Automated classification of maxillary and mandibular dental views on intraoral photographs: a comparative benchmark study and mobile proof-of-concept

Abstract

Artificial intelligence is increasingly explored in dentistry to improve workflow efficiency and support image-based analysis. This study benchmarks deep learning (DL) and machine learning (ML) approaches for classifying pediatric dental views using a publicly available dataset of 9,562 intraoral images from children aged 1–14 years, covering eight maxillary and mandibular view classes. Under 10-fold cross-validation, MobileNetV2 achieved the highest performance among DL models (accuracy 95.18%, F1-score 0.95, AUC 0.997), followed by InceptionV3 (93.76%) and Xception (93.07%). Among ML methods, Logistic Regression achieved 93.33% accuracy with an AUC of 0.996. A symmetry-aware architecture, DentSym, was further proposed, achieving an average accuracy of 98.92% with balanced precision, recall, and F1-score. Model interpretability was examined using Grad-CAM, indicating that predictions were based on relevant dental regions. The highest-performing model was integrated into a prototype iOS application for real-time classification as a proof of concept. However, as cross-validation was performed at the image level due to the absence of patient identifiers, the reported performance should be interpreted as an upper-bound estimate under the current experimental setting. The study provides baseline reference results for this dataset and highlights the potential of explainable, mobile-based AI systems for future dental applications.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12903-026-09089-6.

Introduction

Artificial intelligence (AI) is increasingly shaping healthcare by providing automated and precise diagnostic tools. Dentistry, long dependent on human expertise, is now adopting AI for image classification, anomaly detection, and treatment planning. These advances are supported by curated and labelled datasets that enable models to learn and generalize effectively [1]. Although peripheral in earlier practice, automatic classification of dental images has clear clinical relevance today [2]. It facilitates preliminary screening, assists non-specialists in underserved regions, and ensures standardized inputs for downstream diagnostic models, for example in distinguishing maxillary from mandibular views [3].

The role of machine learning and deep learning in dentistry has been widely examined. These technologies support diagnostic automation, improve visualization, and enhance clinical decision-making [4]. Dental images remain central, as they are critical for health assessment, early abnormality detection, and treatment planning [5]. Traditional practice relies on manual evaluation of radiographs. AI systems, in contrast, can analyze thousands of images within minutes, minimizing human error and saving time, thus providing significant assistance in routine care [6].

Notable contributions include the work of Seino et al., who applied deep learning for view classification and tooth detection in intraoral images. Their dataset contained 1,450 images from 290 cases across five views. Images were resized and augmented before training. A fine-tuned ResNet152 achieved 100% accuracy in view classification, while YOLOv5s trained on 215 cases reached ~ 93% accuracy in tooth detection after post-processing [7]. The same dataset was later employed in related research [8], but few comparative studies exist. Evidence from these works indicates that deep learning is increasingly explored for dental imaging, with applications in dentition stage recognition, diastema detection, and developmental anomaly assessment. Several architectures have been explored, including InceptionResNetV2 [9], Xception [10], InceptionV3 [11], MobileNetV2 [12] and ResNet-101 [11]. Adnan et al., developed a smartphone caries detection app (7,465 images) using YOLOv5s and Transformer, achieving 90.7% precision, 85.6% sensitivity, and visual explanations with Grad-CAM [13]. Xiong et al. proposed ToothNet (YOLOX) for detecting caries and sealants on 1,020 images, achieved AUCs of 0.925 (caries) and 0.902 (sealants), achieving higher performance than less-experienced clinicians [14]. Zhang et al., trained an SSD-based CNN on 3,932 consumer-camera photographs, reported 85.65% AUC and 81.9% sensitivity, showing feasibility for population-level caries screening [15]. Yoon et al., applied cascade R-CNN on 24,578 intraoral images, reached mAP of 0.880 for tooth numbering and 0.769 for caries classification [16]. Thakuria et al., introduced OralSegNet (UNet variant) trained on 538 lesion images, achieved Dice scores of 0.953 (validation) and 0.852 (test) for oral lesion segmentation [17]. Ryu et al., used ResNet and VGG on 3,136 occlusal images, obtained 92.2% accuracy for extraction prediction and Cohen’s kappa of 0.73 for crowding classification [18]. Ghorbani et al. trained YOLOv8-based framework on 3,215 occlusal photographs, achieved F1-score of 97.76%, with errors mainly in rare primary mandibular canines [19]. Nguyen et al. presented SegmentAnyTooth on 5,000 annotated images, Dice coefficients exceeded 0.92 across intraoral views, demonstrating improved segmentation performance compared to standard segmentation networks [20]. Pang et al., Vision Transformer with Mask R-CNN graded tooth wear severity, achieved 93% accuracy, F1 up to 0.97, reducing grading time to seconds [21]. Kartbak et al. predicted cephalometric parameters from 990 intraoral–cephalogram pairs, reported accuracies up to 99% for interincisal angle and Wits appraisal [22]. Liu et al., AI-assisted analysis of 5,951 adolescent 3D dental models, identified rapid arch growth before age 13–14, with geographic differences affecting development [23]. Ragodos et al., CNNs identified ten dental anomalies from cleft patients’ intraoral photos, results comparable to experienced dentists [24]. Takahashi et al., trained CNNs on 1,904 intraoral photographs, detected 11 types of restorations with mAP 0.80, metallic prostheses recognized more accurately [25]. Mouncif et al., proposed RNN-based system classifying 3D tooth models converted into 2D, enabled accurate, fast forensic tooth identification [26]. Arian et al. used synthetic-to-real domain adaptation on 3D intraoral scans with PointNet, improved IoU from 0.35 to > 0.80, avoiding manual labeling [27].

The present study builds on these foundations by applying the dataset to develop AI-based tools that integrate computer science with clinical dentistry. Beyond evaluating deep learning models, it provides a detailed dataset analysis and demonstrates AI’s potential to address practical challenges in dental care. This work responds to the growing demand for AI in pediatric dentistry, where automation of basic tasks such as view classification can reduce diagnostic time, improve accuracy, and extend accessibility to underserved areas. Annotated datasets also hold value for education, supporting both dental training and AI development by reducing reliance on manual labeling.

While view classification alone is not diagnostic, it constitutes an essential preliminary stage for advanced AI-driven systems. Accurate identification helps ensure that downstream models, such as those for anomaly detection, caries assessment, or orthodontic evaluation, are applied appropriately. For instance, a mandibular-specific model cannot be used on maxillary images. By addressing this initial step, the study provides a scalable approach for multi-stage workflows, particularly in teledentistry where images are often captured by non-specialists.

This work provides baseline reference results for the dataset and highlights dental view classification as a preparatory step for more advanced tasks such as segmentation, anomaly detection, and treatment planning. Even simple automated processes may serve as building blocks for more complex AI-based dental systems. The findings suggest the potential of AI in dental imaging while requiring further validation in broader clinical contexts.

The main contributions of this study are as follows:

A large pediatric dental dataset was systematically evaluated using both machine learning and deep learning models, providing baseline performance results for this task.

The proposed DentSym architecture achieved consistent performance, maintaining a balanced trade-off between precision, recall, and F1-score across all experimental settings.

Grad-CAM was employed to support interpretability, indicating that the model focused on relevant dental regions.

A proof-of-concept mobile application was developed to demonstrate the feasibility of real-time, on-device classification.

The study highlights dental view classification as a preparatory step for downstream AI-based dental applications.

Materials and methods

This study addresses the classification of pediatric dental images using the dataset “Varying views of maxillary and mandibular aspects of teeth”, which contains 9,562 images of healthy teeth from children aged 1–14 years. The dataset, publicly available through Data in Brief and Mendeley Data [8], was divided into ten folds for cross-validation. Both DL and classical ML approaches were tested. DL models included Xception, MobileNetV2, InceptionV3, ResNet-101, and InceptionResNetV2. ML models consisted of Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbors (k-NN), and Logistic Regression (LR). For DL, features were learned directly from images, whereas ML models were trained on 1,000 extracted features. Performance was assessed using accuracy and confusion matrices averaged across folds.

Alongside these baselines, a new architecture, DentSym, was introduced and trained under the same 10-fold scheme. DentSym achieved higher accuracy than both DL and ML comparators. Model interpretability was supported through Gradient-weighted Class Activation Mapping (Grad-CAM), which visualized the image regions most influential in decision-making. To illustrate practical use, the DentSym model was deployed into a prototype iOS mobile application capable of classifying new dental images in real time. The overall workflow is shown in Fig. 1.

Fig. 1.

Maxillary and mandibular aspects view of teeth classification flow diagram

Model performance was evaluated using 10-fold cross-validation. Standard performance metrics, including accuracy, precision, recall, F1-score, and the Area Under the ROC Curve (AUC), were calculated. The model with the highest numerical performance under the evaluated conditions was selected for integration into a proof-of-concept mobile application.

Dataset

This dataset, specifically focused on children aged 1 to 14 years, offers unique contributions to pediatric dental research. By providing labelled 9,562 images of primary and permanent dentition, it enables researchers to study developmental growth and track dental health milestones. The dataset’s healthy teeth images can also serve as a baseline for identifying anomalies or developmental delays, enhancing pediatric dentistry’s ability to monitor and assess dental health. Image captured from eight different angles under varying conditions based on view and arch: maxillary (upper) and mandibular (lower) arches photographed from front, right, left, and occlusal (biting surface) angles [8]. Image samples for the dataset shown in Fig. 2. For each class the total number of images presented in the illustration.

Fig. 2.

Sample images from the “Teeth or Dental Image Dataset”

All images were resized to 224 × 224 pixels to match the input requirements of the pretrained models and normalized to pixel values between 0 and 1. No advanced data augmentation techniques such as rotation or flipping were applied at this stage.

Machine learning models

Besides the deep learning approaches, several machine learning algorithms were tested on the same dataset for comparison. The models included SVM, RF, k-NN, and LR. SVM was applied as a margin-based classifier, RF as an ensemble of decision trees, k-NN as a distance-based method, and LR as a statistical model for multiclass prediction. These methods were selected because of their frequent use in medical image studies and their ability to provide baselines against which deep learning models can be assessed. The block diagrams of the applied algorithms are shown in Figure S1.

For the classical machine learning models, specific parameters were optimized to ensure fair evaluation. The SVM was run as a C-SVM with a cost value of 1.0, epsilon of 0.10, complexity bound of 0.50, and a radial basis function (RBF) kernel with γ set automatically; an iteration limit of 100 and numerical tolerance of 0.0010 were applied. The RF classifier was constructed with 100 trees, limiting the depth of individual trees to three, avoiding splits with fewer than five samples, and balancing the class distribution. The k-NN algorithm used three neighbors, Euclidean distance as the metric, and uniform weighting. Finally, LR was performed with ridge (L2) regularization, penalty strength C = 1.0, and without class balancing.

Parameters used for each ML model presented in Table 1.

Table 1.

Parameters for machine learning models

Model	Key Parameters
SVM	Type = C-SVM; Cost (C) = 1.0; Regression loss epsilon (ε) = 0.10; Kernel = Radial Basis Function (RBF) with γ = auto; Iteration limit = 100; Numerical tolerance = 0.0010.
RF	Number of trees = 100; Do not split subsets smaller than 5; Replicable training enabled; Class distribution balanced.
k-NN	Number of neighbors = 3; Distance metric = Euclidean; Weight = Uniform.
LR	Regularization = Ridge (L2); Strength C = 1.0 (medium); Class distribution balancing.

Deep learning models

The use of deep learning has had a significant impact on the field of medical image analysis, enabling improved classification performance through the automatic learning of hierarchical features from complex data. In dentistry, deep learning models are increasingly applied to analyze dental images and identify relevant patterns in a systematic manner. In this study, several widely used convolutional neural network (CNN) architectures were evaluated for the classification of dental images, including Xception, MobileNetV2, InceptionV3, ResNet-101, and InceptionResNetV2. These architectures have demonstrated effective feature extraction capabilities in image analysis tasks and provide a diverse set of design characteristics for comparative evaluation.

To avoid making assumptions in their conclusions, each of these deep learning models was trained as well as evaluated using a systematic approach. The dataset was subjected to 10-fold cross validation, meaning that it was divided into 10 parts for training and testing the models cyclically. This method improves the concept and lowers the possibility of overfitting, since every subset of the data can be used as a holdout once to be tested and all the remaining as training. A batch size of 8 was selected to leverage speed and stability of gradient updates and a fixed learning rate of 0.0001, 30 epochs and an optimizer as ‘adam’ was adopted to maintain smooth convergence of the optimizing algorithm.

All classification experiments were conducted on a laptop with the following system configuration: 12th Gen Intel(R) Core(TM) i7-12650 H processor (2.30 GHz), 64.0 GB of RAM, and an 8 GB dedicated graphics card. The system runs on Windows 11 Education Edition, version 24H2 (OS build: 26100.4484), with the Windows Feature Experience Pack 1000.26100.128.0, installed on 25 November 2024. The classification models and related scripts were implemented using Python version 3.9.21 (https://www.python.org/downloads/release/python-3921/).

The models used represent a diverse range of architectures and strengths:

Xception is a depthwise separable convolutional network, which excels in capturing spatial features while reducing the number of parameters, making it efficient for high-resolution image analysis (Figure S2) [28, 29].

MobileNetV2 is designed for efficiency, using depthwise separable convolutions and inverted residuals, making it suitable for resource-constrained environments while maintaining competitive accuracy (Figure S3) [30, 31].

InceptionV3 builds on the success of earlier Inception architectures, using factorized convolutions and auxiliary classifiers to improve performance and mitigate overfitting in deep networks (Figure S4) [32].

ResNet-101 incorporates deep residual learning with 101 layers, allowing the model to learn complex features without degradation in performance, making it a gold standard for deep image classification tasks (Figure S5) [33, 34].

InceptionResNetV2 combines the Inception module’s multi-scale feature extraction capabilities with the power of residual connections, enabling faster convergence and better handling of vanishing gradients in deep networks (Figure S6) [35].

Below is the layer count and parameter details for each model included in your code (Table 2), summarized for clarity:

Table 2.

Used models and their parameters

Model	Total Layers	Total Parameters
Xception	133	~ 22.9 million
MobileNetV2	154	~ 3.5 million
InceptionV3	315	~ 23.9 million
ResNet-101	347	~ 44.7 million
InceptionResNetV2	782	~ 55.9 million

These models were selected because of their widespread use and reported effectiveness in medical image analysis tasks, as well as their architectural diversity. Not only does the research assess worth of classification done by these architectures but also explains their usefulness in the dataset which is an important aspect for industrial advancement of DL towards dental research and applications.

Dental symmetry (DentSym) model

The DentSym model is a specialized deep learning architecture designed to leverage the anatomical left-right symmetry inherent in dental images to improve classification accuracy. This is achieved by processing an input image and its horizontally flipped counterpart simultaneously through a single, shared-weight feature extractor. The architecture is composed of three primary parts: a shared convolutional backbone, a feature fusion module, and a dual-headed classification system with a symmetric consistency loss.

Backbone feature extractor

The backbone of the proposed model was implemented as a custom convolutional neural network (CNN), designed from scratch without pre-trained weights. Its architecture was inspired by efficient designs such as EfficientNet and is composed of Mobile Inverted Bottleneck Convolution (MBConv) blocks. Each block expands the feature dimension through a 1 × 1 convolution, applies a depthwise 3 × 3 convolution to reduce computation, and projects the features back using a final 1 × 1 convolution. Squeeze-and-Excitation (SE) modules were integrated to improve channel-wise attention, and a swish activation function was adopted throughout. Residual connections were added where input and output dimensions matched, allowing stable training of deeper layers. The final feature representation was obtained through global average pooling of the last convolutional layer.

Feature fusion and dual-headed classification

The distinguishing feature of the DentSym model is its handling of input symmetry. An input image is passed through a shared-weight backbone to produce a feature vector, denoted as f1. A horizontally flipped version of the same image is created and passed through the same backbone to produce a second feature vector,  f2. These two feature vectors are then fused to create a comprehensive representation. This is accomplished by concatenating  f1 and  f2 along with their sum, difference, and element-wise product: []. This concatenated vector, which explicitly encodes both symmetric and asymmetric features, is then passed to the primary classification head. The primary head consists of a fully connected layer with a swish activation, a dropout layer for regularization, and a final dense layer that produces the classification logits.

Simultaneously, a second, auxiliary classification head takes only the flipped image’s feature vector,  f2, as input. This head follows a similar structure to the primary head but is used specifically to enforce a consistency loss during training.

Training and loss functions

The model is trained to minimize a combined loss function that includes both a standard classification loss and a symmetric consistency loss. The main classification loss is a Sparse Categorical Cross-Entropy calculated from the logits of the primary head. The symmetric consistency loss is computed as a Kullback–Leibler (KL) divergence between the logits of the primary head and those of the auxiliary head after permutation. The permutation (FLIP_PERM) is a predefined mapping that swaps left and right classes (e.g., “Upper Left” becomes “Upper Right”). This formulation encourages consistent predictions under horizontal flipping, allowing the model to account for orientation-related variations. The total loss is defined as a weighted sum of the two components, where the consistency term acts as a regularization mechanism. This dual-loss formulation guides the model to learn symmetry-aware feature representations for dental image classification. The model architecture and description are presented in Figs. 3 and 4.

Fig. 3.

Dental symmetry (DentSym) model architecture diagram

Fig. 4.

DentSym model description

Cross validation

The dataset was evaluated using 10-fold cross-validation. Although patient-level identifiers were not available, care was taken to minimize data leakage by ensuring that visually similar or duplicate images were not distributed across different folds. All preprocessing and normalization steps were applied after splitting to avoid information leakage. Nevertheless, strict subject-level independence could not be fully guaranteed. In this approach, the dataset is divided into 10 equal subsets, and the model is trained and tested 10 times. In each iteration, one of the 10 subsets is considered the testing set while the other 9 are classified as the training set. This process ensures that each data point is used both for training and for testing. By calculating the mean performance across the 10 folds, this approach provides a more stable estimate of model performance. Figure 5 shows the illustration for the cross validation [36, 37].

Fig. 5.

Cross validation based on 10 iteration experiments

Performance metrics and confusion matrix

Several measures such as accuracy, precision, recall, F1 score and area under the curve were used in this study to carry out complete evaluation of the different machine learning and deep learning models’ performances. These measures were notably selected to facilitate the overall evaluation of the classification models by addressing different performance dimensions. In addition, confusion matrices [38] were also created in order to analyse the classification outcomes of each class in more detail (Fig. 6). The formula for each metric is presented in Table 3.

Fig. 6.

Multi class confusion matrix

Table 3.

Metrics and formulas evaluation

Performance Metrics	Formulas (Binary Class)	Formulas (multi-class)
Accuracy
Precision
Recall
F1-Score

In a multi-class classification problem with n classes, the confusion matrix provides a detailed breakdown of the model’s performance for each class. When focusing on a specific class C (where 0 ≤ k ≤ n), four possible outcomes can be identified. True Positives (TP), shown in dark cyan-blue, represent the instances correctly classified as belonging to class C. True Negatives (TN), highlighted in bright blue, are the instances correctly identified as not belonging to class k. False Positives (FP), depicted in very pale blue, occur when instances are incorrectly classified as belonging to class C, even though they belong to a different class. Lastly, False Negatives (FN), represented in very light blue, occur when instances that actually belong to class k are misclassified as belonging to another class [29].

Accuracy is the ratio of correctly classified instances to the total number of instances, hence gives direct communication of the performance of the model. Accuracy is vital but in cases of some class dominating the whole dataset, it may be misleading (Table 3).

Precision assesses the fraction of true positives out of all the predicted positives and shows the model’s capability of generating false positive errors. High precision is important in cases where false warnings could lead to interventions or actions that are not necessary (Table 3).

Recall (or sensitivity) can be expressed as the ratio of the number of true positives in the results produced by the model to the total number of actual positives — i.e. it assesses the model’s ability to uncover all the relevant cases. High recall helps ensure that important instances, such as in this case specific dental views, will not be missed (Table 3).

F1 Score is the mean value of precision and recall calculated as the harmonic mean thereby facilitating balance in their trade off. This metric is useful in cases where both precision and recall are needed as it takes into account both false positives and false negatives (Table 3).

The Area Under Curve (AUC) assesses whether the model is able to differentiate between different classes based on the threshold applied at different points thus indicating the discriminatory ability of the model. AUC is useful in appreciating the strength of the model, especially in multi-class situations [39–41].

Confusion matrices were generated to provide class-level insight into model predictions. They revealed that most errors occurred between anatomically adjacent views, such as lower left and lower right arches, while occlusal views were classified with high consistency. These results highlight both the strengths and remaining challenges of the tested models in handling visually similar dental regions.

Receiver Operating Characteristic (ROC) curve

Receiver Operating Characteristic (ROC) curves were generated to evaluate the discriminative ability of all models. The proposed DentSym model achieved a mean AUC of 0.99 across 10 folds, with low variation (± 1 SD), indicating consistent performance under the evaluated experimental conditions. The curves were concentrated in the upper-left region of the plot, corresponding to high sensitivity and low false positive rates. These results suggest that the model is effective in distinguishing between the defined classes within the dataset.

In the multi-class setting, AUC was computed using a one-vs-rest (OvR) approach, where each class was evaluated against all others. The resulting class-wise AUC scores were aggregated using macro-averaging to ensure equal contribution of all classes. For 10-fold cross-validation, AUC was calculated independently for each fold, and the final performance was reported as the mean AUC across all folds.

Gradient-weighted Class Activation Mapping (Grad- CAM)

Grad-CAM was employed to provide visual explanations of the DentSym model’s predictions. The gradients of the predicted class with respect to the final convolutional feature maps were computed, and heatmaps were generated to highlight the regions contributing most to classification. These maps were upsampled and overlaid on the original dental images for qualitative assessment. Grad-CAM was applied across the cross-validation folds to ensure consistency, and representative examples are presented in the Results section (Fig. 15).

Fig. 15.

Grad-CAM samples

Results

In this section, the classification performance of pediatric dental images is reported using machine learning (ML), deep learning (DL), and the proposed DentSym model. The dataset (n = 9,562) was evaluated under a 10-fold cross-validation scheme to obtain stable performance estimates. For each model, standard performance metrics (accuracy, precision, recall, F1-score, and AUC) were computed and averaged across folds.

Machine learning models

The performance of ML classifiers is summarized in Table 4. Logistic Regression (LR) achieved the highest accuracy (93.33%), balanced recall/precision (0.933), F1-score (0.933), and the highest AUC (0.996). k-Nearest Neighbors (kNN) followed with 90.67% accuracy and an F1-score of 0.907. Random Forest (RF) performed moderately (86.97%), while Support Vector Machine (SVM) obtained the lowest accuracy (81.29%) but still maintained an AUC above 0.97.

Table 4.

Results of machine learning models

ML Models	Accuracy (%)	Recall	Precision	F1 Score	AUC
SVM	81.29	0.8129	0.8172	0.8135	0.9783
RF	86.97	0.8697	0.8712	0.8699	0.9859
kNN	90.67	0.9067	0.9075	0.9069	0.9803
LR	93.33	0.9333	0.9334	0.9333	0.9955

The confusion matrices (Fig. 7a–d) highlight systematic misclassifications. SVM showed frequent errors between anatomically symmetric regions, particularly Lower Left vs. Lower Right and Upper Left vs. Upper Right. RF reduced some of these confusions but remained less consistent than kNN. The kNN model achieved stable recognition of occlusal and frontal views, though lateral errors persisted. LR delivered the most balanced results across all classes, with fewer symmetric confusions. ROC curves (Fig. 8a-d) confirmed high discriminative ability across all ML models (AUC > 0.97), reflecting consistent but varying overall accuracy.

Fig. 7.

Confusion matrix based on machine learning models

Fig. 8.

ROC curve for machine learning models

Deep learning models

Results of DL architectures are reported in Table 5. MobileNetV2 achieved the highest performance (accuracy 95.18%, F1-score 0.952, AUC 0.998), compared to InceptionV3 (93.76%) and Xception (93.07%). InceptionResNetV2 and ResNet-101 performed less effectively, with accuracies of 90.61% and 86.69%, respectively.

Table 5.

Classification results based on deep learning models

DL Models	Accuracy (%)	Recall	Precision	F1 Score	AUC
Xception	93.07	0.9307	0.9319	0.9308	0.9955
MobileNetV2	95.18	0.9518	0.9523	0.9519	0.9975
InceptionV3	93.76	0.9376	0.9398	0.9371	0.9969
ResNet-101	86.69	0.8669	0.8775	0.8666	0.9897
InceptionResNetV2	90.61	0.9061	0.9121	0.9059	0.9942

The confusion matrices and ROC curves are shown in Figs. 9, 10, 11, 12, 13 and 14. Across all DL models, occlusal views were classified with very high accuracy, while most misclassifications occurred between lateral pairs (Lower Left vs. Lower Right, Upper Left vs. Upper Right). MobileNetV2 consistently minimized these errors compared to deeper networks, demonstrating its effectiveness under the evaluated dataset conditions. Its lightweight design and high AUC (0.9975) indicate its feasibility for prototype mobile implementation.

Fig. 9.

Confusion Matrix and ROC Curve for Xception

Fig. 10.

Confusion Matrix and ROC Curve for MobileNetV2

Fig. 11.

Confusion Matrix and ROC Curve for InceptionV3

Fig. 12.

Confusion Matrix and ROC Curve for ResNet-101

Fig. 13.

Confusion Matrix and ROC Curve for InceptionResNetV2

Fig. 14.

DentSym Confusion Matrix

DentSym model

The proposed DentSym model achieved the highest overall results (Table 6). Under 10-fold cross-validation, it obtained a mean accuracy of 98.92% (± 0.99), macro-F1 score of 0.989, precision of 0.989, recall of 0.989, and AUC of 0.999. Fold-wise performance was consistently above 96%, with most folds exceeding 99%, confirming the stability of the architecture.

Table 6.

DentSym based model’s performance metrics results

Model	Accuracy (%)	Recall	Precision	F1 Score	AUC
DentSym	98.92	0.9896	0.9891	0.9889	0.999

DentSym achieved higher numerical performance compared to other models under the same experimental conditions; however, these comparisons are descriptive, as no statistical significance testing was conducted.

The confusion matrix (Fig. 14) shows very high classification performance across all eight dental views. Misclassifications were rare and primarily involved neighboring regions such as Lower Right vs. Lower Front and Upper Front vs. Upper Right. Grad-CAM heatmaps further confirmed that the model’s attention focused on clinically relevant structures, particularly occlusal and labial surfaces, rather than background regions (Fig. 15). This supports the interpretability of the model’s predictions. Figure 15a presents the Grad-CAM visualization of the lower teeth, whereas Fig. 15b illustrates the upper teeth.

Table 7 presents a comparative summary of all evaluated models, consolidating their performance metrics in a single view. Under the evaluated experimental conditions, DentSym achieved higher numerical performance across the reported metrics compared to the other evaluated models. However, these comparisons are descriptive, as no statistical significance testing was performed.

Table 7.

Comparison presentation of models outcomes

Model	Accuracy (%)	Recall	Precision	F1 Score	AUC
DentSym	98.92	0.9896	0.9891	0.9889	0.9990
MobileNetV2	95.18	0.9518	0.9523	0.9519	0.9975
InceptionV3	93.76	0.9376	0.9398	0.9371	0.9969
LR	93.33	0.9333	0.9334	0.9333	0.9955
Xception	93.07	0.9307	0.9319	0.9308	0.9955
kNN	90.67	0.9067	0.9075	0.9069	0.9803
InceptionResNetV2	90.61	0.9061	0.9121	0.9059	0.9942
RF	86.97	0.8697	0.8712	0.8699	0.9859
ResNet-101	86.69	0.8669	0.8775	0.8666	0.9897
SVM	81.29	0.8129	0.8172	0.8135	0.9783

Mobile application prototype

To demonstrate the real-world applicability of the trained models, particularly the proposed DentSym architecture, a prototype mobile application was developed. Following systematic evaluation of multiple DL and ML models, DentSym was selected for deployment due to its higher observed accuracy, balanced performance across classes, and computational efficiency.

The application was built for iOS devices using Swift and Apple’s CoreML framework, allowing seamless integration of the trained model for on-device inference. This design ensures rapid classification without internet connectivity, addressing concerns of latency, privacy, and accessibility. The system supports two modes of image input: (i) selection from the photo library and (ii) real-time image capture through the device camera. Images are preprocessed and passed through the DentSym model, with the resulting classification displayed to the user. If the image cannot be processed, an error prompt is provided.

The current version (v1.0.0) is a research prototype intended solely for demonstration and evaluation. It classifies images into one of eight dental views—Lower Front, Lower Left, Lower Occlusal, Lower Right, Upper Front, Upper Left, Upper Occlusal, and Upper Right—consistent with the experimental dataset. A disclaimer accompanies outputs to clarify that predictions are supportive rather than diagnostic. Screenshots of the application interface are presented in Fig. 16. The individual interfaces shown in Fig. 16 correspond to the workflow stages as follows: the application splash screen is presented in Fig. 16a, followed by the home screen in Fig. 16b. The image upload and gallery selection interfaces are shown in Fig. 16c and d, respectively. The camera activation and image capture screens are presented in Fig. 16e and f. The image confirmation and preprocessing preview interfaces are illustrated in Fig. 16g and h. Finally, the classification output and result visualization screens are presented in Fig. 16i, j, k, and l.

Fig. 16.

Smart dental view classifier: application interface

Although preliminary, the application illustrates a practical pathway from research to translation. Such a tool has potential utility in teledentistry, where non-specialists may capture dental images in underserved areas, and in education, where automated classification can support training and annotation. Future development will focus on extending compatibility to Android devices, incorporating additional models, and enhancing usability features such as multilingual support, storage of classification history, and real-time feedback.

Discussion

This study demonstrates that the classification of pediatric dental views can be effectively approached using artificial intelligence methods, with both classical ML and DL models achieving high performance under the evaluated experimental conditions. The systematic benchmarking provides baseline reference values for this publicly available dataset and highlights specific challenges, particularly those arising from anatomical similarity between lateral views.

Among ML approaches, LR achieved the most consistent results, with accuracy exceeding 93%. This suggests that, when informative features are available, relatively simple models can capture relevant discriminative patterns in dental images. However, the observed confusion between symmetric classes, such as lower left and lower right, indicates the limitations of linear decision boundaries in handling subtle spatial differences. Instance-based methods such as kNN partially reduced these errors but remained less consistent compared to DL-based approaches.

Within the DL models, MobileNetV2 achieved the highest numerical performance among the evaluated architectures. This observation suggests that lightweight and parameter-efficient models may generalize effectively on datasets of this size. Nevertheless, persistent misclassifications between adjacent views indicate that conventional architectures relying primarily on global features may not fully capture fine-grained anatomical distinctions. It should be noted that these comparisons are descriptive, as no statistical significance testing was performed.

The proposed DentSym model was designed to address these challenges by incorporating symmetry-aware feature learning. Under the same evaluation protocol, DentSym achieved higher numerical performance compared to the baseline models, with balanced precision, recall, and F1-scores across classes. The reduction in confusion between symmetric regions suggests that explicitly modeling left–right relationships can be beneficial for this task. Grad-CAM visualizations further indicated that the model focused on relevant dental structures, particularly occlusal and labial regions, supporting the interpretability of the learned representations. However, these findings should be interpreted within the constraints of the current dataset and evaluation setup.

The integration of DentSym into a mobile application prototype illustrates a potential pathway for translating such models into practical tools. The use of on-device inference through CoreML demonstrates the feasibility of real-time image classification without reliance on network connectivity. However, this implementation should be considered a proof-of-concept demonstration, and not as evidence of readiness for clinical deployment. Further validation, usability testing, and regulatory considerations would be required before real-world application.

Several important limitations must be considered when interpreting the results. First, cross-validation was performed at the image level due to the absence of patient identifiers in the dataset. As a result, it cannot be guaranteed that images from the same individual were not included in both training and testing folds. This may introduce a degree of information leakage and could lead to optimistic performance estimates. Therefore, the reported results should be interpreted as indicative of model capability under the current experimental conditions rather than as a fully independent patient-level evaluation.

Second, the dataset consists exclusively of healthy pediatric dental images acquired under specific conditions. The models were not evaluated on pathological cases, images from different institutions, or data obtained using varied imaging devices. This limits the broader applicability of the findings, and model performance may vary in more diverse or clinically realistic scenarios. External validation on independent datasets is therefore necessary to further evaluate model performance under diverse conditions.

Third, no statistical significance testing was conducted to compare model performance. As such, observed differences between models should be interpreted as numerical and exploratory rather than statistically confirmed.

Accordingly, the reported performance values may represent an upper-bound estimate under the current experimental setup rather than performance achievable under fully independent clinical conditions.

Previous studies in dental image analysis have reported favorable performance using DL approaches; however, differences in datasets, class definitions, and experimental protocols limit direct comparison. In this context, the present study contributes by providing reproducible baseline results on a publicly available dataset. The DentSym model further extends existing approaches by introducing symmetry-aware learning tailored to the anatomical characteristics of dental structures.

This work suggests that automated dental view classification is a feasible and potentially useful preparatory step for downstream AI-based dental analysis. Accurate view identification may support the correct application of subsequent models, such as those for anomaly detection or orthodontic assessment. While the results are promising within the current experimental framework, further validation under clinically diverse conditions is required before broader application can be considered.

From a clinical perspective, automated view classification may contribute to the standardization of intraoral image organization, particularly in teledentistry settings where images are often captured by non-specialists. By ensuring consistent categorization prior to further analysis, such systems may help reduce downstream processing errors and improve workflow efficiency. However, these potential benefits remain to be validated in real-world clinical environments.

Conclusion

Accurate classification of dental image views represents an important preparatory step for the development of automated workflows in dental image analysis. Correct identification of maxillary and mandibular views helps ensure that subsequent tasks, such as anomaly detection, caries assessment, or orthodontic evaluation, are applied to appropriate inputs. In this study, both classical machine learning and deep learning models were systematically evaluated on a publicly available pediatric dataset of 9,562 intraoral images. The findings indicate that machine learning approaches provide useful baseline performance, with Logistic Regression achieving an accuracy of 93.3%, while deep learning models demonstrated higher numerical performance under the same experimental conditions. Among the evaluated architectures, MobileNetV2 achieved the highest results among standard deep learning models. These findings are limited to the present dataset and experimental setup, and should not be interpreted as evidence of clinical readiness. The proposed DentSym architecture further achieved improved numerical performance, with balanced precision, recall, and F1-scores across classes. Grad-CAM visualizations suggested that the model’s predictions were based on relevant dental regions. In addition, the integration of DentSym into a prototype iOS application demonstrated the feasibility of real-time, on-device classification as a proof of concept. However, the reported results should be interpreted with caution. Cross-validation was performed at the image level due to the absence of patient identifiers, and therefore a fully independent patient-level evaluation could not be ensured. This may introduce a degree of information leakage and lead to optimistic performance estimates.

Furthermore, the dataset consists only of healthy pediatric cases, and external validation on diverse populations, imaging conditions, and pathological cases was not conducted. Future work should focus on patient-level evaluation, external validation across multiple datasets, inclusion of clinically diverse cases, and the application of statistical significance testing. Exploration of alternative architectures, such as attention-based or transformer models, may also further improve performance. This study provides baseline reference results for dental view classification on a public dataset and highlights the potential of explainable, mobile-based AI systems, while emphasizing the need for further validation before practical deployment can be considered.

Abbreviations

AI

Artificial Intelligence

AUC

Area Under the Curve

CNN

Convolutional Neural Network

DL

Deep Learning

iOS

iPhone Operating System

ML

Machine Learning

OvR

One-vs-rest

ROC

Receiver Operating Characteristic

YOLO

You Only Look Once

Authors' contributions

Conceptualization, E.T.Y., M.K., M.A., and Y.G.; methodology, E.T.Y., M.K., M.A., and Y.G.; software, E.T.Y.; validation, E.T.Y., M.K., M.A., and Y.G.; formal analysis, E.T.Y., M.K., M.A., and Y.G.; resources, E.T.Y., M.K., M.A., and Y.G.; writing—original draft preparation, E.T.Y., M.K.; writing—review and editing, E.T.Y., M.K., M.A., and Y.G.; visualization, E.T.Y., M.K., M.A., and Y.G.; supervision, M.K.; project administration, E.T.Y. and M.K.; funding acquisition Y.G.;. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the Deanship of Scientific Research, the Vice Presidency for Graduate Studies and Scientific Research, King Faisal University, Saudi Arabia, under Grant No. KFU261547.

Data availability

The “Teeth or Dental image dataset” is available in data in brief, and the dataset can be downloaded directly from the Mendeley Data through this link: [https://data.mendeley.com/datasets/6zsnhrds9t/1](https://data.mendeley.com/datasets/6zsnhrds9t/1).

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Conflict of interest

The authors declare no conflicts of interest.