Artificial intelligence for detecting dental ankylosis in primary molars using panoramic radiographs—a retrospective study

1. Introduction

Dental ankylosis is an eruptive abnormality, histologically characterized as a fusion between dentin or cementum and bone resulting in the obliteration of the periodontal ligament space [1]. The fusion of bone and the tooth root leads to a vertical halt in tooth eruption, resulting in infraocclusion or possibly impaction of the affected tooth. This condition may occur during tooth eruption at any stage, whether before full emergence or after arrival at the occlusal plane [1, 2]. A significant incidence of dental ankylosis has been documented among children aged seven to eleven, with rates ranging from 1.3% to 38.5% [3, 4].

Two major theories explain dental ankylosis’ etiology: the first emphasizes local factors, while the second focuses on genetic factors, although no genes have yet been identified [5]. Abnormal activation of the periodontium in mice, which is known to increase Wnt signaling, may influence cancer stem cell behavior. A correlation has also been proposed between high levels of cellular cementum and alveolar bone accumulation and ankylosed teeth development [6].

The early detection of dental ankylosis is based on clinical (the presence of infraocclusion of the respective tooth, percussion testing, and loss of tooth mobility) and preclinical findings. Because ankylosis of primary molars is associated with severe clinical outcomes, such as infraocclusion of tooth, tipping of neighbouring teeth into infraocclusion space, which may cause arch space loss, exfoliation and eruption impairments for permanent successors and an asymmetric dental pattern [1, 7].

The ankylosed tooth is unable to exert the post-eruptive motion, thus it could be located in infraocclusion from 1 mm to complete retention beneath the gingival tissue [1, 8]. In severe ankylosis cases, percussion testing and assessment of tooth mobility are not applicable [7]. Therefore, radiographic assessment is crucial for diagnosing ankylosis. Bone and root surfaces will be visible on X-ray images, as well as the absence of periodontal space [1, 7].

Panoramic radiography is a vital imaging tool in dental practice for diagnosing and planning treatment for dental and maxillofacial conditions. A wide range of advantages are associated with this procedure, including its simplicity, affordability, minimal radiation exposure, reduced patient discomfort, and also the capacity to display a comprehensive bilateral view of oral structures [9, 10]. Further, interpreting dental radiographs in children can be more challenging and time-consuming. The mixed dentition, low-density areas in dental follicles, and possible pathological periapical inflammation may all be contributing factors [11, 12, 13].

Thus, disease identification can be transformed into a standardized image categorization task through artificial intelligence (AI), leveraging machine algorithms endowed with reasoning capabilities and cognitive functions, particularly relying on deep learning [10, 11]. The medical field is rapidly utilizing such technology. Particularly, image analysis using convolutional neural networks (CNNs) has demonstrated the potential to enhance practitioners’ reliability and CNNs learn from the statistical patterns present in imagery by iteratively processing pairs of images and corresponding image labels, typically supplied by medical professionals. Eventually, CNNs become capable of evaluating unseen data [14].

Artificial intelligence has the potential to reduce dentists’ clinical workload and enhance diagnostic accuracy, enabling earlier detection and more effective treatment alternatives, while mitigating errors stemming from fatigue and facilitating timely intervention, thereby conserving time [9, 11]. Deep learning-driven artificial intelligence algorithms have largely replaced specific conventional machine learning tasks in computer vision, such as classification, segmentation, and detection [10, 11, 15]. AI has been used in dentistry for a variety of purposes and across a wide range of modalities [10, 16, 17, 18, 19]. This includes detecting caries using bitewing or periapical radiographs [11, 17, 20], determining ectopic eruption [11] and maxillary sinusitis with panoramic radiographs [21], diagnosing osteoarthritis on cone-beam computed tomography (CBCT) [22], identifying lateral mandibular deviation through cephalograms [18], and detecting impacted supernumerary teeth [16]. Additionally, AI has effectively identified white spot lesions [23] and hypoplastic lesions on photographic images [24].

Dental ankylosis is challenging to diagnose. Although panoramic radiography is essential for its detection, image interpretation can be complex [7, 11]. Furthermore, traditional radiographic interpretation is inherently subjective and susceptible to human error, influenced by fatigue and experience levels. A missed or delayed diagnosis of dental ankylosis can lead to more severe clinical outcomes and potentially more complex and costly interventions later on [25, 26]. The use of AI in medical imaging, particularly deep learning models, has shown promise in improving diagnostic accuracy and efficiency [11, 15]. We identified a significant gap in the literature and in clinical practice regarding the use of AI for detecting dental ankylosis, particularly in primary molars using panoramic radiographs. This gap highlights the timeliness and relevance of this study.

This study aimed to address the limitations of traditional diagnostic methods such as time constraints, high costs, and potential human errors. The objective was to develop an accurate deep learning model for detecting ankylosed primary molars from panoramic radiographs.

In our study, the null hypothesis (H₀) posited that there is no significant difference between the performance of our AI model and chance performance. Thus, if H₀ is true, the model’s predictive ability is no better than random guessing. To test this hypothesis, we employed several evaluation metrics, including accuracy, AUC (Area Under the ROC (Receiver Operating Characteristic) Curve), sensitivity, and F1-Score, as previously used in deep learning or machine learning studies [10, 11, 16, 27, 28]. Each metric is statistically connected to H₀ by setting up confidence intervals and performing hypothesis tests.

2. Materials and methods

This retrospective study was approved by the Scientific Research Ethics Committee of Karadeniz Technical University Faculty of Dentistry (Protocol No. 2023/18, dated 19 June 2023). Given the retrospective nature of the analysis, informed consent from individual patients was not obtained. However, to ensure confidentiality and compliance with ethical standards, all patient data was anonymized before analysis. To ensure privacy, personal identifiers were removed, and data were aggregated. A retrospective analysis was conducted on pediatric patients (aged 4 to 15) who had panoramic radiographs taken between January 2015 and August 2023 at the Karadeniz Technical University Faculty of Dentistry (Trabzon, Turkey). Ankylosis was determined by analyzing panoramic X-ray images during their classification to determine the presence or absence of ankylosis (Fig. 1).

Fig. 1.Decision flow of the system (the relevant figure illustrates the overall workflow of the study methodology). VGG16: Visual Geometry Group 16-layer.

All panoramic radiograph images were taken with the Sirona Orthophos XG3 (Dentsply Sirona, Bensheim, HE, Germany) panoramic device (62 kV, 8 mA, 14.1 seconds) and saved in Joint Photographic Experts Group (JPEG) format. Sirona Orthophos XG3 (Dentsply Sirona, Bensheim, Germany) panoramic radiography device used in this study provides high-resolution imaging and is widely used in dentistry practices. The device uses a 0.5 mm X-ray tube and digital sensor technology to provide clear and detailed panoramic images. In addition, there is a special pediatric mode that reduces radiation dose for pediatric patients and allows detailed examination of teeth and jaw structures with various imaging programs [29]. Exclusion criteria in the study included radiographs with excessive blurriness, inadequate exposure, motion artifacts, or obscure or unclear anatomical structures. Alternatively, radiographs of children aged 4 to 15 years containing ankylosed primary molars met the inclusion criteria. Children who did not exhibit dental ankylosis were designated as controls (Table 1). The VGG Image Annotator (VIA) program was used to label teeth identified with dental ankylosis.

This table illustrates the gender and age distribution of patients in the study who had radiographs with and without ankylosed teeth.

Sample size was calculated based on a previous study [11] with alpha error = 0.05 and beta error = 0.20. Thus, it was decided to use at least 200 images. To avoid data loss (such as poor-quality images), at least 240 panoramic radiographs were used. Ultimately, due to the limited number of ankylosed primary teeth cases between January 2015 and August 2023, 268 panoramic radiographs were used: 98 cases and 170 controls.

Candidate digital panoramic radiographic images were screened based on medical records. Two experienced pediatric dentists, one with 6 years of experience (NY) and the other with 16 years of experience (TT) independently classified the images as cases or controls using the same monitor and environment. When the two dentists could not reach a consensus on a diagnosis, despite discussions, the image was excluded.

Due to the relatively limited number of X-rays in the target dataset, various data augmentation techniques were applied. Initially, a network structure was created and evaluated using the classical CNN (AT&T Bell Labs, Murray Hill, NJ, USA) method. Although data augmentation was applied, CNN performed relatively poorly. Therefore, transfer learning was used to improve performance. For the purpose of transfer learning, the performance of well-known and current neural network models, such as Inception V3 (Google, Mountain View, CA, USA), MobileNet V2 (Google, Mountain View, CA, USA) and Visual Geometry Group 16-layer (Visual Geometry Group-VGG, Oxford, UK) (VGG16), which had previously demonstrated high performance, was examined (Fig. 2) [30, 31, 32]. Images identified as ankylosis and those known to be non-ankylosis were processed sequentially for data augmentation and preprocessing Then, classification was performed using deep learning models. Finally, the classification performance was evaluated.

Fig. 2.Steps of the ankylosis detection methodology.

Experiments were conducted using Python (Phyton Software Foundation, USA) programming language on an i5 processor computer with Windows (Microsoft Corporation, Redmond, WA, USA) operating system.

2.1 Data preprocessing

Panoramic dental X-ray images contain extraneous spaces or bones visible outside the jaw. Therefore, a cropping process was applied to the left-right and top-bottom regions to obtain an image where only the jaw structure is visible. This eliminated unwanted areas and provided a clearer image. Visual examples of the cropping parameters used, along with before-and-after images, would enhance clarity and aid reproducibility (Fig. 3). Image pixels were normalized between [0–1] (Pnorm).

$\[ P_{norm}=\frac{P-P_{min}}{P_{max}-P_{min}} \]$

Fig. 3.Examples of images generated randomly for data purposes. (a) Randomly generated image examples for data augmentation copy not augmented. (b) Randomly generated image examples for data augmentation copy vertical flipped and rotated. (c) Randomly generated image examples for data augmentation copy channel shifted and rotated. (d) Randomly generated image examples for data augmentation copy vertical shifted and rotated. (e) Randomly generated image examples for data augmentation vertically flipped and channel shifted, and rotated. (f) Randomly generated image examples for data augmentation vertically shifted and rotated.

80% of the cropped images have been designated for training (10% for validation and 10% for testing). Separate datasets for training, validation, and testing were created. Data augmentation was then applied separately to each dataset. This ensured that augmented images from the training dataset did not appear in the validation and test sets. The data augmentation process (horizontal shift, random vertical shift, shear transformation, random zoom, horizontal flip, vertical flip, channel shift) was performed to generate 20 new images from each original image.

The augmented images were resized to 156 × 156 pixels. The choice of pixel size was based on both model performance and processing efficiency. Resizing standardizes image size, making them easier to process. Additionally, separate lists were created for training, validation, and testing: train_images, train_labels, valid_images, valid_labels, test_images and test_labels [9, 10, 11]. The data was shuffled to prevent memorization. This improved the model’s ability to generalize, leading to better results.

3. Results

3.1 Loss and accuracy of the models’ results

When examining the training loss and accuracy graphs for CNN, it was understood that the training loss value consistently decreased and converged to a loss value of around 0.25. Similarly, the training accuracy value steadily increased and converged to around 0.90. This indicates that the training data was well learned. A well-trained model was also shown in the loss and accuracy graphs for the VGG16 training set. However, after the 10th epoch, the loss and accuracy values did not significantly improve for the validation value. Additionally, Inception V3’s loss and accuracy graphs also indicated the model was well trained. However, after the 10th epoch, the loss and accuracy values did not significantly improve. The loss and accuracy graphs for the MobileNet V2 training set showed that the model learned the training data well (Fig. 4).

Fig. 4.Training accuracy and loss figure. (a) Training accuracy of each model; (b) Training loss of each model. The training accuracy and loss plots show that VGG16 and Inception V3 reach faster convergence and higher accuracy, whereas CNN demonstrates the lowest performance among the tested models. Notably, VGG16 achieves the best training results, featuring the highest accuracy and the lowest loss. CNN: Classical convolutional neural network; VGG16: Visual Geometry Group 16-layer.

A consistent decrease in loss and accuracy was not observed in the CNN validation set. Upon reaching the 25th epoch, the loss value increased, while the accuracy value decreased. In this case, overfitting began after the 25th epoch. Therefore, the best validation accuracy was 0.71. VGG16 validation accuracy and loss values did not consistently decrease after the first 10 epochs and even deteriorated. In this case, the model overfitted the training data. Therefore, the network model at the best point was saved. For Inception V3, the validation accuracy and loss values showed irregularities. While the training data consistently improved, the validation graph suggested that the Inception V3 model might have overfitted. Therefore, the network model at the best point was saved.

For MobileNet V2, instability in loss and accuracy values was observed in the validation data after the 5th epoch. The accuracy and loss values did not consistently decrease after the first 10 epochs and even showed adverse progress in some epochs. In the last 2 epochs, an increase was observed (Fig. 5). Essentially, no significant improvement in validation performance was noticed after the epoch of increasing gaps between training and validation, which indicates overfitting. This study employed the dropout method to prevent overfitting. To reduce the overfitting tendency of the model, the dropout value was initially set at 0.7 in all models. However, different dropout values (0.3, 0.5, 0.7) were tested and the most suitable value was determined.

Fig. 5.Validation accuracy and loss figure. (a) Validation accuracy of each model; (b) Validation loss of each model. The validation accuracy plot indicates VGG16 achieves the highest accuracy, but the validation loss plot reveals VGG16 also exhibits increasing loss, suggesting overfitting; in contrast, CNN shows the lowest validation accuracy among the models. CNN: Classical convolutional neural network; VGG16: Visual Geometry Group 16-layer.

3.2 Classification of models

Sensitivity, specificity, accuracy, AUC and F1-Score results obtained for all models used in the study are shown in Table 3. An analysis of the ROC curve (Fig. 6) and Table 3 values revealed that non-ankylosed images were well-detected, but the CNN model was found less effective at detecting ankylosed images. The CNN model detected ankylosis with a 0.28 sensitivity value.

Table 3.Comparison of the models according to accuracy, AUC score and specifity metrics.

	CNN	VGG16	MobileNet V2	Inception V3
Sensitivity	0.28	0.89*	0.77	0.79
Specificity	0.96	0.91	0.92	0.98*
Accuracy	0.71	0.90	0.86	0.91*
AUC	0.79	0.95	0.88	0.96*
F1-Score	0.42	0.87*	0.81	0.86
*It denotes the maximum values across all rows. AUC: Area Under Curve; CNN: Classical convolutional neural network; VGG16: Visual Geometry Group 16-layer.

Fig. 6.ROC curve and AUCs of the models. The validation accuracy plot shows that VGG16 attains the highest accuracy; however, the validation loss plot indicates that VGG16 also experiences increasing loss, signaling potential overfitting. In contrast, CNN exhibits the lowest validation accuracy among all the models. CNN: Classical convolutional neural network; VGG16: Visual Geometry Group 16-layer.

Apart from CNN, sensitivity/specificity/accuracy/AUC/F1-Score values in Table 3 were differentiated as followed: VGG16 (0.89/0.91/0.90/0.95/0.87), Mobile NetV2 (0.77/0.92/0.86/0.88/0.81) and Inception V3 (0.79/0.98/0.91/0.96/0.86). According to the highest and lowest values for each metric, the sensitivity values ranged from VGG16 (0.89) to CNN (0.28), the specificity values from Inception V3 (0.98) to VGG16 (0.91), the accuracy values from Inception V3 (0.91) to CNN (0.71), the AUC values from Inception V3 (0.96) to CNN (0.79), and the F1-Score values from VGG16 (0.87) to CNN (0.42). All the highest values when comparing the metrics among the groups were summarized in Table 3.

The table reveals the highest and lowest values for each metric: sensitivity values range from the lowest in CNN to the highest in VGG16, specificity values vary from Inception V3 as the highest to VGG16 as the lowest, accuracy values go from Inception V3 at the highest to CNN at the lowest, AUC values also range from Inception V3 as the highest to CNN as the lowest, and F1-Score values vary from VGG16 at the highest to CNN at the lowest (Higher value indicates better performance).

4. Discussion

Dental ankylosis is an eruption anomaly characterized by the fusion of the tooth root and alveolar bone, resulting in the obliteration of the periodontal ligament space. This condition is particularly prevalent among children aged 7 to 11 years, and early diagnosis is crucial to prevent serious complications such as vertical bone loss, dental asymmetry, and loss of arch space [1, 2, 3, 4, 5, 7].

Dental experts recommend combining clinical and radiographic examinations for ankylosed primary teeth diagnosis, with panoramic radiographs considered a routine approach [1]. Previous studies reported that CNNs can classify and segment periapical, bitewing, CBCT, and panoramic radiographs to detect caries and anatomical structures, with panoramic radiography being the most widely used technique [35, 36].

Due to the advantages of AI technologies, dentists are increasingly using them for diagnosis, which improves diagnostic efficiency and enables earlier diagnosis and intervention [37]. Integration of AI-enabled systems into clinical applications optimizes workflow and enhances cost-effectiveness in dental practices. With AI-based imaging and diagnostic systems, the time spent on each patient can be reduced, improving clinical efficiency. Furthermore, AI-based applications in specific areas such as caries detection reduce treatment costs by minimizing missed lesions. AI integration also streamlines administrative tasks like scheduling appointments, updating patient records, and responding to routine inquiries. By automating these processes, clinical staff can concentrate on addressing more complex patient needs, ultimately saving time and resources [38].

To determine the best AI model, various models’ suitability has been tested against study methods and disease diagnosis criteria. While more advanced models exist than those used in this study, they require higher computational power. This study used models that can be easily applied to clinical settings and could potentially improve their performance with the addition of more images [10, 11, 39].

Dentists’ decision-making during diagnosis has been facilitated using machine learning, with deep learning methods being employed to detect dental caries, supernumerary teeth, dental anomalies, and other dental conditions [10, 11, 40]. In this study, the effectiveness of AI, specifically the deep learning approach, in detecting ankylosis on dental panoramic x-ray images was evaluated. Aside from the sensitivity and F1- Scores of CNN, the H₀ was rejected (when considering 50% as borderline) indicating that AI-assisted deep learning models performed better than random predictions.

In earlier studies, a Region of Interest (ROI) was extracted from the original X-ray image for the deep learning model to predict diagnosis [11, 16, 20]. In our study, cases of ankylosis are localized to specific regions within the images, suggesting that training on the entire image set may not effectively learn the diagnosis of ankylosis. Therefore, during deep learning, cropping was performed on both ankylosed and non-ankylosed images to focus on ROIs. This approach aims to enable deep learning to easily identify the most relevant regions for classification, thereby enhancing the model’s ability to accurately detect cases [11, 41].

It is often easier to train direct classification systems than segmentation, which requires human experts to delineate contours or record coordinates for every sample. However, human annotation can introduce errors and biases. Probability estimates in each category can be affected by these errors. Despite significant differences in diagnostic probabilities, a positive diagnosis could still be made [11, 42]. However, in comparison to segmentation-based models, classification-based models can be further improved by retraining them with manually labeled data [11].

In this study, similar to previous research, four different structure models—CNN, VGG16, Inception V3 and MobileNet V2 were used, and metrics such as accuracy, AUC score, sensitivity, specificity, and F1-Score were evaluated [10, 11, 16, 35, 39, 43]. As observed in our study, training loss values decreased, which is consistent with the results of a study using deep learning algorithms to detect supernumerary teeth in dentistry [10]. Thus, based on the training accuracy and loss values obtained in this study, it was concluded that all models could be generally trained to acceptable levels. However, the validation loss and accuracy values differed from those reported in the literature [10]. Alongside these differences, overfitting was observed in all models at various epochs, with optimal validation values identified, but no significant improvement in validation performance after the onset of overfitting epochs. Study designs, sample types, and sizes may explain the variability in results observed in this study compared to previous findings [10].

Sensitivity is considered a crucial measure in disease screening to avoid false-negative diagnoses [10]. As observed by the sensitivity results, CNN exhibited the lowest value (0.28) compared to VGG16 (0.89), MobileNet V2 (0.77) and Inception V3 (0.79). Because of the low sensitivity of the condition, there is a risk of missing truly ankylosed cases when using the CNN model.

Findings in the dental literature report sensitivity values for VGG16 at 0.58, 0.74 and 0.85 and specificity values at 0.70, 0.79 and 0.83. For Inception V2 and/or V3, sensitivity values were reported at 0.82 and 0.90 and a specificity value at 0.78. For MobileNet V2, sensitivity values were 0.62 and 0.89, and specificity was 0.85 [10, 11, 23, 39, 44]. Similarly, our sensitivity and specificity findings were consistent with previous findings [10, 11, 23, 39, 44]. According to Mine et al. [10], when testing for presence of supernumerary teeth in early mixed dentition using deep learning, VGG16 had the highest sensitivity value (0.85), with all models, including AlexNet, VGG16, and Inception V3, showed high sensitivity values. In our study, VGG16 model had a higher sensitivity value (0.89) than Inception V3 (0.79). Also, MobileNet V2 showed a slightly lower value (0.77) than Inception V3. To avoid missing true positives, it is advantageous to use Inception V3, specifically VGG16, in comparison to MobileNet V2.

Higher accuracy values indicate the model’s overall ability to produce correct results [16, 36] and higher AUC values demonstrate the model’s ability to perform effective classification [45]. In this study, Inception V3 exhibited slightly higher accuracy (0.91) and AUC (0.96) values than VGG16 (0.90 and 0.95), respectively. MobileNet V2 (0.86 and 0.88) and CNN (0.71 and 0.79) exhibited lower values than both Inception V3 and VGG16. Compared to MobileNet V2 and CNN, Inception V3 and VGG16 may offer better detection of true positive and negative cases.

A high F1-Score value indicates the model’s overall high performance and suggests balanced classification [11]. Overall, VGG16 achieved the highest performance in this metric, while Inception V3 showed very close F1-Score performance. VGG16 and Inception V3 models are both capable of classifying ankylosed and non-ankylosed images with high performance and balance. VGG16 and Inception V3 models generally excel in performance over MobileNet V2 due to their more complex architectures. These deeper models, however, may be overfitted due to limited data sizes. Even though MobileNet V2 is slightly less performant, it is a suitable option for scenarios that require a lighter model.

According to the literature, VGG16 achieved AUC values of 0.73, accuracy values of 0.57, 0.72, 0.89, and a F1-Score of 0.56; Inception V2 and/or V3 achieved accuracy values of 0.80 and 0.88; and MobileNet V2 achieved accuracy values of 0.62 and F1-Scores of 0.57 and 0.87 [10, 11, 16, 39, 44]. Based on the combined results, the AUC and F1-Scores are aligned with the literature, indicating that VGG16 and Inception V3 could be used to detect ankylosed primary teeth from panoramic images. However, other studies using similar models have reported metric rankings that differ from ours [16, 39]. Model performance may vary based on research designs, explaining the discrepancy. Future developments in AI-assisted diagnostic models for dental ankylosis could greatly benefit from hybrid approaches and ensemble learning techniques. By combining the strengths of different deep learning architectures, hybrid models can enhance diagnostic accuracy by leveraging complementary feature extraction capabilities. Likewise, ensemble learning techniques that aggregate predictions from various models can enhance classification reliability and minimize model bias. By integrating these techniques, AI-based diagnostic systems can deliver more reliable and consistent results and ultimately aid clinical decision-making in pediatric dentistry. These approaches should be assessed in future studies in larger datasets and real-world clinical settings to optimize their applicability [46, 47].

For AI systems to be used effectively and reliably, clinical staff must undergo appropriate training and calibration processes. Dentists and auxiliary staff should be trained on AI-based tools principles, limitations and interpretation skills. Training ensures that AI systems are used effectively and safely to ensure patient safety and diagnostic accuracy. In addition, AI models need regular calibration and updates to perform optimally. This process ensures the models are constantly updated, ensuring accuracy and adaptability. AI integration into dental practices requires both technology infrastructure and the human factor to work in harmony [48].

Finally, this study explores the use of artificial intelligence to detect dental ankylosis in primary molars through the analysis of panoramic radiographs, highlighting its potential for early diagnosis, enhanced accuracy, and improved efficiency in dental care. However, this study has various limitations, including diagnosing ankylosis solely based on panoramic radiographic images, increasing case numbers statistically, and overfitting in some models. Possible biases in this study may be due to specific patient demographics, insufficient data diversity in data used in model training, variability in expert labeling, and potential model tuning biases. The limited sample size and use of panoramic radiographs alone could negatively influence dental ankylosis diagnosis. Future research should utilize three-dimensional imaging methods such as CBCT and larger sample sizes for conditions such as ectopic eruptions or hypoplasia lesions. Such adaptations could further enhance diagnostic accuracy and efficiency across a wide spectrum of dental and maxillofacial anomalies, providing valuable tools for practitioners and addressing gaps in current clinical practices. The versatility and relevance of the model extend beyond dental ankylosis in this suggestion.

5. Conclusions

Artificial intelligence (AI) methodologies offer significant potential in pediatric dentistry by reducing clinical workload and enhancing diagnostic accuracy, particularly in the early detection of dental ankylosis. This study emphasizes the effectiveness of deep learning models—specifically VGG16, Inception V3, and MobileNet V2—in identifying ankylosis from panoramic radiographic images. Compared to CNN, VGG16 showed relatively greater sensitivity and accuracy in detecting ankylosed images, highlighting the importance of model selection in dental diagnosis. Based on these findings, applying advanced AI algorithms to diagnostic workflows could enhance precision and patient outcomes by enabling earlier interventions. The limitations of certain models, such as CNN, should also be considered by dental practitioners, who may need to further optimize these models for reliable clinical use.

Availability of data and materials

The data presented in this study are available on request from the corresponding author. The data are not publicly available for ethical purposes.

Author contributions

NY, MHB and TT—conceived the ideas. NY, MHB and MA—collected the data. NY, MHB, MA and TT—analysed and checked the data; and led the writing.

Ethics approval and consent to participate

The present retrospective study received approval from Scientific Research Ethics Committee of Karadeniz Technical University Faculty of Dentistry (Protocol No. 2023/18). The informed consent is exempted according to rules of the Scientific Research Ethics Committee of the Faculty of Dentistry at Karadeniz Technical University.

Acknowledgment

Not applicable.

Funding

This research received no external funding.

Conflict of interest

The authors declare no conflict of interest.