual segmentation. The accuracy was evaluated by comparing the segmentation maps generated from the first and second segmentation for each single‐rooted tooth, using the same accuracy metrics mentioned earlier. Afterwards, these results were compared to those obtained for AI segmentation within each accuracy metric. The time required to carry out the manual segmentation was recorded using a digital stopwatch.

Timing analysis

The average time required to obtain the segmentation maps using manual, AI and R‐AI approaches was investigated. For this task, the same sample used for comparing the manual and automatic approaches was used (i.e. 30% of the teeth used in the testing data set).

1. Manual method: The time spent to perform manual segmentation from importing the CBCT data set to generating the 3D model.

2. AI method: The online platform automatically provided the time spent to perform the automatic segmentation until obtaining the 3D model.

3. R‐AI method: The time required to perform manual refinements was added to the time taken to perform the AI segmentation method.

Statistical analysis

Statistical analysis was performed using SPSS software (version 24.0, IBM Corp., Armonk, NY). Descriptive statistics of the accuracy metrics and time analysis data was carried out using mean and standard deviation (SD) values.

To assess the normality of the data, the Shapiro–Wilk test was utilized, considering the residual values. Since the quantitative parameters showed a normal distribution (p > .05), parametric statistical tests were employed depending on the number of groups compared. The mean values of each accuracy metric were compared using a one‐way analysis of variance (anova), followed by the Tukey post hoc test, with the type of tooth as the study factor. Additionally, a paired t‐test was employed to compare the performance of manual segmentation and AI segmentation for each accuracy metric. For time analysis, the time spent to generate the 3D root canal models using different segmentation methods (manual, AI and R‐AI) was compared using one‐way anova with Tukey post hoc test. For each statistical test, the power analysis was performed as follows: for anova, it considered the minimum difference between groups, the standard deviation within each group and the number of teeth in each group. For the paired t‐test, it considered the mean difference between paired observations, their standard deviation and the number of teeth for each accuracy metric. Based on these parameters, a statistical power ranging from 75% to 99% was achieved. Throughout all conducted analyses, the significance level was established at 0.05 (α = 5%).

RESULTS

Table 3 presents the AI‐driven tool's performance in generating 3D root canal models within each tooth group based on accuracy metrics. The AI‐driven tool demonstrated excellent performance, yielding high values of IoU (ranging from 80% ± 14 to 86% ± 7), DSC (ranging from 89% ± 6 to 93% ± 4), recall (ranging from 84% ± 11 to 93% ± 5), precision (ranging from 90% ± 5 to 94% ± 4) and accuracy (ranging from 98% ± 0.8 to 99% ± 0.4). Furthermore, the tool exhibited low values of 95% HD ranging from 0.10 ± 0.03 mm to 0.13 ± 0.06 mm. These results represent accurate automatic segmentation of root canals with an ideal superposition with the AI and R‐AI 3D models, indicating that the AI required only minor refinements (Figure 4).

TABLE 3.

Performance of the AI‐driven tool for automated root canal segmentation according to teeth group based on accuracy metrics.

Teeth group	IoU (%)	DSC (%)	Recall (%)	Precision (%)	Accuracy (%)	95% HD (mm)
	Mean (SD)	Mean (SD)	Mean (SD)	Mean (SD)	Mean (SD)	Mean (SD)
Upper central incisors (n = 20)	83 (8)	90 (5)	88 (7) AB	94 (4)	99 (0.4)	0.13 (0.06)
Upper lateral incisors (n = 20)	80 (14)	89 (8)	84 (11) B	92 (6)	99 (0.6)	0.12 (0.04)
Upper canines (n = 20)	81 (9)	89 (6)	88 (6) AB	90 (6)	98 (0.8)	0.10 (0.03)
Lower incisors (n = 20)	81 (9)	89 (5)	88 (7) AB	90 (5)	99 (0.6)	0.12 (0.04)
Lower canines (n = 20)	86 (6)	93 (4)	93 (4) A	92 (4)	99 (0.9)	0.11 (0.08)
Lower premolars (n = 20)	86 (7)	93 (4)	93 (5) A	92 (4)	99 (0.5)	0.12 (0.04)
p value	.11	.09	<.001	.16	.31	.32

Note: Statistical power analysis showed a range from 75% (for the 95% HD metric) to 98% (for the recall metric). Distinct uppercase letters indicate a statistically significant difference between the groups of teeth (p < .05).

Abbreviations: DSC, dice similarity coefficient; HD, Hausdorff distance; IoU, intersection over union; n, sample size; SD, standard deviation.

FIGURE 4.

Comparison of AI and R‐AI three‐dimensional models for different types of single‐rooted teeth, employing colour mapping in frontal and lateral views. The areas coloured in red and yellow indicate substantial differences, highlighting discrepancies between the AI and R‐AI 3D models.

The accuracy metrics were not significantly affected by the tooth type (p > .05), except for the recall metric (p < .05). The highest recall values were observed for the lower canines (93% ± 4) and premolars (93% ± 5), whereas the lowest value was recorded for the upper lateral incisors (84% ± 11; Table 3).

Table 4 shows the comparison between manual and AI‐driven methods for performing the root canal segmentation. The automated approach exhibited superior performance compared to the manual segmentation method (p < .05; Figure 5). This superiority was evidenced by lower 95% HD values (0.13 ± 0.02 mm) and higher values of IoU (84% ± 2), DSC (91% ± 4), precision (94% ± 4) and accuracy (99% ± 0.6). However, there were similar recall values between the two segmentation methods (p > .05). Figure 6 illustrates the types of errors (under and/or over segmentation) found in 3D models obtained by automated segmentation of root canals.

TABLE 4.

Comparison between the manual and AI segmentation methods for single‐rooted canal segmentation.

Metrics	Manual	AI	p‐value
	Mean (SD) (n = 54)	Mean (SD) (n = 54)
IoU (%)*	74 (10)	84 (2)	.003
DSC (%)*	85 (7)	91 (4)	.004
Recall (%)	87 (11)	89 (7)	.50
Precision (%)*	86 (10)	94 (4)	.01
Accuracy (%)*	98 (0.02)	99 (0.6)	.02
95% HD (mm)*	0.24 (0.06)	0.13 (0.02)	<.001

Note: Statistical power analysis showed a range from 78% (for the recall metric) to 93% (for the 95% HD metric). Asterisk (*) indicates a statistically significant difference between the segmentation methods within each accuracy metric (p < .05).

Abbreviations: AI, artificial intelligence; DSC, dice similarity coefficient; HD, Hausdorff distance; IoU, intersection over union; n, sample size; SD, standard deviation.

FIGURE 5.

Three‐dimensional (3D) models of an upper canine root canal segmentation employing colour mapping, comparing (a) the manual method and (b) the AI method. The areas highlighted in red and yellow indicate regions with significant differences, representing discrepancies between the first and second manual segmentation for the manual method and between the AI and R‐AI 3D models for the AI method.

FIGURE 6.

Three‐dimensional models illustrating various types of errors found in AI‐driven segmentation. The white three‐dimensional (3D) models represent AI predictions, whereas the red ones represent the 3D model after the refinements made by the expert. Errors are indicated by arrows. (a), Undersegmentation in the apical third of the root canal for an upper left canine tooth; (b), Oversegmentation in the pulp chamber and apical third of a first lower left premolar tooth; and (c) under‐ and oversegmentation in the pulp chamber and root canal of an upper left central incisor tooth.

Regarding the time analysis (Figure 7), the manual method showed the lengthiest consumption of time (2262.4 ± 679.1 s) for conducting the segmentation compared to the AI‐driven approach (41.8 ± 12.2 s) and R‐AI (94 ± 64.7 s), which did not differ from each other (p < .05).

FIGURE 7.

Timing analysis according to the segmentation method. Distinct uppercase letters indicate a statistically significant difference among the segmentation methods (p < .05). Statistical power analysis of 99%. AI, artificial intelligence; R‐AI, refined artificial intelligence.

DISCUSSION

The 3D segmentation of the pulp chamber and root canal is a crucial step in the endodontic digital workflow. However, several limitations are commonly associated with classical segmentation methods (Lin et al., 2021; Wang et al., 2023). To address these challenges, the present investigation aimed to develop and validate a unique AI‐driven tool for the automatic segmentation of endodontic structures (i.e. pulp chamber and root canal) in single‐rooted teeth on CBCT scans. The current findings revealed that the CNN model could provide highly accurate automated segmentation of these fine anatomical structures with minimal working time, demonstrating significant differences compared to manual segmentation performed by a human. Thereby, the null hypothesis established was rejected.

The performance of the AI‐based algorithm was evaluated by using accuracy metrics. The high IoU values (ranging from 80% ± 14 to 86% ± 7), DSC (ranging from 89% ± 6 to 93% ± 4) and low values of 95% HD (ranging from 0.10 ± 0.03 mm to 0.13 ± 0.06 mm) evidence the excellent performance attributed to the AI‐driven tool developed in this study for obtaining accurate 3D models of root canals in single‐rooted teeth. To the best of our knowledge, there are few previous investigations that have explored the effectiveness of AI‐driven tools for automatically segmenting root canals on CBCT scans. Duan et al. (2021) reported adequate performance for a CNN model in segmenting teeth and pulp chambers of single and multi‐rooted teeth. However, their image data set was acquired from only one CBCT device using similar acquisition parameters, making comparison challenging with the current investigation. Lin et al. (2021) reported acceptable performance (DSC = 87.49%, 95% HD = 1.99 mm) for AI‐based automatic segmentation of root canals in single‐rooted premolars. However, the authors stated that the developed CNN model needed optimization, especially for accurate segmentation of root canals in the apical third. Still, these results should be carefully interpreted, as only single‐rooted premolars were included in the sample. Another study revealed that it was not possible to perform automatic segmentation of narrow root canals in mandibular molars using a 3D U‐Net neural network on CBCT scans (Lin et al., 2022). On the other hand, Zhang et al. (2021) attributed acceptable results (DSC = 0.95) to an AI algorithm for automated segmentation of root canals. However, no information related to the type of tooth included in the data set was reported. Recently, Wang et al. (2023) proposed a CNN model for the automatic segmentation of single‐rooted teeth and their root canals on CBCT scans. However, due to the limited data set used during the training (5 CBCT scans, 21 teeth), validation (2 CBCT scans, 8 teeth) and testing (7 CBCT scans, 27 teeth) of the AI networks and also due to the unknown CBCT device used, generalizations and comparisons with the findings of the present investigation are difficult.

The current findings evidence the excellent performance of the AI algorithm in obtaining highly accurate 3D models of root canals across all types of single‐rooted teeth, including incisors, canines and premolars. This is in contrast to the results of previous aforementioned studies. However, when evaluating the recall metric, slightly lower values were found for the root canals of the upper lateral incisors (84% ± 11) compared to canines (93% ± 4) and lower premolars (93% ± 5). Nevertheless, this statistical difference was relatively minor and might not have a significant impact on the tool's performance in clinical practice. These findings could be attributed to the complex root anatomy of the upper lateral incisors, characterized by the presence of a long and narrow root canal, especially in the apical third, along with a pronounced radicular curvature in the palato‐distal direction (Wang et al., 2022).

Automated segmentation of root canals driven by AI exhibited notably superior performance (IoU = 84% ± 2, DSC = 91% ± 4, precision = 94% ± 4, accuracy = 99% ± 0.6, and 95% HD = 0.13 ± 0.02 mm) compared to manual segmentation performed by human. However, no statistical differences were found when considering the recall metric between both segmentation methods. These results further reinforce the premise that the anatomical complexity of the upper lateral incisor root canals presents an intrinsic challenge for segmentation cannot be attributed as a limitation to the AI‐driven tool. Consequently, these findings highlight the clinically acceptable results attributed to the CNN model in this study for obtaining accurate 3D models of root canals, even in teeth with challenging root anatomy.

Previous studies have demonstrated that precise segmentation of fine head and neck anatomical structures required a large training sample size when employing deep learning algorithms (Fang et al., 2021). In the present study, a data set of 69 CBCT scans, containing 103 teeth, was employed to train and validate the CNN model. Several approaches were implemented to mitigate the challenge of overfitting. To achieve this, the approaches included ensuring data set heterogeneity, partitioning the data for continuous monitoring of model performance, employing data augmentation techniques, such as rotations, flips and zooms, and incorporating early stopping as a regularization method. While these measures were essential for optimizing the training process and ensuring an adequate data set size, the success of the approach cannot be attributed solely to them. Prior training of the tested online cloud‐based AI platform for tooth segmentation, which utilized a substantial sample of 175 CBCT scans encompassing 500 teeth, as reported by Shaheen et al. (2021) and Fontenele et al. (2022), played a pivotal role in the development of the current model. This prior knowledge substantially improved the feature extraction capabilities of the AI network. By applying insights from earlier tooth segmentation training, particularly those related to the voxels representing tooth structure, the learning process for segmenting the pulp chamber and root canals was significantly enhanced. This accelerated learning, grounded in previous training, enabled the study to reach a saturation point in the learning curve with a smaller sample size, thereby optimizing the overall results achieved.

The time consumed to perform segmentation is a key point in the digital workflow (Fontenele et al., 2022, 2023). However, existing literature lacks information regarding the time required for root canal segmentation. This investigation pioneers the evaluation and comparison of the time required for root canal segmentation using manual, AI and R‐AI methods. The AI segmentation showed the fastest performance (41.8 ± 12.2 s), being 54 times faster than manual segmentation performed by humans (2.262.4 ± 679.1 s). Additionally, the current findings revealed that the R‐AI approach exhibited low working time (94 ± 64.7 s), similar to AI‐driven segmentation. These results highlight the high accuracy of the 3D models obtained by the AI method and clearly suggest that in cases where adjustments to the segmentation map are needed, such refinement will be minimal and not require significant working time.

The data sets used in this study for training, validation and testing the CNN model encompassed a diverse range of single‐rooted teeth, providing a comprehensive representation of clinical scenarios. These teeth varied in the number of root canals, ranging from one to two, and exhibited a wide range of widths, lengths and curvatures, particularly in the apical third. This anatomical diversity within the sample set was essential for adapting the CNN model to the numerous anatomical variations of root canals, ultimately resulting in accurate automatic segmentation outcomes, as shown in Table 3. The AI‐driven tool initially trained on simpler cases, such as teeth with a single, wide and straight root canal. This approach, which prioritized less complex root canal configurations, facilitated a progressive learning curve for the CNN model, similar to a curriculum‐based learning strategy (Shan et al., 2021). As a result, the tool demonstrated improved performance in delivering highly precise automatic segmentations, even for root canals with complex anatomical morphology.

To further assess the robustness of the CNN model, future studies should focus exclusively on testing its performance on challenging cases, particularly teeth with highly complex root canal configurations. This approach will allow researchers to effectively evaluate the model's ability to accurately segment challenging anatomical features. In addition, incorporating regression model analyses or subgroup analyses into these studies will provide valuable insights into how variations in root canal number, length, diameter and the presence of calcifications affect the performance of the CNN model. Furthermore, investigating the performance of the model across diverse patient populations and imaging settings is crucial to ensure its reliability and applicability in various clinical scenarios.

Periapical radiography serves as the primary imaging modality for diagnosis and planning in Endodontics (Hilmi et al., 2023). However, complex clinical cases may necessitate a CBCT scan to aid in establishing a more accurate diagnosis and precise treatment plan (Pinto et al., 2023). In addition, requesting a CBCT scan is a mandatory step in the digital workflow for guided endodontic access planning in challenging cases, such as calcified root canals (Chaves et al., 2023). Nevertheless, it is essential to note that a CBCT scan requires a higher radiation dose than a periapical radiograph (Pauwels, 2015). Hence, clinicians must bear in mind that a CBCT device should provide an image with effective diagnostic quality while requiring the lowest radiation dose possible, following the ALADAIP principle (as low as diagnostically acceptable being indication‐oriented and patient‐specific; Oenning et al., 2018).

The image data set for this study was acquired from two CBCT devices, which may limit the generalization of the current results to different devices. Thus, future investigations should focus on optimizing the AI‐driven tool using data from various CBCT devices. A portion of the sample included in the present study comprises CBCT scans with large FOV, which could be considered a potential limitation. In fact, visualizing small anatomical structures like root canals is challenging in images acquired with a large FOV due to lower spatial resolution (Pinto et al., 2023). However, to achieve greater generalization of AI networks, it is necessary to include CBCT scans with larger FOVs (Fontenele et al., 2022, 2023). This choice has shown no negative impact, as evidenced by the overall excellent performance of the CNN model in this study, which achieved precise and fast automated segmentation of root canals. However, it is important to note that the CBCT scans used in this investigation were sourced from a single hospital database, and the demographic characteristics of the patients remain unknown. Therefore, conducting multi‐center studies is highly recommended, as they can provide a broader and more diverse data set for training and testing AI algorithms. Additionally, CBCT scans with a high artefact expression from materials with high density were excluded; consequently, our results cannot not be extrapolated to this clinical scenario. However, it is important to emphasize that some CBCT scans included were not entirely artefact‐free, considering the presence of some teeth with coronal and root fillings (e.g. metallic restorations, gutta‐percha, sealer). Thus, further research is needed to test the CNN model's performance within a clinical setting with a higher expression of artefacts.

The groundbreaking results of this study reveal the remarkable high accuracy and time efficiency achieved by a unique AI‐based tool for automatically segmenting single‐rooted teeth on CBCT scans, surpassing human performance. These outcomes not only suggest the need for further research in automating segmentation for bi‐rooted premolars and molars but also indicate the transformative potential of the developed AI‐driven tool in revolutionizing the digital workflow within endodontics. Particularly, beneficial in cases requiring guided endodontic access, this advancement may facilitate fast and precise localization of narrow root canals, thereby reducing the risk of root perforations and preserving the dentin structure. Such technologies ensure greater consistency in endodontic diagnosis and safer treatment planning. Moreover, the tool's potential application in the pre‐clinical training of undergraduate and postgraduate students, specifically in accessing root canals, stands as a notable advantage. Lastly, clinicians can leverage the highly accurate 3D models of root canals to improve communication with patients, ultimately enhancing treatment acceptance.

CONCLUSION

The unique AI‐based tool demonstrated an exceptional accuracy and fast performance in segmenting pulp chambers and root canals in single‐rooted teeth on CBCT images, surpassing the human performance. These outcomes hold significant for reshaping the digital workflow in Endodontics, fostering greater consistency and predictability in treatments with minimal loss of tooth hard tissue structures.

AUTHOR CONTRIBUTIONS

Airton Oliveira Santos‐Junior: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Software, Validation, Visualization, Roles/Writing—original draft, Writing—review and editing. Rocharles Cavalcante Fontenele: Conceptualization, Data curation, Formal analysis, Project administration, Investigation, Methodology, Software, Validation, Visualization, Roles/Writing—original draft, Writing—review and editing. Frederico Sampaio Neves: Conceptualization, Data curation, Investigation, Methodology, Software, Validation, Visualization, Roles/Writing—original draft. Mário Tanomaru‐Filho: Conceptualization, Funding acquisition, Resources, Supervision, Visualization, Roles/Writing—original draft, Writing—review and editing. Reinhilde Jacobs: Conceptualization, Funding acquisition, Resources, Supervision, Methodology, Visualization, Roles/Writing—original draft, Writing—review and editing.

FUNDING INFORMATION

The study was supported by the São Paulo State Research Foundation (FAPESP), Brasil (grant nos. 2020/11012‐3, 2021/11496‐3 and 2022/13774‐3).

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interst.

ETHICS STATEMENT

The present study was approved by the Local Ethical Review Board under protocol number S67798.

Santos‐Junior, A.O. , Fontenele, R.C. , Neves, F.S. , Tanomaru‐Filho, M. & Jacobs, R. (2025) A novel artificial intelligence‐powered tool for automated root canal segmentation in single‐rooted teeth on cone‐beam computed tomography. International Endodontic Journal, 58, 658–671. Available from: 10.1111/iej.14200

DATA AVAILABILITY STATEMENT

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