Deep learning-based prediction of buccal, middle and palatal orientations of impacted maxillary canines using panoramic radiographs: a pilot study

Abstract

Objective

The aim of this pilot study was to compare the performance of deep learning models in detecting and classifying the buccal, middle, and palatal orientations of impacted maxillary canines using panoramic radiographs.

Methods

A total of 200 panoramic radiographs were collected from patients’ records with a history of unilateral and/or bilateral impacted maxillary canines. The images were preprocessed for resizing, bit depth, and file format. Data augmentation was performed by horizontal flipping to increase the total dataset size from 200 to 400 images. The images were randomly divided into three subsets as 80% training, 10% validation and 10% testing. The annotation was made by two experienced dentists. The dataset was trained and tested using two state-of-the-art deep learning models based on the YOLO11x and YOLO12x architectures. Model performance was evaluated using standard metrics, including precision, recall, F1-score, mAP@0.5, and mAP@0.5:0.95. In addition, confusion matrices and visual predictions from the test images were analyzed to assess class-specific performance and localization accuracy.

Results

The YOLO11x model demonstrated superior performance compared to the YOLO12x model. During the training phase, the mAP@0.5 values were 0.723 and 0.639, respectively, while in the testing phase, they were 0.468 and 0.435, respectively. While both models achieved accurate localization in a limited number of test samples, the YOLO12x model exhibited a higher rate of false negatives. Visual inspection revealed that palatal orientations were classified more reliably, whereas classification performance was lower for buccal orientations and particularly limited for middle orientations.

Conclusion

This pilot study presented certain challenges in accurately classifying the spatial orientation of impacted maxillary canines due to methodological limitations (e.g., small sample size, preliminary nature). While satisfactory performance was achieved for palatal orientations, reduced classification accuracy and frequent misclassifications were observed for buccal and middle orientations. Given the inherent limitations of panoramic imaging, future studies incorporating larger datasets and 3D imaging techniques would be warranted to improve classification accuracy and clinical applicability.

Introduction

Impacted canine is a significant orthodontic condition that frequently poses challenges to clinicians and requires comprehensive treatment planning in daily practice [1]. Among impacted teeth, the maxillary canine is the second most commonly seen, following the mandibular third molars [2]. The etiology involves multiple factors, including genetic predisposition, tooth size-arch length discrepancies, persistence or premature loss of deciduous canines, and various pathological conditions [1, 2]. The prevalence of maxillary impacted canines varies across populations, but generally ranges between 1% and 3% [2, 3]. Impacted canines are approximately twice as common in females compared to males [2], and unilateral impaction is observed more frequently than bilateral one [2, 3]. Impacted canines can potentially cause several complications, including adjacent root resorption, cyst formation, and both dental and skeletal malocclusions [4, 5]. Among Turkish population, bilateral impacted canines are more commonly associated with skeletal Class III malocclusions [5]. Furthermore, the eruption of the maxillary canine plays a critical role in shaping the dental arch form, occlusal guidance, and overall facial aesthetics [1].

Early diagnosis and accurate localization of impacted canines are essential for developing an effective treatment strategy and minimizing potential complications. For buccally displaced maxillary canines, corrective treatment approaches include surgical uncovering and traction. For palatal impaction, surgical uncovering and a variation of open exposure (Gopex technique), spontaneus eruption and traction may be considered as alternative treatment options [6]. Depending on the position, clinical management may require invasive procedures such as full-thickness flap elevation and/or bone removal [7]. The use of temporary anchorage devices can eliminate potential risks [8].

Panoramic radiographs have traditionally been considered the initial imaging modality in diagnosing impacted canines due to their low radiation exposure, wide availability, and cost-effectiveness [9]. However, panoramic imaging presents notable limitations, primarily due to its two-dimensional (2D) nature and the superimposition of anatomical structures, which can make accurate spatial localization of impacted canines challenging. Although panoramic radiographs are useful for valuating the mesio-distal position and vertical depth of impacted teeth, they are often insufficient in differentiating between labially and palatally impacted canines [10]. In this context, three-dimensional (3D) imaging techniques, such as cone-beam computed tomography (CBCT), overcome these limitations by providing precise spatial localization, identifying root resorption in adjacent teeth, and assessing their relationship with surrounding anatomical structures [11]. However, the routine use of CBCT is limited by its higher radiation dose and cost [12]. Therefore, the present study focuses on exploring the potential of deep learning-based approaches for the localization and classification of impacted canines using panoramic radiographs, which remain a widely used and more accessible imaging method.

In recent years, artificial intelligence (AI), particularly deep learning, has made groundbreaking advances in the field of medical imaging. Among these, the You Only Look Once (YOLO) family of algorithms has gained widespread application in medical object detection due to its high accuracy and real-time processing capabilities [13, 14]. YOLO models automatically detect and classify specific structures in images and have evolved significantly, with newer versions offering improved performance, speed, and robustness, making them particularly suitable for clinical applications [13]. An overview of systematic reviews concluded that recent AI models demonstrated high accuracy (over 90%) in detecting caries, periapical lesions, bone loss, and tooth identification, highlighting AI’s strong potential to assist dentists [15]. YOLO-based approach, in particular, has been successfully implemented for the detection of localization of all impacted teeth and third molar on panoramic radiographs, thereby enhancing clinical workflows through automation and standardization of diagnostic processes [16].

Accurate determination of the spatial orientation of impacted canines is clinically crucial, as buccal and palatal impactions require different surgical and orthodontic management. Despite advances in AI, current chatbot- and image-based systems have not achieved satisfactory diagnostic reliability, underlining the need for deep learning–based methods [17]. Alenezi et al. (2025) compared various convolutional neural network (CNN) architectures and reported that adding an automatic cropping step increased the AUC-ROC value of the GoogLeNet-based model to 0.99 in distinguishing impacted and normal canines. However, these approaches focus solely on binary classification (impacted/non-impacted) and do not evaluate clinically important distinctions such as orientation [18].

To the best of our knowledge, studies applying the latest YOLO11x and YOLO12x deep learning architectures, particularly YOLO12x, to dental radiographs were scarce in the literature. Therefore, this study aimed to evaluate the performance of these next-generation YOLO models in detecting buccal, middle and palatal orientations of impacted canines using panoramic radiographs. The null hypothesis was that there would be no difference between YOLO11x and YOLO12x deep learning models in classifying buccal, middle, and palatal orientations of impacted maxillary canines on panoramic radiographs.

Materials & methods

Collection of panoramic radiographs

This study was conducted in a retrospective design in accordance with the ethical principles of the Declaration of Helsinki regarding medical research with human subjects. The study protocol was approved by the Ethics Committee of Pamukkale University (Date/Protocol No: 12.08.2025/15). Patient data were obtained retrospectively from archival records, and all procedures complied with principles of confidentiality and anonymity.

The archival records of the Department of Orthodontics between 2020 and 2023 were reviewed to identify patients with impacted maxillary canines. A total of 2003 patient records were reviewed, comprising 1086 females and 917 males, aged 13 to 47 years. Selection criteria were determined as follows: (I) availability of pre-treatment digital panoramic radiographs; (II) presence of unilateral and/or bilateral impacted maxillary canines; (III) patients undergoing fixed orthodontic treatment without prior early intervention; (IV) availability of intraoral photographs and clinical records both before and after orthodontic treatment to confirm the final position of impacted canine.

Digital records were excluded if they had poor image quality or artifacts, large cystic lesions or other pathologies associated with the impacted teeth, or insufficient clinical information regarding the final position of the impacted canine. All panoramic radiographs were obtained using the same dental imaging device (OP200D; Instrumentarium Corp., Imaging Division, Tuusula, Finland) with standardized scanning parameters (60 kV, 6.3 mA, 14.1 s exposure time). The final dataset included 200 panoramic radiographs (115 females and 85 males) with a mean age of 14.47 ± 0.17 years.

Pre-processing and ground truth annotation

A total of 200 panoramic radiographs were used in this study. Since the original file extensions of the images were different, they were first converted to .jpg format to ensure consistency. Then, the images were resized to 1024 × 532 pixels using the bilinear interpolation method, which was the closest value to the average pixel ratio in the data set, and at the same stage, they were converted to 8-bit grayscale format to standardize them.

Annotation was performed using bounding boxes via the browser-based makesense application (https://www.makesense.ai/). All images were categorized into “Buccal”, “Middle”, or “Palatal” orientations by two experienced dentists who are experts in their field, based on the clinical reports obtained after the eruption of impacted teeth. Every annotation was finished with consensus; no mathematical calculation was performed. In this respect, buccal orientation meant that maxillary canine was located towards the lip (labial); palatal orientation meant that maxillary canine was located toward the palate; middle orientation meant that tooth was positioned centrally, erupting from the crest of the alveolar ridge.

To increase data diversity, the data augmentation process was applied after the dataset was split into training (80%), validation (10%), and test (10%) sets. Within each subset, all images were horizontally mirrored using a Python-based script, thereby doubling the number of examples in each group and increasing the total data count from 200 to 400. This approach was chosen to prevent data leakage and preserve the independence of the test group by ensuring that each subset contained its own augmented samples. Horizontal flipping creates positional and structural variations without distorting the anatomical structure, thereby improving the model’s ability to generalize spatial features. As this was a pilot study, no power analysis was performed; the dataset was limited to the available 200 radiographs that fulfilled the inclusion criteria to assess methodological feasibility.

Figure 1 illustrates how the annotation process is performed on a panoramic radiograph and its augmented form.

Fig. 1.

Panoramic radiograph annotated with a bounding box using makesense, representing an impacted canine with defined orientation and its augmented version

Transfer learning of deep learning models

The pretrained YOLO11x and YOLO12x architectures were employed using a transfer learning approach to classify maxillary impacted canines according to their spatial orientation on panoramic images. Both models were among the most recent and advanced versions of the YOLO family and were trained on the COCO (Common Objects in Context) dataset prior to adaptation for dental image classification.

For this study, the models were reconfigured with three output classes (buccal, middle, and palatal). The extra-large (x) versions of both models were selected due to their deeper network architecture, which allows more efficient learning of complex radiographic features and higher classification accuracy for fine anatomical structures.

After labeling, the images were divided into training, validation, and test subsets, then compressed into a.zip file and uploaded to Google Drive. Within each subset, all images were horizontally flipped to increase dataset diversity. In addition, the built-in data augmentation techniques of the YOLO architecture, including mosaic, blur, CLAHE and random grayscale, were automatically applied during training. Model training was conducted on Google Colab Pro + using an NVIDIA A100 GPU. The training process employed the default YOLO parameters with the AdamW optimizer, where the learning rate was approximately 0.0014, momentum 0.9, and weight decay 0.0005. The mini-batch size was set to 4 and the number of epochs to 600, as this relatively small batch size allows for more frequent weight updates in small datasets, thereby supporting more accurate model learning.

Performance evaluation metrics

The models’ performance was evaluated separately during training and testing stages. The confusion matrices were generated to compare the model predictions with the ground-truth labels annotated by expert dentists, and both the confusion matrices and other performance visualizations were automatically produced by the YOLO framework. This allowed us to assess how accurately each model classified buccal, middle, and palatal orientations.

The basic criteria used in the evaluation process were True Positive (TP), False Positive (FP), False Negative (FN), and True Negative (TN) [19]. TP demonstrated the actual positives predicted correctly by the model; FP indicated the incorrect prediction of a class that did not actually exist; FN identified the classes that the model missed but actually existed; and TN which was not considered in the context of object detection.

To evaluate the model’s classification success more comprehensively, the metrics precision (Eq. 1), recall (Eq. 2), F1-score (Eq. 3), and mean Average Precision (mAP) (Eq. 5) were calculated. Precision represented the ratio of correct detections to all positive predictions; recall represented the ratio of correctly detected TPs to the total number of TPs and total number of FNs; and the F1-score represented the harmonic mean of these two values. The model’s detection performance for each class was measured by Average Precision (AP), which represented the area under the precision-recall curve (Eq. 4). These values were averaged to obtain mAP, which indicated the overall performance of the model. In this context, mAP was calculated in two different ways: mAP@0.5, when the Intersection over Union (IoU) threshold was fixed at 0.5; mAP@0.5:0.95, when the IoU threshold was changed in increments of 0.05 from 0.5 to 0.95.

The class-wise and overall performance of the model were evaluated using the training/validation losses reported by YOLO and the epoch-wise evolutions of precision, recall, mAP@0.5 and mAP@0.5:0.95.

1

2

3

4

5

Results

Training and testing performance results

During the model training process, early stopping occurred for both architectures, and not all epochs were completed. The YOLO11x model achieved its best performance at epoch 71 and completed training after a total of 171 epochs, requiring approximately 40 min (0.668 h). The best model file (best.pt) obtained as a result of training was 114.4 MB in size. On the other hand, the YOLO12x model performed best at epoch 49, completed training in a total of 149 epochs, and requiring approximately 46 min (0.762 h). The size of the best model file obtained for YOLO12x was 119.1 MB. Using the models obtained, experiments were conducted on test data that the system had never seen before in the Google COLAB environment.

The training-validation performance graphs for both models are shown in Fig. 2. These graphs revealed that the YOLO11x model exhibited a steady decrease in loss and a consistent improvement in success metrics, suggesting a more stable learning process. Conversely, the YOLO12x model demonstrated an increase in validation loss and greater metric fluctuations toward the later training stages, indicating a higher tendency toward overfitting. Based on these observations, YOLO11x provided a more balanced and reliable training performance.

Fig. 2.

The training-validation graphs obtained for YOLO11x (top graph) and YOLO12x (bottom graph)

The confusion matrices generated for training and testing processes are displayed in Fig. 3. The analyses revealed that the YOLO11x model demonstrated superior performance in both class discrimination and object detection compared to YOLO12x. For the training data, YOLO11x achieved high classification accuracy for the palatal orientation, while maintaining a low misclassification rate across all classes. In the test data, meaningful accuracy rates were achieved for both the buccal and palatal classes; however, inter-class confusion persisted, particularly in the middle orientation. By contrast, YOLO12x model misclassified a significant number of actual object samples as background, resulting in a high false negative rate. Specifically, in the test data, 19 middle-class and 16 buccal-class examples were incorrectly predicted as background, suggesting insufficient learning of class-level distinctions and lower detection reliability. As a result, the YOLO11x model demonstrated a more balanced and stable classification performance in terms of class distinction and background error control.

Fig. 3.

Confusion matrices obtained during training and testing processes (a) YOLO11x training (b) YOLO11x testing (c) YOLO12x training (d) YOLO12x testing

The training and testing performance results are demonstrated in Tables 1 and 2, respectively. When examining Table 1, the YOLO11x model generally achieved higher success compared to the YOLO12x model. Across all orientations, the YOLO11x model achieved 0.653 precision, 0.802 recall, 0.720 F1-score, 0.723 mAP@0.5, and 0.458 mAP@0.5:0.95. In the same scope, the YOLO12x model remained at lower values with 0.565 precision, 0.683 recall, 0.618 F1-score, 0.639 mAP@0.5, and 0.427 mAP@0.5:0.95. When examined by class, the YOLO11x model achieved 0.704 precision, 1.000 recall, 0.826 F1-score, 0.933 mAP@0.5, and 0.618 mAP@0.5:0.95 in the palatal orientation. In terms of middle orientation, it achieved 0.851 precision, 0.479 recall, 0.613 F1-score, 0.597 mAP@0.5, and 0.350 mAP@0.5:0.95. It also offered a more balanced performance with 0.403 precision, 0.929 recall, 0.562 F1-score, 0.639 mAP@0.5, and 0.405 mAP@0.5:0.95 in terms of buccal orientation.

Table 1.

Training performance results

Train	Class	Images	Instances		Precision	Recall	F1-score	mAP@0.5	mAP@0.5:0.95
YOLO11x	Buccal	14		14	0.403	0.929	0.562	0.639	0.405
	Middle	12		12	0.851	0.479	0.613	0.597	0.350
	Palatal	14		14	0.704	1.000	0.826	0.933	0.618
	All	40		40	0.653	0.802	0.720	0.723	0.458
YOLO12x	Buccal	14		14	0.393	0.714	0.507	0.564	0.345
	Middle	12		12	0.664	0.333	0.444	0.528	0.378
	Palatal	14		14	0.637	1.000	0.778	0.826	0.559
	All	40		40	0.565	0.683	0.618	0.639	0.427

Table 2.

Testing performance results

Test	Class	Images	Instances	Precision	Recall	F1-Score	mAP@0.5	mAP@0.5:0.95
YOLO11x	Buccal	14	14	0.270	0.898	0.415	0.436	0.311
	Middle	20	20	0.288	0.400	0.335	0.359	0.225
	Palatal	14	14	0.567	0.571	0.569	0.533	0.327
	All	40	48	0.375	0.623	0.468	0.443	0.288
YOLO12x	Buccal	14	14	0.296	1.000	0.457	0.419	0.285
	Middle	20	20	0.386	0.900	0.540	0.392	0.263
	Palatal	14	14	0.388	0.786	0.520	0.494	0.278
	All	40	48	0.357	0.895	0.510	0.435	0.275

On the other hand, the YOLO12x model achieved 0.637 precision, 1.000 recall, 0.778 F1-score, 0.826 mAP@0.5, and 0.559 mAP@0.5:0.95 in the palatal orientation. It achieved 0.664 precision, 0.333 recall, 0.444 F1-score, 0.528 mAP@0.5, and 0.378 mAP@0.5:0.95 in terms of middle orientation. It also achieved 0.393 precision, 0.714 recall, 0.507 F1-score, 0.564 mAP@0.5, and 0.345 mAP@0.5:0.95 in the buccal orientation. Overall, YOLO11x achieved higher classification stability and better adaptation to the training data, particularly in the palatal orientation. Furthermore, when overall F1 scores were evaluated, the YOLO11x model demonstrated a more balanced performance between precision and recall (0.720 vs. 0.618), thus achieving more robust success across classes.

When examining the performance on the test data in Table 2, the YOLO11x model achieved higher values in terms of overall accuracy and average success metrics; the model’s test results across all classes were 0.375 precision, 0.623 recall, 0.443 mAP@0.5, and 0.288 mAP@0.5:0.95, respectively. In contrast, the YOLO12x model achieved 0.357 precision, 0.895 recall, 0.435 mAP@0.5, and 0.275 mAP@0.5:0.95. While YOLO12x showed a clear advantage in the recall metric, YOLO11x was more successful in terms of precision and mAP values. In terms of F1 score, the YOLO12x model (0.510 vs. 0.468) was slightly ahead. This showed that high recall values partially compensated for low precision. When examined by class, the YOLO11x model offered remarkable accuracy in the palatinal orientation with 0.567 precision, 0.571 recall, 0.569 F1-score, 0.533mAP@0.5, and 0.327 mAP@0.5:0.95. It achieved 0.270 precision, 0.898 recall, 0.415 F1-score, 0.436mAP@0.5, and 0.311 mAP@0.5:0.95 in the buccal orientation. It also achieved 0.288 precision, 0.400 recall, 0.335 F1-score, 0.359 mAP@0.5, and 0.225 mAP@0.5:0.95 in terms of middle orientation. The test performance of the YOLO12x model was 0.296 precision, 1.000 recall, 0.457 F1-score, 0.419 mAP@0.5, and 0.285 mAP@0.5:0.95 in the buccal orientation. It achieved 0.386 precision, 0.900 recall, 0.540 F1-score, 0.392 mAP@0.5, and 0.263 mAP@0.5:0.95 in terms of middle orientation. It was measured as 0.388 precision, 0.786 recall, 0.520 F1-score, 0.494 mAP@0.5, and 0.278 mAP@0.5:0.95 in the palatal orientation. These findings indicated that while YOLO12x could detect more examples in the test data, YOLO11x offered a more balanced and reliable performance with lower false positive rates.

When all results were examined, it could be said that palatal orientation was detected better than others in panoramic images, partial success was achieved in the buccal class, but limited discrimination performance was obtained in the middle class. Considering the overall performance, the YOLO11x model provided more reliable results than YOLO12x in the classification of these three orientations. These results indicated that YOLO11x achieved more consistent localization performance, especially for palatal impactions.

Visual results

When performance evaluation criteria were examined, it was observed that the YOLO11x model yielded more successful results compared to the other model. When the test images belonging to this model were examined, it was observed that the system provided high accuracy in detecting the localization of impacted maxillary canine teeth and achieved correct localization in all three sample cases illustrated in Fig. 4. However, when the test metrics were evaluated overall, it was understood that the performance in terms of class separation was limited and that uncertainties between classes persisted, especially in the buccal and middle orientations.

Fig. 4.

True positive predictions and localizations of the YOLO11x model. Blue: Buccal orientation, Cyan: Middle orientation, Grayish White: Palatal orientation

In visual evaluations, it was observed that the model provided high discriminative power by anatomically correctly positioning teeth located in the posterior maxillary region in examples belonging to the palatal orientation. However, it was observed that confusion in class separation may occur in some images due to the more uncertain position of the tooth in the buccal and especially middle classes.

Discussion

This study comprehensively evaluated the performance of the YOLO11x and YOLO12x deep learning models developed for classifying the spatial orientation of impacted maxillary canine on panoramic radiographs. To the best of our knowledge, this was among the first studies to explore the diagnostic potential of the latest YOLO architectures, particularly YOLO12x, for dental imaging applications. Recently, the researchers have emphasized that AI-based studies on impacted canines should incorporate expert-labeled datasets and relevant clinical parameters to improve model accuracy and reliability [20]. Consistent with this recommendation, our study utilized a dataset annotated by two experienced dentists and included clinical confirmation of impacted maxillary canine positions, thus enhancing the validity of the training and testing data.

Previous Al-based classification studies have demonstrated that panoramic radiographs can be used effectively for detecting impacted canines [21, 22]. In a recent study based on the deep learning-based automatic landmark localization, Zhang et al. [23] reported that the accuracy of predictive models such as those by Arnautska [24] and Margot et al. [25] showed good performance compared to Alqerban et al. [26]. In this study, we did not use predictive landmark-based models; instead, we directly classified impacted canine based on ground-truth positional data derived from clinical records. This approach provided a more clinically reliable evaluation of the models’ diagnostic performance.

In the present study, YOLO11x consistently outperformed YOLO12x in terms of precision, balanced performance, and lower misclassification rates. Therefore, the null hypothesis was rejected. The lower performance of YOLO12x may be attributed to its deeper architecture and larger number of parameters, which may limit its generalization capability on small datasets. While YOLO12x achieved higher recall values, this was accompanied by a significant increase in false positives, making it less reliable for clinical applications. Importantly, both models demonstrated their highest accuracy in palatal orientations, moderate success in buccal orientations, and limited classification performance in middle orientations, where overlapping anatomical structures reduced spatial differentiation. This difficulty was not merely a limitation of the model but also a fundamental problem inherent in 2D imaging.

A notable finding was the performance gap between training and testing results for both models. The sharp drop in test performance suggested potential overfitting to the training data or increased anatomical variability in the test set. Besides, higher recall combined with lower precision during the testing phase suggested that both models tended to overestimate positive classes, potentially increasing the false positive rate in clinical decision-making. The low mAP@0.5 values (YOLO11x: 0.443, YOLO12x: 0.435) during testing indicated that these models were not yet adequate enough for clinical routine.

Our findings also highlighted the limitations of panoramic radiographs for accurate spatial localization of impacted canines. Due to their 2D nature and structural superimpositions, differentiating buccal from middle orientations remained challenging. This was consistent with the conclusions of Nagpal et al. [27], who reported that panoramic radiography alone may be insufficient for reliable localization of impacted maxillary canines. In a different manner, the performance of AI-based chatbots was also demonstrated to have inadequate accuracy for clinical practice in identifying labiolingual localization of impacted canines using only panoramic radiographs [28].

Based on a literature review, CBCT remains the gold standard for precise localization of impacted maxillary canines due to its 3D imaging capabilities, allowing better visualization of anatomical landmarks and potential root resorption [29]. Similarly, Araujo et al. [30] demonstrated successful prediction of palatally impacted maxillary canines using supervised machine learning applied to CBCT images, achieving high diagnostic accuracy. However, higher radiation dose, cost, and limited accessibility of CBCT, especially in younger patients, reduced its practicality for routine screening. Therefore, panoramic radiographs have still accepted as a more reliable diagnostic method for predicting the spatial orientation of impacted canines [12]. While this study did not incorporate CBCT data due to the nature of our dataset, the results indicated that panoramic radiographs, when combined with deep learning algorithms, could provide clinically useful insights despite their inherent limitations.

Limitations

This study had several limitations. The use of 2D panoramic radiographs introduced challenges related to anatomical overlap and reduced discriminative power between buccal and middle orientations. The relatively small dataset size limited the models’ generalizability. The imbalance in the number of samples per class may have affected the models’ ability to learn minority classes effectively. Briefly, dataset imbalance and anatomical overlap were identified as major factors affecting model accuracy and class separation.

While YOLO-based architectures achieved moderate success, the testing-phase results showed that further improvements, such as integrating 3D imaging data or combining panoramic and CBCT datasets, would be necessary to enhance classification performance. Therefore, future studies should focus on larger and balanced datasets, advanced deep learning architectures, and integrating 3D data to enhance diagnostic precision and improve clinical applicability.

Conclusion

In this study, deep learning models based on YOLO11x and YOLO12x were evaluated for their ability to classify buccal, middle, and palatal orientations. YOLO11x demonstrated superior overall performance, achieving higher precision, better balance between recall and accuracy, and lower misclassification rates compared to YOLO12x. The superior performance of YOLO11x, especially for palatal orientations could assist clinicians in early identification and treatment planning of impacted canines. However, moderate accuracy in buccal and middle classes suggested that CBCT imaging may still be required for precise localization. Given these findings, panoramic radiographs have certain limitations in determining the spatial orientation of impacted maxillary canines due to their 2D structure and anatomical superimpositions.

Authors’ contributions

All authors contributed to the study’s conception and design. SÇ and ST contributed to data collection and annotating images. The images were annotated by SÇ and ST. MÜÖ contributed to algorithm design. Coding, and determination of performance evaluation metrics. All authors contributed equally to article manuscript writing, literature review, and article design. All authors approved the final version of the manuscript.

Funding

No funding was declared.

Data availability

The datasets supporting this study’s findings are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The study protocol was approved by the Ethics Committee of Pamukkale University (Date/Protocol No: 12.08.2025/15) and conducted in accordance with the Helsinki Declaration. Consent to participate is not available. This is a retrospective study, and the ethics committee does not require informed consent in retrospective studies.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.