Evaluation of an AI-based segmentation model for mandibular fracture segmentation on CBCT using volumetric and 2D slice-based validation approaches Authors: Ilgın Arı, Ersagun Kara, Arzum Yılmaz, Ecem Usman, Alper Aktaş Published: 2026 ABSTRACT AbstractObjectiveThis study aimed to evaluate the diagnostic performance of an artificial intelligence (AI)-based segmentation model for mandibular fracture segmentation on cone-beam computed tomography (CBCT) scans using both volumetric (3D) and 2D slice-based validation approaches.Materials and methodsCBCT data from 92 patients, comprising 142 mandibular fracture sites, were retrospectively collected. Ground truth annotations were established through manual segmentation by two oral and maxillofacial surgeons. The AI model was developed using the nnU-Net v2 architecture in 3D full-resolution mode. The dataset was divided into training (90%) and test (10%) subsets. Model performance was evaluated by comparing AI-generated segmentations with expert annotations using accuracy, Dice similarity coefficient (DICE), Jaccard index, precision, recall, average surface distance (ASD), and the 95% Hausdorff distance. These metrics were calculated for both volumetric (3D) and 2D slice-based validation approaches.ResultsVolumetric (3D) evaluation yielded a DICE of 0.605, a Jaccard index of 0.446, a precision of 0.61, and a recall of 0.663. In contrast, 2D slice-based validation resulted in a DICE of 0.492, a Jaccard index of 0.350, a precision of 0.60, and a recall of 0.484. The average surface distance was 4.88 mm for the 3D model and 7.85 mm for the 2D model, while the 95% Hausdorff distance was 16.48 mm and 22.04 mm, respectively.ConclusionThe AI-based segmentation model demonstrated a moderate level of performance in segmenting mandibular fracture regions on CBCT scans. Although overall segmentation accuracy was limited, the volumetric (3D) validation approach consistently outperformed the 2D slice-based method. These findings suggest that volumetric evaluation provides a more comprehensive and reliable framework for assessing AI-assisted mandibular fracture segmentation on CBCT images.Keywords: Mandibular fracture, Artificial intelligence, Deep learning, Segmentation volumetric analysis, 2D slice-based validation Introduction Mandibular fractures are among the most commonly encountered injuries in maxillofacial trauma, and accurate diagnosis is essential for effective treatment and surgical planning [1]. Cone-beam computed tomography (CBCT) has become a widely preferred imaging modality for the evaluation of mandibular fractures due to its high-resolution three-dimensional imaging capabilities and relatively low level of artifacts [2] CBCT may offer important advantages, particularly in the assessment of fractures in anatomically complex regions such as the condyle, ramus, and angle [3]. Despite these advantages, the interpretation of CBCT images remains subject to human-related limitations [4, 5]. Diagnostic accuracy may be affected by interobserver variability, perceptual errors, and clinician fatigue [6–8]. In this context, artificial intelligence (AI)-based decision support systems have emerged as promising tools with the potential to improve diagnostic consistency and efficiency.⁶⁻⁸ AI models trained on large datasets may assist in identifying fracture lines more reliably, thereby reducing oversight and enhancing the standardization of diagnosis in clinical practice [3, 9–12]. Most studies investigating AI-assisted analysis of mandibular fractures have focused on panoramic radiographs [13–15]. However, panoramic imaging has notable limitations, including low spatial resolution and anatomical superimposition, which may adversely affect diagnostic performance, particularly in posterior mandibular regions [9, 10, 13, 15–18]. In the limited number of studies utilizing CBCT data, two-dimensional (2D) slice-based validation methods have predominantly been employed. Similarly, most previous studies based on CT imaging have also relied on two-dimensional analyses. Although these approaches can provide useful diagnostic information, they may be insufficient to fully capture the complex three-dimensional morphology and spatial continuity of mandibular fractures [3, 8, 11]. The number of studies investigating volumetric (3D) validation approaches for evaluating AI model performance remains limited in literature. Volumetric analysis may allow a more holistic assessment of fracture regions within the anatomical volume and may contribute to a more detailed understanding of fracture geometry and spatial relationships. In this context, volumetric evaluation metrics such as the DICE and the Jaccard index may facilitate a more quantitative assessment of spatial overlap, thereby supporting more detailed analyses of fracture characteristics. Accordingly, volumetric evaluation approaches in CBCT-based mandibular fracture analysis may offer a more comprehensive assessment compared with conventional slice-based methods. Although 2D slice-based validation approaches are widely used in the literature, studies that systematically evaluate AI models using volumetric ground truth comparisons remain limited. In the present study, an AI-based segmentation model was developed for the analysis of mandibular fractures on CBCT images, and its performance was evaluated using both volumetric (3D) and 2D slice-based validation approaches. In this respect, the study aims to provide a comparative evaluation of AI model performance across two validation strategies, with particular attention to the potential clinical relevance of volumetric analysis. This complementary evaluation framework may contribute to a broader understanding of algorithm performance and may provide further insight into its potential clinical applicability. Accordingly, this study aimed to explore potential differences between volumetric (3D) and 2D slice-based validation approaches in the evaluation of AI-assisted mandibular fracture segmentation. Materials and methods This retrospective study was approved by the Ethical Committee of Hacettepe University and conducted in accordance with the Declaration of Helsinki. Data collection The CBCT scans used in this study were obtained using an i-CAT Next Generation device (Imaging Sciences International, Hatfield, PA, USA). Image analyses were performed using i-CAT Vision software (version 1.9.3.14, Imaging Sciences International, Hatfield, PA, USA), and linear measurements were conducted. The imaging parameters were configured as follows: 120 kVp tube voltage, 0.2–0.3 mm³ voxel size, 3–7 mA tube current depending on the voxel size, 1 mm slice thickness, different FOV sizes (16 × 10–16 × 13 cm), and variable exposure times. CBCT scans with a slice thickness of 1 mm were retrieved in DICOM format from the archives of the Department of Oral and Maxillofacial Surgery, Hacettepe University. An example of fracture case presented Fig. 1 as axial, sagittal and transverse planes and Fig. 2 as three-dimensional image. Patients with documented mandibular fractures due to maxillofacial trauma between 2015 and 2025 were included. The study included patients who met the following criteria: (i) age between 18 and 80 years, (ii) confirmed mandibular fracture within the previous 15 days after trauma, (iii) absence of systemic bone disease or congenital facial deformities, and (iv) no history of radiotherapy or chemotherapy. Only linear mandibular fractures were included in the study, while comminuted fractures were excluded to ensure consistency in annotation. Only patients with mandibular fractures were included; non-fractured cases were excluded. Cone-beam computed tomography (CBCT) images demonstrating mandibular fractures in (A) axial, (B) coronal, and (C) sagittal planes Three-dimensional (3-D) cone-beam computed tomography (CBCT) images of mandibular fractures are presented from (A) frontal, (B) left lateral, and (C) right lateral views A total of 92 patients with mandibular fractures, comprising 142 fracture sites, were included in the study, including 23 females and 69 males. The causes of mandibular fractures were falls in 36 patients, sports-related injuries in 7 patients, interpersonal violence in 28 patients, and traffic accidents in 21 patients. The age of the patients ranged from 18 to 64 years, with a mean age of 29.63 ± 12.09 years. The overall anatomical distribution of fracture regions and displacement status within the entire sample is presented in Table 1. Condylar (n = 48) and angle (n = 29) regions were the most frequently affected sites, followed by the parasymphysis (n = 26). Table 1Distribution of mandibular fracture sites according to anatomical regionsAnatomical regionDisplaced (n)Non-displaced (n)TotalCondyle311748Ramus044Angle121729Corpus61622Parasymphysis52126Symphysis21113Total5686142 Distribution of mandibular fracture sites according to anatomical regions Of the total sample, 83 patients (128 fractures) were allocated to the training and validation sets, and 9 patients (14 fractures) formed the independent test set. The specific distribution of fracture regions and displacement patterns within the test group is detailed in Table 2. Table 2Fracture distributions in the test patientsAnatomical regionDisplaced (n)Non-displaced (n)TotalCondyle224Ramus000Angle134Corpus011Parasymphysis134Symphysis011Total41014 Fracture distributions in the test patients Non-displaced mandibular fractures are defined as fractures with no or minimal separation between bone fragments, typically maintaining anatomical alignment and cortical continuity, and are often managed conservatively [19]. In contrast, displaced mandibular fractures are characterized by visible separation and misalignment of fracture segments, often requiring reduction to restore anatomical positioning [20]. Annotation and data preparation Fracture regions were manually segmented by experienced oral and maxillofacial surgeons based on the hypodense discontinuity observed in CBCT images. Instead of enforcing a strict single-voxel-width definition, the segmentation was performed by delineating the clinically relevant fracture gap along the visible discontinuity. Consistent annotation principles were applied across all cases. Manual segmentation of mandibular fracture regions was performed on CBCT images by-slice by an oral and maxillofacial surgeon (I.A.) using 3D Slicer software (Slicer 5.8.0-win.amd 64) [21] (Fig. 3). All annotations were reviewed by a second expert (A.A.). Final annotations were established through a consensus-based approach to ensure anatomical consistency and clinical relevance. Cases in which consensus could not be reached were excluded from the study to maintain the reliability and consistency of the reference standard. The fracture lines were carefully delineated slice-by-slice on axial images, with additional verification on sagittal and coronal planes to ensure anatomical consistency. The segmentation masks were generated by outlining the fracture boundaries within the mandibular bone, and each case was reviewed by an experienced oral and maxillofacial surgeon (A.A.) to ensure the accuracy of the annotations. The finalized segmentation masks were exported in NIfTI format and used as ground truth labels for training and validation of the deep learning model. The dataset was divided into training and test sets, with 90% of the data allocated for training and validation and the remaining 10% reserved as an independent test set. Given the relatively limited number of cases, a larger proportion of the data was allocated to the training set to improve the model’s learning capacity and generalization performance. A hold-out validation strategy was used during training, in which a portion of the training dataset was reserved as a validation subset to monitor model performance and prevent overfitting. A total of 83 CBCT scans were used for training and validation, and 9 volumes were reserved for testing. In addition to voxel-level evaluation, a case-based analysis was performed using DICE calculated for each individual test case to assess segmentation performance at the patient level. Cone-beam computed tomography (CBCT) images showing expert-annotated fracture regions in (A) axial, (B) coronal, and (C) sagittal planes. In the two-dimensional (2D) slice-based evaluation, these annotations are compared with AI-generated segmentations to assess model performance AI model training (nnU-Net) The segmentation model was developed using the nnU-Net v2 framework, which is specifically designed for biomedical image segmentation and automatically configures preprocessing and training pipelines based on dataset characteristics. Preprocessing was determined through the nnU-Net dataset fingerprinting mechanism, without any manual intervention. Image properties such as voxel spacing, median image size, and intensity distribution were analyzed to define optimal configuration parameters. Accordingly, images were processed using an automatically determined patch size (512 × 512), based on GPU memory constraints and target spacing. Intensity normalization was performed using z-score normalization derived from the 0.5 and 99.5 percentiles of foreground voxels, as defined within the nnU-Net pipeline. The model was trained using a combination of DICE and Cross-Entropy loss functions. Optimization was performed using stochastic gradient descent with Nesterov momentum (0.99) and a polynomial learning rate schedule. To improve generalization, data augmentation techniques including Gaussian noise, Gaussian blur, rotation, scaling, and mirroring were applied via the batchgenerators framework [22]. The model was trained for 1000 epochs, and the best-performing checkpoint based on validation performance was selected as the final model. The initial learning rate was set to 0.01 and updated using a polynomial decay schedule throughout training. The batch size was automatically determined based on GPU memory constraints and was set to 2 for 3D training and 12 for 2D training. Dropout was not applied, as the nnU-Net architecture relies on instance normalization and data augmentation for regularization. Early stopping was not used during training. Data augmentation was automatically configured by the nnU-Net v2 framework based on dataset fingerprinting and probabilistically applied during training. Spatial transformations included random rotations, scaling, flipping, and elastic deformations. For 3D training, random rotations between − 30° and + 30° were applied along the x, y, and z axes, whereas 2D training included random in-plane rotations between − 180° and + 180°. Random scaling between 0.7 and 1.4 and flipping with a probability of 0.5 along each axis were also applied. Elastic deformation was additionally used to increase structural variability. Intensity-based augmentations included brightness adjustment with intensity scaling factors between 0.7 and 1.3, contrast variation between 0.65 and 1.5, and gamma correction with gamma values ranging from 0.7 to 1.5. In addition, Gaussian noise and Gaussian blur were applied to improve model robustness against image variability. For 3D full-resolution (3D_fullres) training configuration the model was trained using the nnU-Net v2 framework with the 3D_fullres configuration. The training process consisted of 1000 epochs, each including 250 iterations (a total of 250,000 iterations). Optimization was performed using stochastic gradient descent (SGD) with Nesterov momentum (0.99) and a weight decay of 3 × 10⁻⁵. The initial learning rate was set to 0.01 and updated using a polynomial decay (poly) learning rate schedule throughout training. The loss function was defined as a combination of DICE and Cross-Entropy losses, with deep supervision applied across multiple resolution levels. The model input was defined with a patch size of [112, 160, 128] and a batch size of 2. Intensity normalization was performed using CT-specific normalization based on the 0.5 and 99.5 percentiles of foreground voxel intensities. For 2D full-resolution training configuration the model was trained using the nnU-Net v2 framework with the 2D configuration. The training process consisted of 1000 epochs, each including 250 iterations (a total of 250,000 iterations). Optimization was performed using stochastic gradient descent (SGD) with Nesterov momentum (0.99) and a weight decay of 3 × 10⁻⁵. The initial learning rate was set to 0.01 and updated using a polynomial decay (poly) learning rate schedule throughout training. The loss function was defined as a combination of DICE and Cross-Entropy losses, with deep supervision applied across multiple scales. The model input was defined with a patch size of [512, 512] and a batch size of 12. Intensity normalization was performed using CT-specific normalization based on the 0.5 and 99.5 percentiles of foreground intensity values. Evaluation strategy and performance metrics The performance of the segmentation model was evaluated on the independent test dataset using two complementary evaluation strategies: volumetric (3D) evaluation and slice-based (2D) evaluation. For the volumetric (3D) evaluation, expert-labeled three-dimensional ground truth outputs (Fig. 4) were used as reference standards and compared with the AI-generated fracture segmentations (Fig. 5). The comparison was performed using standard segmentation performance metrics, including DICE, Jaccard index, accuracy, precision, recall, ASD, and the 95% Hausdorff distance. This approach evaluates the spatial agreement between the predicted and reference segmentations within the entire fracture volume. Three-dimensional (3-D) ground truth outputs labeled by experts are presented in (A) frontal, (B) left lateral, and (C) right lateral views Three-dimensional (3-D) artificial intelligence (AI)-generated fracture segmentations are presented in (A) frontal, (B) left lateral, and (C) right lateral views For the slice-based (2D) evaluation, expert-annotated fracture regions (Fig. 3) were used as reference standards and compared with the AI-detected fracture boundaries (Fig. 6). In this approach, the predicted fracture areas were evaluated against expert annotations using the same performance metrics to ensure methodological consistency Specifically, 3D volumes were decomposed into individual axial slices. For each slice, boundary points were extracted from both the predicted and ground truth segmentation masks. The 95% Hausdorff distance was computed as the maximum distance from a point on one boundary to the closest point on the corresponding boundary, representing the worst-case segmentation error for each slice. The ASD was calculated as the mean distance between all boundary points of the predicted mask and the ground truth boundary, reflecting the overall contour accuracy. To quantify the model’s 2D spatial performance, the final reported values were obtained by averaging the metric results across all slices. Two-dimensional (2D) slice-based artificial intelligence (AI)-detected fracture boundaries are presented in (A) axial, (B) coronal, and (C) sagittal planes In this study, model performance was evaluated based on voxel-level segmentation comparisons. The segmentation outputs generated by the AI model were compared voxel-wise with the reference (ground truth) annotations manually defined by expert clinicians. Based on this comparison, voxels correctly identified as fracture in both the reference annotations and model output were classified as true positives (TP), while voxels labeled as fracture in the reference but not detected by the model were considered false negatives (FN). Voxels incorrectly predicted as fracture by the model in regions where no fracture was present in the reference were defined as false positives (FP), whereas voxels correctly identified as non-fracture in both the reference and model output were classified as true negatives (TN). These voxel-level classifications were then used to compute performance metrics [23]. The DICE quantifies the spatial overlap between predicted and reference segmentations and is calculated as: Considering the thin nature of fracture structures, overlap-based metrics such as DICE may be sensitive to small spatial deviations; therefore, surface-based metrics (ASD and 95% Hausdorff Distance) were also included to provide a more robust evaluation. The Jaccard index evaluates the agreement between predicted and ground truth masks. DICE and Jaccard metrics were both used to provide a more comprehensive evaluation of segmentation performance, as they capture overlap with different sensitivity levels. Accuracy represents the proportion of correctly classified pixels or voxels among all evaluated samples. Precision indicates the proportion of correctly predicted positive cases among all predicted positives, whereas recall (sensitivity) represents the proportion of actual positive cases correctly identified by the model. Boundary-based metrics were also used to assess segmentation accuracy. The ASD represents the mean spatial distance between the surfaces of the predicted and reference segmentations, while the 95% Hausdorff Distance reflects the largest boundary deviation between the two segmentations after excluding extreme outliers [13]. Statistical analysis This study did not aim to perform a hypothesis-driven statistical comparison between independent groups. Instead, the performance of a single AI segmentation model was evaluated using two different validation approaches (volumetric 3D and axial slice-based 2D) on the same test dataset. Since the evaluated metrics were derived from the same cases and were not statistically independent, classical significance testing based on p-values was not applied. Model performance was reported using descriptive and geometric metrics, including the DICE, Jaccard index, accuracy, precision, recall, ASD, and the 95% Hausdorff distance [24–26]. Results Quantitative performance comparison between volumetric (3D) and 2D slice-based evaluation methods is shown in Table 3. The 3D model achieved higher DICE (0.605 vs. 0.492) and Jaccard index (0.446 vs. 0.350) values, reflecting greater spatial overlap with expert annotations. It also demonstrated superior recall (0.663 vs. 0.484) and smaller boundary errors, indicated by a lower average surface distance (4.88 mm vs. 7.85 mm) and 95% Hausdorff distance, DICE (16.48 mm vs. 22.04 mm). Case-based evaluation (Table 4) further supported these findings, showing more consistent segmentation accuracy for the 3D model across test patients. Overall, volumetric validation provided a more accurate and clinically relevant performance assessment than 2D slice-based analysis. Table 3Comparison of quantitative performance metrics between volumetric (3D) and axial-slice (2D) validation of the AI modelMetrics3D model2D modelAccuracy0.990.99DICE0.6050.492Jaccard index0.4460.350Precision0.610.60Recall0.6630.484Avg. surface distance (mm)4.887.85Hausdorff distance 95 (mm)16.4822.04 Comparison of quantitative performance metrics between volumetric (3D) and axial-slice (2D) validation of the AI model Table 4Comparison of patient-wise DICE scores, location and type (displaced and non-displaced) of fractures between volumetric (3D) and axial-slice (2D) validation of the AI model in test group. (Displaced fracture locations were shown in red, while non-displaced locations were shown in green)Patient ID3D Model2D ModelLocation and type of fractures10.7070.627Parasymphysis20.6110.632Angle30.2690.031Angle, Condyle40.7670.683Parasymphysis50.7110.695Parasymphysis, Condyle60.6680.634Angle, Condyle70.6410.273Condyle80.530.434Corpus, Symphysis90.5430.424Parasymphysis, Angle Comparison of patient-wise DICE scores, location and type (displaced and non-displaced) of fractures between volumetric (3D) and axial-slice (2D) validation of the AI model in test group. (Displaced fracture locations were shown in red, while non-displaced locations were shown in green) According to the case-based evaluation, non-displaced fractures and fractures located in the condylar region demonstrated lower DICE values compared with other fracture locations and fracture types. One of the lowest DICE values was observed in patient 3 in both the 2D and 3D models. In this patient, the fracture sites were located in the angle and condylar regions, and both fractures were non-displaced. Another low DICE value was observed in patient 7 (Fig. 7). Unlike patient 3, the 2D model demonstrated a DICE value of 0.273, whereas the 3D model demonstrated a higher DICE value of 0.641. In this case, the fracture was also located in the condylar region and was non-displaced. Cone-beam computed tomography (CBCT) images showing (A) the expert-annotated fracture region, (B) the two-dimensional (2D) slice-based artificial intelligence (AI)-detected fracture region in the condylar region at the axial plane, (C) the three-dimensional (3D) ground-truth segmentation annotated by experts, and (D) the three-dimensional (3D) AI-generated fracture segmentation Additional condylar fractures were observed in patients 5 and 6. Although the fracture regions were located in the condyle, the DICE values (0.711–0.695 and 0.668–0.634, respectively) were higher than the mean DICE values of the 3D and 2D models (0.605 and 0.492, respectively). This difference may be related to the displaced nature of the fractures compared with those observed in patients 3 and 7. Furthermore, fractures in similar anatomical regions (angle and condyle) were observed in patients 3 and 6; however, the fractures in patient 3 were non-displaced, whereas those in patient 6 were displaced. In patient 8 and patient 9 (Figs. 1, 2, 3, 4, 5 and 6), DICE values were slightly lower than the mean DICE values in both the 3D and 2D models. In these patients, fracture sites were located in relatively more detectable regions, including the parasymphysis, symphysis, angle, and corpus. Therefore, the relatively lower DICE values may be associated with the non-displaced nature of the fractures rather than fracture location alone. Higher DICE values were observed in patients 1 (Fig. 8), 2, and 4, with patient 4 demonstrating slightly higher values than patients 1 and 2. This finding may also be associated with fracture displacement, as patient 4 had displaced fractures, whereas patients 1 and 2 had non-displaced fractures. Cone-beam computed tomography (CBCT) images showing (A) the expert-annotated fracture region, (B) the two-dimensional (2D) slice-based artificial intelligence (AI)-detected fracture region in the parasymphysis region in the axial plane, (C) the three-dimensional (3D) ground-truth segmentation annotated by experts, and (D) the three-dimensional (3D) AI-generated fracture segmentation Discussion This study aimed to evaluate the diagnostic performance of an AI-based segmentation model for mandibular fracture analysis on CBCT images by comparing volumetric (3D) and 2D slice-based validation approaches. The findings suggest that volumetric validation may provide a more comprehensive representation of model performance compared with conventional 2D evaluation. The higher DICE, Jaccard, and recall values obtained with the 3D model, together with lower average surface distance and 95% Hausdorff distance, indicate a more consistent spatial representation of fracture morphology. When compared with the existing literature, the performance observed in the present study appears to follow a similar trend but remains relatively lower in absolute terms. Studies using larger datasets have reported precision and recall values exceeding 0.90 with YOLO-based models applied to CT images [12]. Similarly, Wang et al. and Mao et al. reported accuracy values above 95% for fracture classification tasks using CT data [8, 11]. In contrast, the CBCT-based segmentation model developed in this study demonstrated lower precision (0.61) and recall (0.66) values. This difference may not only be attributed to the inherent limitations of CBCT imaging, such as lower contrast resolution and increased noise, but also to differences in task design, as most previous studies have focused on detection or classification rather than voxel-level segmentation. In addition, the relatively limited sample size may have contributed to the observed differences in performance. Most AI-based mandibular fracture studies have been conducted on conventional CT data, which provide higher contrast resolution and reduced noise compared to CBCT [8, 11, 12]. However, CBCT offers finer voxel resolution and better spatial detail at a lower radiation dose, making it preferable for dentomaxillofacial applications despite increased scatter and limited soft-tissue contrast. These modality-related differences directly influence the delineation of fracture lines, particularly in regions such as the condyle, where thin cortices and overlapping anatomical structures complicate segmentation. In our test dataset, ten of the fourteen fractures were non-displaced. Non-displaced fractures often present with subtle and indistinct radiographic findings on CBCT, making them inherently more difficult to identify [27, 28]. This increased diagnostic difficulty may pose challenges for both clinicians and AI-based models. In the present study, a substantial proportion of fractures in the test dataset were non-displaced, which may have influenced the observed performance outcomes. Non-displaced fractures often present with subtle and indistinct radiographic features on CBCT images, making them more challenging to identify. This may introduce diagnostic difficulty for both clinicians and AI-based models. Accordingly, lower DICE values were observed for non-displaced fractures in the condylar region compared with other anatomical regions. In contrast, DICE values obtained from displaced condylar fractures were found to be comparable to those of other regions. Similarly, fractures located in the symphysis, parasymphysis, corpus, and angle regions demonstrated DICE values close to the overall mean, regardless of displacement status. The relatively lower performance observed in non-displaced condylar fractures may be associated with the unique anatomical and biomechanical characteristics of this region. Condylar fractures often occur under compressive forces and may present with minimal displacement and subtle cortical discontinuities, which may reduce their visibility on CBCT images [29, 30]. This reduced visibility may complicate both manual annotation and automated segmentation processes, thereby influencing performance metrics. Furthermore, the limited representation of such subtle fracture patterns in the training dataset may have constrained the model’s ability to effectively learn these features. Nevertheless, the model demonstrated a reasonable level of performance in other anatomical regions as well as in displaced condylar fractures. Also, the limited representation of such challenging cases in the training dataset may have restricted the model’s ability to adequately learn these fracture patterns. Consequently, the relatively lower precision and recall values observed compared with CT-based studies can be attributed not only to the smaller sample size and model complexity, but also to the inherent limitations of CBCT in visualizing subtle fracture lines. Furthermore, unlike most previous studies focusing on detection or classification tasks, the present study emphasized segmentation, which is inherently more demanding as it requires precise voxel-level delineation of fracture boundaries. The comparative analysis between volumetric and 2D slice-based validation approaches highlights an important methodological consideration. Conventional 2D slice-based evaluation may underestimate model performance, as it does not fully capture the contextual information between adjacent slices. In contrast, volumetric validation preserves spatial continuity and may allow a more comprehensive representation of the geometric relationships of fracture planes across three dimensions. In the present study, DICE values in the test group were observed to be higher with volumetric validation compared with 2D slice-based evaluation. For instance, in patient 7, the DICE score increased from 0.273 in the 2D slice-based approach to 0.641 with volumetric validation, which may suggest that volumetric validation has the potential to better capture the spatial continuity of fracture patterns. Similarly, in patient 3, although D values were low in both volumetric (0.269) and 2D slice-based (0.031) evaluations, the volumetric approach yielded relatively higher values. Previous CBCT-based studies, including those employing segmentation methods, have largely relied on 2D slice-based validation rather than true volumetric assessment, which may limit the evaluation of spatial fracture continuity and morphological complexity [3]. In this context, the approach adopted in the present study may provide a more comprehensive three-dimensional evaluation framework and may contribute to a more anatomically consistent representation of fracture morphology. The findings of the present study may further emphasize the potential importance of using volumetric ground-truth annotations when evaluating bone integrity in AI-based diagnostic models. Although the current results cannot be directly compared with large-scale meta-analyses, such as that of Ju et al., which reported pooled sensitivity and specificity values above 0.89 for facial fracture detection using various AI algorithms, the overall trend appears to be consistent with the notion that AI may achieve clinically meaningful diagnostic performance when supported by high-quality and spatially consistent data [7] In particular, the findings of the present study suggest that volumetric (3D) validation may provide a more comprehensive and anatomically representative evaluation of model performance compared with conventional 2D slice-based analysis, thereby indicating its potential clinical relevance for future AI-assisted diagnostic systems. However, it should be noted that most previous studies have relied on CT rather than CBCT data, and none have incorporated volumetric segmentation-based evaluation [6, 31]. In this regard, the present study may contribute to the literature by demonstrating the feasibility and potential diagnostic value of three-dimensional segmentation within CBCT-based workflows. Despite these findings, several limitations of the present study should be considered. The relatively small dataset and single-center design may limit the generalizability of the model. In addition, variations in CBCT scanners, exposure parameters, and reconstruction algorithms may influence image quality and model transferability [2, 32, 33]. Future research may benefit from increasing dataset diversity through multi-institutional collaborations and exploring advanced architectures, including hybrid models incorporating transformer-based components for improved feature extraction. Furthermore, clinical validation studies may be valuable to assess whether AI-assisted segmentation can reduce diagnostic time and inter-observer variability among clinicians. Among the limitations of this study is the inclusion of only fracture-positive cases in the dataset. This may have limited the direct evaluation of case-level specificity. Nevertheless, in line with the study design, model performance was assessed based on voxel-level segmentation accuracy, which allows for a detailed and sensitive evaluation of the model’s ability to delineate fracture lines. Future studies including non-fracture cases may contribute to a more comprehensive assessment of the model’s clinical performance. Furthermore, only linear fracture types were included in the present study, while comminuted fractures were excluded to ensure consistency in the annotation process. Although this approach improved standardization, it may limit the applicability of the findings to more complex fracture patterns encountered in clinical practice. In addition, the inherently thin and discontinuous nature of fracture lines in CBCT images presents challenges for voxel-wise segmentation, particularly in defining precise boundaries. Consequently, overlap-based metrics such as the DICE similarity coefficient may be sensitive to small spatial deviations and should be interpreted with caution. To address this limitation, surface-based metrics, including ASD and 95% Hausdorff distance, were also incorporated to provide a more comprehensive evaluation of segmentation performance. In summary, the findings of the present study suggest that volumetric (3D) validation may provide a more comprehensive and clinically meaningful assessment of AI-based mandibular fracture segmentation on CBCT images compared with conventional 2D slice-based evaluation. Although the overall segmentation performance was moderate and relatively lower than that reported in large-scale CT-based studies, the volumetric approach tended to demonstrate higher Dice similarity scores than the 2D method, potentially indicating improved spatial representation of fracture regions. These findings further suggest that volumetric segmentation approaches may contribute to the reliability of AI-assisted fracture assessment and support the potential role of 3D AI models as complementary tools in CBCT-based mandibular fracture evaluation. Acknowledgements None. Ethical statement The research protocol was approved by the Hacettepe University Health Sciences Research Ethics Committee. Abbreviations Artificial intelligence Cone-beam computerized tomography Dice similarity coefficient Average surface distance Authors’ contributions IA: Study design, idea/concept, writing the article, literature review. AY: Writing the article. EK: Data interpretation, and manuscript revision. EU: Data analysis and preparation of figures and Tables. AA: Idea/concept, writing the article. Funding None. Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate This study was approved by the Ethical Committee of Hacettepe University (Approval No: 2024/20–83). Due to the retrospective nature of the study and the use of anonymized data, the requirement for informed consent was waived. All procedures performed in this study involving human participants were conducted in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.