Open Access is an initiative that aims to make scientific research freely available to all. To date our community has made over 100 million downloads. It’s based on principles of collaboration, unobstructed discovery, and, most importantly, scientific progression. As PhD students, we found it difficult to access the research we needed, so we decided to create a new Open Access publisher that levels the playing field for scientists across the world. How? By making research easy to access, and puts the academic needs of the researchers before the business interests of publishers.
We are a community of more than 103,000 authors and editors from 3,291 institutions spanning 160 countries, including Nobel Prize winners and some of the world’s most-cited researchers. Publishing on IntechOpen allows authors to earn citations and find new collaborators, meaning more people see your work not only from your own field of study, but from other related fields too.
Accurate delineation of dental structures is fundamental to modern dentistry, and specifically endodontics, directly impacting diagnostic reliability and treatment outcomes. This chapter addresses the challenge of precise image segmentation, focusing on the development and validation of workflows optimized for teeth and root canals. While deep learning models, including 3D U-Net and Generative Adversarial Networks (GAN), have shown considerable promise for automated segmentation, their widespread adoption is hindered by dependencies on large, tediously annotated datasets and concerns regarding cross-population generalizability. To circumvent these limitations, this work introduces two semi-automatic pipelines, namely Grow From Seeds (GFS) and Watershed (WS), implemented within the 3D Slicer platform. These methods were specifically designed to reduce the manual effort required for generating reliable ground truth annotations. A comparative dice score analysis was conducted comparing 3D models of different workflows, including manual, GFS, WS, and a contemporary automated segmentation tool, against a manually segmented ground truth of a maxillary second molar. Dice scores reached 0.978 in some cases with GFS and 0.968 with WS. Further statistical testing demonstrated that models produced by the semi-automatic workflows were statistically indistinguishable from manual segmentations, whereas the fully automated approach deviated significantly. Furthermore, we present adaptations of these workflows capable of differentiating between dentin and enamel through separate segmentation layers, enhancing their utility for biomechanical analyses. The findings underscore the efficacy of semi-automatic strategies in balancing accuracy with efficiency, offering a practical pathway for generating high-quality training data in relatively short times, facilitating deep learning models training in endodontics segmentation.
Department of Mechanical Engineering, School of Sciences and Engineering, The American University in Cairo, New Cairo, Egypt
Yousef Abo El Ela
Department of Mechanical Engineering, School of Sciences and Engineering, The American University in Cairo, New Cairo, Egypt
*Address all correspondence to: mobadran@aucegypt.edu
1. Introduction
Image segmentation has been a beneficial approach for dealing with Digital Imaging and Communications in Medicine (DICOM) data with regards to medical, and more specifically dental application [1]. DICOM can be presented in Image segmentation refers to the process of partitioning an image into non-overlapping regions that are homogeneous with respect to intensity, texture, or other features, so that objects of interest can be isolated and quantified. In dental imaging, segmentation is a key prerequisite for tasks such as 3D reconstruction of teeth and jaws, caries and periapical lesion analysis, implant planning, orthodontic assessment, and endodontics treatment planning. This section introduces the working principles of some image segmentation techniques and their integration and application in the field of dentistry.
1.1 Image segmentation types
Segmentation algorithms are commonly grouped into threshold-based, region-based, edge-based, clustering-based, model-based, graph-based, interactive GrowCut‑based, and learning-based methods, each with characteristic advantages, limitations, and dental applications [2, 3]. The following subsections only discusses threshold-based, region-based, GrowCut-based, and learning-based techniques as per their relevance to the work presented in the Chapter.
1.1.1 Threshold-based segmentation
Thresholding is assigning pixels of an image to separate segmentation layers by comparing their intensity values with a threshold derived from the gray-level histogram as shown in Eq. (1).
V(x,y)={Ls,ifv(x,y)≥TLb,ifv(x,y)<TE1
Where V is the output value of each pixel given by its x and y coordinates, Ls is the value assigned to the subject segmentation layer, Lb is the value assigned to the background layer, v is the intensity value of an input pixel, and T is the set threshold value of comparison.
This simple algorithm take any grayscale image as an input and outputs a bitmap image of a segmented image. Figure 1 shows this process applied to a maxillary central incisor.
Figure 1.
An example of thresholding on a maxillary central incisor where (a) is the input image of various intensity levels, and (b) is the output bitmap after selecting only the pixels of the threshold value or brighter.
Thresholding is computationally efficient and easy to implement, however global thresholds perform poorly in the presence of noise, non-uniform illumination, or overlapping intensity distributions between tissues, which are common in clinical CBCT and radiographs [2]. This is why local thresholding techniques are a better match for medical applications despite requiring more complex algorithms.
For dental application, Cone Beam Computed Tomography (CBCT) and micro-Computed Tomography (micro-CT) are the most commonly used types of datasets for 3D dental segmentation [4–6]. On 2D intraoral and panoramic radiographs, thresholding is used both as a standalone method and as a preprocessing step to highlight high-density regions or potential lesions before more advanced segmentation or classification [7]. On CBCT data, threshold-based methods have been used to segment structures with relatively distinct attenuation values, such as enamel and pulp cavities, especially when combined with local contrast enhancement. Workflows have combined marker-controlled watershed (WS) for tooth and pulp partitioning with global thresholding for enamel extraction, exploiting the higher intensity of enamel relative to dentin and surrounding bone to detect enamel regions relevant for caries and structural damage assessment [8].
Overall, threshold-based segmentation in dentistry is most effective in high-contrast scenarios, such as enamel in CBCT, or as a first stage in multi-step algorithms [9].
1.1.2 Region-based segmentation
Region-based methods segment images by grouping pixels into regions that are internally homogeneous according to intensity, texture, or statistical criteria, often starting from seed points provided by the user and expanding via region growing or splitting and merging. WS segmentation interprets the gray-level image as a topographic relief and identifies catchment basins separated by ridges, with marker-controlled variants using predefined internal and external markers to mitigate over-segmentation [10]. Because region-based approaches exploit spatial context, they can better handle gradual intensity transitions than pure thresholding but are sensitive to noise and the selection of seeds or markers [9].
Region‑based approaches, particularly WS and region growing, are widely used for tooth segmentation in CBCT and micro‑CT, where each tooth typically forms a coherent 3D region within surrounding bone. Galibourg et al. evaluated automatic WS‑based segmentation of individual teeth on micro‑CT and CBCT data and showed that for high‑resolution micro‑CT, automatic segmentation produced tooth volumes comparable to semi‑automatic ground truth, while CBCT segmentations preserved overall tooth form but required manual refinement for precise volumetry [11].
1.1.3 Growcut‑based segmentation
GrowCut is an interactive, multi‑label segmentation algorithm based on cellular automata where user‑provided seed pixels with labels for each segmentation layer act as living cells that compete to label their neighbors through an iterative process [12, 13]. Each pixel is initially assigned a strength value based on whether it was user‑labeled (strength = 1.0) or unlabeled (strength = 0). In each iteration, every labeled pixel attacks its 3 × 3 neighborhood, attempting to overwrite neighbor labels based on a feature‑dependent attack strength computed from intensity similarity. An example of this is shown in Figure 2.
Figure 2.
The initial phase of an example showing unlabeled cells in black, object-labeled cells in green, and background-labeled cells in Orange.
The flow of spread is as follows: each user-labelled cell is given a strength of 1, and each unlabeled cell is given a strength is 0, regardless of their labels. For each iteration, each labeled cell attacks its 8 neighbors with an attack strength that is given by Eq. (2).
S=f(ΔI)×SaE2
Where S is the attack strength, ΔI is the difference of intensity between the attacker and defender cells, and Sa is the current strength of the attacking cell. The intensity difference function is shown in Eq. (3).
f(ΔI)=1−|Ia−Id|ImaxE3
Where Ia is the attacker cell’s intensity, Id is the defender cell’s intensity, and Imax is 255 since a gray-level histogram spans from 0 (darkest) to 255 (brightest). If S>Sd where Sd is the defender cell’s strength, the defender cells adopts the label of the attacker cell and changes its strength to the attack strength S. This process continues until no further changes occur, converging to a stable segmentation where labels propagate through regions of similar appearance while respecting intensity boundaries [12]. The example shown in Figure 2 is shown again in Figure 3 after the first iteration. A notable aspect of Cell 4’s segmentation outcome is its adoption of the object label despite the substantial intensity disparity (ΔI = 178) with its attacking Cell 5. This occurs because Cell 4 receives no competing attacks during the initial iteration. In subsequent iterations, however, Cell 4 becomes vulnerable to attacks from its newly labeled neighbors, particularly Cell 2 which shares far greater intensity similarity (ΔI = 2). This competitive dynamic results in Cell 4’s relabeling to background.
Figure 3.
The example shown in Figure 2 after the first iteration of the growcut algorithm.
Fast GrowCut is an accelerated variant of the original GrowCut algorithm that reformulates the cellular automaton-based label propagation as a clustering problem in feature space (typically intensity ± spatial position), solved via an efficient approximate method [14]. Rather than iterating through all pixels repeatedly until global convergence as in standard GrowCut, Fast GrowCut constructs a graph where nodes represent pixels, edges encode feature similarity, and an adapted Dijkstra’s shortest-path algorithm computes approximate cluster assignments from user seeds in a single pass. This reduces initial segmentation time by 8–30× and manual user refinement by up to 128× while retaining ≥ 97% of standard GrowCut’s accuracy across tested medical datasets [14]. Grow from seeds (GFS) is the built‑in Segment Editor tool in the software 3D Slicer that implements the Fast GrowCut algorithm [15, 16].
1.1.4 Learning-based segmentation
Learning‑based segmentation trains statistical or machine‑learning models to map image features to pixel‑wise labels, with deep convolutional neural networks (CNNs) and fully convolutional networks (FCNs) now dominating medical image segmentation due to their ability to learn hierarchical features from large datasets. U‑Net, introduced by Ronneberger et al. [17], is a widely used FCN architecture with an encoder–decoder “U” shape and skip connections that propagate fine spatial detail, enabling accurate segmentation with relatively few annotated biomedical images and achieving leading performance on several biomedical challenges. Recent studies of medical image segmentation quantifiably demonstrate that U‑Net and its 2D, 2.5D, and 3D variants [18], Generative Adversarial Network (GAN)‑based models form the backbone of state‑of‑the‑art segmentation across modalities in the dental field [19–24].
Learning-based approaches are still required to be assessed by a human expert to ensure compliance with medical standards to avoid any ethical problems when it comes to incorrect treatment planning [25, 26]. While the automated workflows and ease-of-use of such methodologies are a significant advantage once their models have been fully trained, achieving proper training of such models are a difficult task to achieve due to their high dependence on the quantity and quality of annotated ground truth datasets. In obtaining ground truth, significant time and manual effort take place, which in turn can lead to variability due to the operator’s subjectivity when labeling [27]. Furthermore, generalizing pretrained deep learning models can pose a threat to the model’s efficiency even those acquired from identical anatomical regions or the same scanner manufacturer [28]. Consequently, ongoing access to high-quality ground truth annotations proves essential for performance optimization as they constitutes the cornerstone for both training and rigorous validation of deep learning architectures in CBCT tooth segmentation.
Semi-automatic approaches balance the accuracy and precision of manual segmentation while reducing the labor and time required to achieve these accurate results. A few attempts along this direction has been made. For instance, Morell et al. [29] introduce a workflow that treats the root canal as a distinct segmentation layer. The process starts by loading CBCT dataset into ITK-SNAP software and modifying the histogram to enhance tissue visibility. The region of interest then magnified via cropping. The “Snake” function is employed to restrict segmentation to the target teeth. The “Threshold” segmentation mode is activated, with the upper limit set to maximum and the lower boundary calibrated for each patient’s case. Seeds are subsequently placed along the tooth’s long axis near the pulp canal and chamber, enabling the segmentation to propagate and automatically incorporate hard dental tissues. Following automatic segmentation, manual adjustments are required to distinguish structures with similar radiodensity, such as bone and cementum. This is done by modifying label opacity and using the “Paintbrush” to add or subtract regions across the three anatomical planes. The “Scalpel” is then applied to separate individual teeth into distinct labels, followed by final manual refinements, particularly in areas of tooth contact. The software subsequently computes the volume of each segmented tooth label in cubic millimeters. The study acknowledges several CBCT imaging challenges and the specific acquisition parameters used. Issues reported that may affect segmentation accuracy include field of view, voxel size, partial-volume effects, surrounding artifacts, scattered X-rays, subjective user decisions, proximity of adjacent anatomical structures, and patient motion during scanning. Consequently, the correctness of the results may exhibit considerable variability. Another limitation is the time required for the procedure; the manual refinement step is quite labor-intensive, making it impractical for busy clinicians. Furthermore, according to the authors’ own performance metrics, accuracy ranges from “good to excellent,” which may be quantitatively insufficient for applications that demand high precision.
Verykokou et al. [30] present another semi-automatic segmentation methodology that integrates thresholding, manual sample specification, and the GFS method within 3D Slicer (version 4.11.20210226). The process begins by cropping the region of interest from the CBCT dataset, followed by defining three segments: teeth, alveolar bone, and “other.” For each segment, seeds are manually placed on a subset of CBCT images, guided by data-dependent thresholds that serve as masks. The GFS method is then initiated, and the results undergo visual inspection and iterative correction until a satisfactory segmentation is achieved. The segmented teeth and alveolar bone are subsequently exported as 3D STL models, which are refined in Geomagic Wrap 2017 for point cloud editing, mesh generation, and optional merging into a single model. Limitations of this study include neglecting the root canal and modeling the tooth as a solid entity. This is an acceptable shortcut given the authors’ goal of designing 3D scaffolds for periodontal regeneration, however, for endodontics-oriented segmentation objectives, this workflow is not entirely sufficient. The authors also recommend micro-CT scanning to enhance accuracy over CBCT and encourage validation of their method on additional datasets.
The main problem of the mentioned work is the lack of attention to endodontics-oriented applications such as endodontics-based VR simulators which need to make a clear distinction between tooth structures and pulp chambers. This composition is crucial for accurate and effective training. The proposed methods presented in this work mainly targets the pulp chamber and root canals recognition so that the output 3D models are acceptable for endodontics-related adaptations. Furthermore, semi-automated methods are a viable solution to the issue of training data availability discussed in Section 1.1.4. Good quality, annotated ground truth as training data is the backbone of any well-performing deep learning-based model for image segmentation; therefore, streamlining the training process of such models can be achieved by offering large, labeled datasets using the proposed methods.
This section introduces two semi-automatic workflows specifically designed with endodontics in mind. Thorough documentation of these methods are detailed here for ease of repeatability. Also presented is a quantitative comparison between the output of these methods and those of an automated state-of-the-art workflow, DentalSegmentator by Dot et al. [22], against the output of completely manual segmentation, which is considered a gold standard and a benchmark for correct segmentation [31]. Only the results for a maxillary second molar from 3 different datasets are shown here as molars present a generalized test for difficult-to-segment multi-rooted teeth, however the full results of 3 datasets and 2 teeth anatomies are available in a previous work [32]. Figure 4 shows the unaltered dataset.
Figure 4.
The original CBCT scans of dataset 1 showing the selected tooth from: (a) the axial plane, (b) the coronal plane, and (c) the sagittal plane.
The comparison presented here is a Kruskal-Wallis test to determine if there is any significant deviations between the control group, which is the manual segmentation, the proposed semi-automated methods (GFS and WS), and the automated workflow (DentalSegmentator). A Nemenyi test is then carried out to determine which of the test groups is the deviated one.
3.1 Materials
The experiment is achieved using an Acer ConceptD 3 Ezel laptop (Acer, New Taipei City, Taiwan) on Windows 11 having an i7-10750 H processor (Intel, California, United States), GTX1650 Ti (NVIDIA, California, United States),16 GB of RAM, and an external mouse as the input device. 3D Slicer [15, 16] (version 5.8.1) was utilized as the segmentation software as it is open-source and a large selection of community-built plug-ins that are easy-to-use and effective including the state-of-the-art automated tool mention earlier by Dot et al. [22].
The datasets chosen are CBCT scans uploaded on a publicly available database of medical scans to preserve anonymity [33]. Dataset 1 is an example of a clear and simple dataset with no major defects or abnormal patient-specific case except for an amalgam restoration. It was also taken with a bite block between the mandibular and maxillary teeth to avoid overlapping of entities. Dataset 2 is more challenging for the automated tool as it does not contain any mandibular teeth, the maxillary ones are of an adolescent as their apexes are still open, there is a newly erupted molar, and an impacted molar can be seen. Dataset 3 is the most challenging as the patient has orthodontic braces. Since they make direct contact with the teeth, delineating such featureless overlap would be a great challenge for all workflows. Figure 5 shows the 3D rendering of all 3 datasets.
Figure 5.
3D volume rendering of: (a) dataset 1, (b) dataset 2, and (c) dataset 3.
3.2 Unified pre-processing procedures
For manual and semi-automatic approaches, pre-processing steps are beneficial for streamlining the workflows and unifying an enhanced starting point. After importing the dataset, pre-processing steps are three-fold: volume cropping, histogram adjustment, and thresholding. Volume cropping is specifically important for semi-automatic methods to significantly reduce the amount of pixels that the algorithms should run through. It is a straightforward step using the built-in “Crop Volume” module inside 3D Slicer. Histogram adjustment is particularly advantageous for semi-automated methods as it accentuates any differences between brighter and darker pixels and therefore facilitates the algorithms labelling processes. For dataset 1, a window of 3,000 at level 1,600 in the “Volumes” module was effective without losing any valuable data due to aggressive histogram adjustment. Lastly, a thresholding mask was used to limit the selection of pixels only to non-dark ones and therefore save time by avoiding meticulous pixel selection in clearly distinguished parts, such as enamel and background. It is to be noted that while aggressive pre-processing might ease the selection in subsequent steps for manual or algorithmic workflows, slightly conservative parameters should be used as to not eliminate any important anatomical details from the original scan; having a slightly noisier set of pixels or larger mask of selection than needed is definitely advantageous over a noise-free picture with missing details or a mask that does not allow desired pixels selection. The preprocessing steps are shown in Figure 6.
Figure 6.
The preprocessing steps effects on (a) original CBCT of dataset 1, (b) after volume cropping, (c) after histogram adjustment, and (d) after threshold masking.
3.3 Manual segmentation
After modifying the input dataset using the preprocessing steps, the main workflow can start to take shape. Completely manual segmentation is carried out at first to obtain a ground truth as a benchmark and 6 additional 3D models of the same tooth per dataset to then compare them to the benchmark, totaling 21 models. In 3D Slicer, manual pixel selection can be performed using either a paintbrush or a drawing tool. The paintbrush features a circular selection area with an adjustable radius, whereas the drawing tool allows users to trace the outline of a region of interest to be incorporated into the active segmentation layer. These tools are shown in Figure 7.
Figure 7.
Manuals selection tools: (a) paint tool while selecting a slice, (b) the result of paint tool, (c) draw tool while selecting, and (d) the result of draw tool.
Both selection methods produce identical output; however, depending on geometry, one tool may be more effective than the other. Observations indicated that the brush tool was more time-efficient when dealing with large, uninterrupted pixel clusters with clear separation of the background in terms of pixel intensity where thresholding is most effective. In contrast, the drawing tool proved more adaptable for non-uniform slices, particularly near the alveolar bone, where the threshold mask was ineffective at separating tooth pixels from bone pixels. The painting process begins by selecting a plane of the 3 and a slice at one end of the target tooth which does not necessarily need to be the final slice, as another plane view may offer better visibility of the tooth’s boundaries. Tooth pixels were painted slice by slice in the first plane, followed by the same procedure in a second plane, and then a third. After completing the initial segmentation stage, another segmentation layer is created with a label “non-tooth” to serve as a scrap layer for manual refinement. Alternating cycles of trimming and fine painting were carried out from other viewpoints using the “Erase” tool in the software which deselects pixels or painting unwanted pixels while selecting the “non-tooth” layer.
3.4 Proposed semi-automatic segmentation
The next step now is to acquire 6 additional 3D segmentations of the same tooth per dataset using the proposed methodologies, namely GFS and WS.
3.4.1 Grow from seeds (gfs)
Unlike the fully manual painting approach described in Section 3.3, this technique does not require marking every slice; instead, roughly ten slices per plane view suffice as the “seeds” for the algorithm to start with as described in Section 1.1.3. Moreover, the painting process itself is relatively coarse and does not require capturing every pixel within a slice. This method is applied to label both the “tooth” and “non-tooth” layers, as illustrated in Figure 8 by blue and red colors, respectively.
Figure 8.
The grow from seeds workflow: (a) manual seeding process. (b) initial iteration of the algorithm with some mislabeling shown by yellow arrows.
Next, the GFS algorithm is initialized within 3D Slicer, producing a preview, which is usually shown in slightly faded colors, of how the software interprets each segment based on the input pixels. Following this, a visual inspection is necessary to evaluate whether the algorithm has correctly labeled the intervening slices between those manually painted. Imperfect labeling is to be expected; some of this mislabeling is shown in Figure 8(b) by the yellow arrows. Consequently, iterative corrections are made to the input slices to help guide the algorithm toward more accurate predictions of segment membership. A balance must be struck between achieving perfect segmentation and obtaining a reasonably accurate result quickly.
3.4.2 Watershed (WS)
Similar to GFS, this workflow only requires seeding in a limited number of slices. Narrow and axial paintbrush strokes were used as a technique to segment the “tooth” layer in the sagittal plane. At points where teeth came into contact, seeds were applied to both segments every three slices to differentiate between the two adjacent regions; the same approach was used in the subgingival area. It is important to note that the algorithm did not incorporate the thresholding step described in Section 3.2. Therefore, it was necessary to completely encircle the tooth using the “non-tooth” segment, as demonstrated in Figure 9(a).
The Watershed tool, available from the “SegmentEditorExtraEffects” extension, was then used to initialize the algorithm, with the object scale parameter set to 0.2. This value was chosen based on the user’s subjective judgment to ensure efficient results. Increasing the object scale value produced a smoother selection, as this parameter plays a key role in edge detection; however, higher values inevitably lead to a loss of fine detail in the segmented model.
Figure 9.
The watershed workflow: (a) manual seeding using longitudinal strokes and encapsulating technique to compensate for lack of thresholding. (b) the initial output result of the algorithm with mislabeling shown by a yellow arrow.
Iterative refinements similar to those in the previous workflow were used to improve the algorithm’s selection, though the process required considerably less time since the initial output of this algorithm was substantially cleaner in terms of noise and mislabeling, an example of which is shown in Figure 9(b) by a yellow arrow.
3.5 Unified post-processing procedures
Following the generation of 3D models, the resolution inherent to each scanning device may produce sharp edges and surface irregularities that deviate from the true anatomical form of the exported model. Thus, a smoothing or polishing step must be applied to all techniques. Through 3D Slicer, a median filter with a kernel size of 3 × 3 × 3 pixels was applied to all 3D models. The choice of this particular kernel size was driven by the fact that it represents the mathematically minimal amount of smoothing, thereby posing no risk of removing details resulting from the segmentation process. It is also worth noting that the smoothing effect was applied exclusively to the “tooth” layer to eliminate imperfections on the outer tooth layer and within the root canal. The decision to withhold the median filter from the “non-tooth” layer stems from the observation that the root canal portion of this layer is often less than 3 pixels wide; applying the kernel could entirely erase the root canal and inadvertently solidify the “tooth” layer. The result of this smoothing process is shown in Figure 10.
Figure 10.
A 3D model obtained using GFS: (a) before smoothing and (b) after [32].
3.6 Dentalsegmentator automated tool
The extension, developed by Dot et al. [22], features a remarkably straightforward interface, offering only a single adjustable parameter related to surface smoothness. For the datasets used in this study, this parameter was retained at its default value of 0.5. Additionally, users have the option to run the model on the CPU if their computer lacks a CUDA-compatible GPU, though this naturally results in longer processing times.
The plugin’s output, shown in Figure 11, is not ready to be compared directly to the previously attained 3D models from other methods. The tool works by segmenting the datasets into 5 separate layers: maxilla and upper skull, mandible, upper teeth, lower teeth, and mandibular canal. Moreover, the teeth layers do not individually segment each tooth which is an acknowledged limitation by the authors [22]. Therefore, a manual step was carried out to cut out everything from the upper teeth layer except the target maxillary second molar to be compared. This was done by erasing the contact lines between it and adjacent teeth and then choosing to “keep selected island” in the “islands” tool from the segment editor of 3D Slicer.
Figure 11.
The output of the dentalsegmentator tool in 3D and 3 planes.
3.7 Obtaining roots of 3D teeth models
After obtaining 6 smoothed 3D models from each of the manual, semi-automated methods, and automated method, they were separated at a consistent height level using an isolation and island removal technique similar to that described in Section 3.5, ensuring clean cuts along the same plane. This is mainly to test the accuracy of the workflows in endodontics-oriented applications, which is an oversight of most previous similar work as mentioned in Section 2. After carrying out this step, the teeth models obtained, totaling 150 models, are now summarized in Figure 12.
Figure 12.
A summary of all 3D models acquired through various methods.
A dice similarity score, which is a widely used metric in image segmentation for validation [34], was obtained for each 3D model against the benchmark model of the respective anatomy, whether whole teeth or only the subgingival part. These scores was achieved using 3D Slicer’s community-built extension “SlicerRT” [35]. For the entire tooth models, a summary box-and-whiskers graph of this anatomy for all 3 datasets is shown in Figure 13.
Figure 13.
A summary plot of the dice score comparison results for the whole tooth.
As for the subgingival part anatomy models, a summary box-and-whiskers graph for all 3 datasets is shown in Figure 14.
Figure 14.
A summary plot of the dice score comparison results for the subgingival tooth part.
Table 1 shows the summary of recorded dice scores for each anatomy and method.
Method
Anatomy
Dice Scores
Lowest
Average
Highest
Manual
Whole Tooth
0.973
0.978
0.983
Subgingival Part
0.923
0.944
0.965
GFS
Whole Tooth
0.957
0.965
0.978
Subgingival Part
0.904
0.930
0.9543
WS
Whole Tooth
0.940
0.956
0.965
Subgingival Part
0.876
0.931
0.951
Automated
Whole Tooth
0.907
0.936
0.942
Subgingival Part
0.887
0.905
0.927
Table 1.
A summary of the dice scores for all trials.
A Kruskal-Wallis test was then performed and a p-value < 0.001 was achieved for both sets of comparisons. This meant, for both comparisons, at least one of the groups deviated significantly from at least one of the others. To determine which groups are unsimilar from the manual group, Nemenyi’s post hoc test was performed, and it stated that GFS models consistently did not differ from the manual models in all datasets, WS models only deviated on the whole tooth anatomy in datasets 1 and 2, and models acquired through the automated workflow consistently deviated from the manual models in all datasets. These conclusion are drawn after comparing the q-stat values of each group comparison to the critical q-value from the Studentized Range Q table [36]; according to the 0.05 alpha level and the number of trials, the critical q-value is 3.633. The summary of the post hoc results can be seen in Table 2 for all models and all datasets.
Dataset
Group 1
Group 2
Tooth Anatomy
q-stat
Statistically different?
Dataset 1
Manual
GFS
Whole
2.078
No
Subgingival
1.612
No
Manual
WS
Whole
4.157
Yes
Subgingival
1.501
No
Manual
Automated
Whole
6.235
Yes
Subgingival
5.196
Yes
Dataset 2
Manual
GFS
Whole
2.078
No
Subgingival
3.233
No
Manual
WS
Whole
4.503
Yes
Subgingival
3.349
No
Manual
Automated
Whole
5.889
Yes
Subgingival
5.889
Yes
Dataset 3
Manual
GFS
Whole
2.309
No
Subgingival
3.406
No
Manual
WS
Whole
3.406
No
Subgingival
1.617
No
Manual
Automated
Whole
6.062
Yes
Subgingival
4.907
Yes
Table 2.
A summary of the post hoc results according to dice scores.
Moreover, the time taken for each trial was recorded in minutes and can be seen in Table 3. It is clear that the average times of non-manual workflows are significantly shorter, allowing for more efficient and streamlined approaches to create more 3D models in the same time span.
Dataset
Average time of trials in minutes for each workflow
Manual
GFS
WS
Automated
Dataset 1
52.5
30.8
15.5
7.7
Dataset 2
31
10
8.2
5.5
Dataset 3
43.7
19
17.7
5.8
Table 3.
The individual and average times, in minutes, of each trial in each of the workflows.
As can be seen from the results in Section 4. GFS is a viable alternative to manual segmentation according to our statistical analysis. WS was also an efficient workflow, most of the time, in replicating decently similar anatomies, especially considering the time savings compared to manual or GFS workflows. The automated tool, however, did not provide accurate enough 3D models in comparison to the other workflows. GFS, therefore, can be recommended to be used as a workflow for ground truth generation for deep learning-based models to provide much larger training datasets. There are several disadvantages in the automated workflow; the inability for individual teeth segmentation can be inconvenient for some use cases. Furthermore, the segmentation of the root canals as part of the tooth and create solid 3D models instead of hollow ones definitely is a downside for endodontics-oriented applications, this can be seen in Figure 15.
Figure 15.
Incorrect labeling of root canals as part of the tooth by the automated tool.
A beneficial part of the semi-automated methods is the ability to be modified for a wide variety of techniques. As an example, to differentiate between enamel and dentin, a new layer could be created after the original workflow and a new threshold range be assigned to limit the lower value of which the mask is applied to; this is because enamel is much more radio-opaque than dentin. The user can then reselect the enamel part in the new layer, overwriting the pre-existing selection in the “tooth” layer. This would ultimately create 2 distinct 3D models as the output which can then be dealt with in separate ways depending on the use case, for example, apply different textures and material properties for different sensations in a VR environment. This output of this technique can be seen in Figure 16.
Figure 16.
The result of using 2 different thresholding masks to differentiate between enamel and dentin in: (a) axial plane, (b) sagittal plane, and (c) 3D view.
Several limitations should be acknowledged in this work. The sample size, while sufficient for statistical comparison, remains modest and may not fully capture the anatomical variability present in broader clinical populations. Despite the sample size in our extended, previous research being larger and demonstrating similar conclusions, more trials should be attempted for guaranteed clinical recommendation [32]. Additionally, the proposed workflows inherently involve operator-dependent decisions, such as seed placement, threshold selection, and iterative correction, which could influence reproducibility. No formal inter-operator variability analysis was conducted; therefore, the reported time and accuracy metrics reflect the performance of a single experienced user. Further testing should include multiple operators to quantify consistency across different skill levels and to establish more generalizable guidelines for clinical adoption.
This chapter has addressed the critical challenge of generating accurate, high-quality ground truth annotations for dental image segmentation, with a particular focus on endodontics-oriented applications where precise delineation of root canals and pulp chambers is essential. Given the limitations of fully automated deep learning approaches, namely their dependence on large, laboriously annotated datasets and concerns regarding cross-population generalizability, we have proposed and validated two semi-automatic workflows, GFS and WS, which are implemented within the open-source 3D Slicer platform.
Quantitative analysis showed that GFS matched manual segmentation accuracy, while WS reduced processing time to 8–18 minutes per trial (40–70% faster) without compromising most results, unlike the fully automated DentalSegmentator, which failed to identify root canals. Clinically, these workflows enable rapid creation of large, annotated datasets for deep learning and can be adapted to differentiate dentin/enamel or isolate subgingival anatomy, supporting biomechanical modeling, VR endodontic simulators, and personalized treatment planning. Future work should extend validation to micro‑CT scans and diverse populations, though current evidence confirms that semi‑automated segmentation offers a robust, efficient, and accessible alternative to manual or fully automated methods.
References
1.KahnCE, CarrinoJA, FlynnMJ, PeckDJ, HoriiSC. DICOM and Radiology: Past, Present, and Future. Journal of the American College of Radiology. 2007;4(9):652–657. DOI: 10.1016/j.jacr.2007.06.004
2.XuY, QuanR, XuW, HuangY, ChenX, LiuF. Advances in Medical Image Segmentation: A Comprehensive Review of Traditional, Deep Learning and Hybrid Approaches. Bioengineering (Basel). 2024;11(10):1034. DOI: 10.3390/bioengineering11101034
3.SalpeaN, TzouveliP, KolliasD. Medical Image Segmentation: A Review of Modern Architectures. In: Computer Vision – ECCV 2022 Workshops: Tel Aviv, Israel. October23–27, 2022. Proceedings, Part VII [Internet]. Berlin, Heidelberg: Springer-Verlag; 2022. p. 691–708. doi:10.1007/978-3-031-25082-8_47
4.ZhengQ, GaoY, ZhouM, LiH, LinJ, ZhangW, et al. Semi or fully automatic tooth segmentation in CBCT images: A review. PeerJ Computer Science. 2024;10:e1994
5.HaberthürD, HlushchukR, WolfTG. Automated segmentation and description of the internal morphology of human permanent teeth by means of micro-CT. BMC Oral Health. 2021;21(1):185. DOI: 10.1186/s12903-021-01551-x
6.CarreraCA, LanC, Escobar-SanabriaD, LiY, RudneyJ, AparicioC, et al. The use of micro-CT with image segmentation to quantify leakage in dental restorations. Dental Materials. 2015;31(4):382–390. DOI: 10.1016/j.dental.2015.01.002
7.MajangaV, ViririS. Dental Images’ Segmentation Using Threshold Connected Component Analysis. Computational Intelligence and Neuroscience. 2021;2021:2921508. DOI: 10.1155/2021/2921508
8.KakehbaraeiS, SeyedarabiH, ZenouzAT. Dental Segmentation in Cone-beam Computed Tomography Images Using Watershed and Morphology Operators. Journal of Medical Signals & Sensors. 2018;8(2):119–124
9.WangL, PengLJ, puGZ, LiG. CBCT image based segmentation method for tooth pulp cavity region extraction. Dentomaxillofacial Radiology. 2018;48(2):20180236. DOI: 10.1259/dmfr.20180236
10.VincentL, SoilleP. Watersheds in digital spaces: An efficient algorithm based on immersion simulations. IEEE Transactions on Pattern Analysis and Machine Intelligence. 1991;13(6):583–598. DOI: 10.1109/34.87344
11.GalibourgA, DumoncelJ, TelmonN, CalvetA, MichettiJ, MaretD. Assessment of automatic segmentation of teeth using a watershed-based method. Dentomaxillofacial Radiology. 2018;47(1):20170220. DOI: 10.1259/dmfr.20170220
12.VezhnevetsV, KonouchineV. “GrowCut” - Interactive Multi-Label N-D Image Segmentation By Cellular Automata. In: Proceedings of the 15th International Conference GraphiCon. 2005June. p. 150–156. Available from: https://www.graphicon.org/html/2005/proceedings/papers/VezhntvetsKonushin.pdf
13.QuasthoffU. Nested Cellular Automata: Continuous Aspects of Discrete Systems. In: PietroneroL, TosattiE, editors. Fractals in Physics [Internet]. Amsterdam, The Netherlands: Elsevier; 1986. p. 465–468. DOI: 10.1016/B978-0-444-86995-1.50086-X
14.ZhuL, KolesovI, GaoY, KikinisR, TannenbaumA. An Effective Interactive Medical Image Segmentation Method Using Fast GrowCut. In: MICCAI Workshop on Interactive Medical Image Computing. 2014September
15.KikinisR, PieperSD, VosburghKG. 3D Slicer: A Platform for Subject-Specific Image Analysis, Visualization, and Clinical Support. In: JoleszFA, editor. Intraoperative Imaging and Image-Guided Therapy [Internet]. New York, NY: Springer; 2014. p. 277–289. DOI: 10.1007/978-1-4614-7657-3_19
16.3D Slicer [Internet]. 3D Slicer image computing platform. Available from: https://slicer.org/[Accessed: 2023-November-12]
17.RonnebergerO, FischerP, BroxT. U-Net: Convolutional Networks for Biomedical Image Segmentation. In: NavabN, HorneggerJ, WellsWM, FrangiAF, editors. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Cham: Springer International Publishing; 2015. p. 234–241. DOI: 10.1007/978-3-319-24574-4_28
18.ÇiçekÖ, AbdulkadirA, LienkampSS, BroxT, RonnebergerO. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In: OurselinS, JoskowiczL, SabuncuMR, UnalG, WellsW, editors. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016. Cham: Springer International Publishing; 2016. p. 424–432. DOI: 10.1007/978-3-319-46723-8_49
19.DumontM, PrietoJC, BrossetS, CevidanesL, BianchiJ, RuellasA, et al. Patient Specific Classification of Dental Root Canal and Crown Shape. In: ReuterM, WachingerC, LombaertH, PaniaguaB, GokselO, RekikI, editors. Shape in Medical Imaging. Cham: Springer International Publishing; 2020. p. 145–153. DOI: 10.1007/978-3-030-61056-2_12
20.WangY, XiaW, YanZ, ZhaoL, BianX, LiuC, et al. Root canal treatment planning by automatic tooth and root canal segmentation in dental CBCT with deep multi-task feature learning. Medical Image Analysis. 2023;85:102750
21.DuanW, ChenY, ZhangQ, LinX, YangX. Refined tooth and pulp segmentation using U-Net in CBCT image. Dentomaxillofacial Radiology. 2021;50(6):20200251. DOI: 10.1259/dmfr.20200251
22.DotG, ChaurasiaA, DuboisG, SavoldelliC, HaghighatS, AzimianS, et al. DentalSegmentator: Robust open source deep learning-based CT and CBCT image segmentation. Journal of Dentistry. 2024;147:105130
23.TianS, WangM, YuanF, DaiN, SunY, XieW, et al. Efficient Computer-Aided Design of Dental Inlay Restoration: A Deep Adversarial Framework. IEEE Transactions on Medical Imaging. 2021;40(9):2415–2427. DOI: 10.1109/TMI.2021.3077334
24.KoikeY, AnetaiY, TakegawaH, OhiraS, NakamuraS, TanigawaN. Deep learning-based metal artifact reduction using cycle-consistent adversarial network for intensity-modulated head and neck radiation therapy treatment planning. Physica Medica: European Journal of Medical Physics. 2020;78:8–14. DOI: 10.1016/j.ejmp.2020.08.018
25.MakhloufA, MaayahM, AbughanamN, CatalC. The use of generative adversarial networks in medical image augmentation. Neural Computing and Applications. 2023;35(34):24055–24068. DOI: 10.1007/s00521-023-09100-z
26.XunS, LiD, ZhuH, ChenM, WangJ, LiJ, et al. Generative adversarial networks in medical image segmentation: A review. Computers in Biology and Medicine. 2022;140:105063
27.Yepes-CalderonF, McCombJG. Eliminating the need for manual segmentation to determine size and volume from MRI. A proof of concept on segmenting the lateral ventricles. PLoS One. 2023;18(5):e0285414. DOI: 10.1371/journal.pone.0285414
28.LiangX, NguyenD, JiangSB. Generalizability issues with deep learning models in medicine and their potential solutions: Illustrated with cone-beam computed tomography (CBCT) to computed tomography (CT) image conversion. Machine Learning: Science and Technology. 2020;2(1):015007. DOI: 10.1088/2632-2153/abb214
29.MorellGFC, ChenK, Flores-MirC. 3D reconstruction of lower anterior teeth from CBCT images: Automatic segmentation with manual refinements. Dental Press Journal of Orthodontics. 2023;28:e232249. DOI: 10.1590/2177-6709.28.3.e232249oar
30.VerykokouS, IoannidisC, AngelopoulosC. Evaluation of 3D Modeling Workflows Using Dental CBCT Data for Periodontal Regenerative Treatment. Journal of Personalized Medicine. 2022;12(9):9. DOI: 10.3390/jpm12091355
31.GanY, XiaZ, XiongJ, LiG, ZhaoQ. Tooth and Alveolar Bone Segmentation From Dental Computed Tomography Images. IEEE Journal of Biomedical and Health Informatics. 2018;22(1):196–204. DOI: 10.1109/JBHI.2017.2709406
32.ElaYAE, BadranM. Development of Semi-Automatic Dental Image Segmentation Workflows with Root Canal Recognition for Faster Ground Tooth Acquisition. Journal of Imaging. 2025;11(10). DOI: 10.3390/jimaging11100340
33.embodi3D.com. Embodi3D Home Biomedical 3D Printing [Internet]. Available from: https://www.embodi3d.com[Accessed: 2024-January-10]
34.ZouKH, WarfieldSK, BharathaA, TempanyCMC, KausMR, HakerSJ, et al. Statistical Validation of Image Segmentation Quality Based on a Spatial Overlap Index. Academic Radiology. 2004;11(2):178–189. DOI: 10.1016/S1076-6332(03)00671-8
Open access peer-reviewed chapter - ONLINE FIRST
