A Comprehensive Deep Learning Framework for Dental Disease Classification

Dental diseases represent a significant global health challenge, affecting billions of people and often leading to serious complications if not identified and treated promptly. Early diagnosis is crucial to prevent the progression of these conditions, yet access to timely dental care remains limited, particularly in resource-constrained areas. The shortage of dental professionals exacerbates this issue, creating a critical need for tools that can assist in early screening and diagnosis to improve health outcomes.

This study leverages technological advancements in AI and deep learning to develop an automated system for detecting common dental diseases. By employing a convolutional neural network (CNN), Residual Networks (ResNet), and Vision Transformer (ViT), the system can analyze dental images and classify them into five prevalent conditions with high accuracy and reliability. The proposed system incorporates data augmentation to enhance generalization, confidence thresholding to identify uncertain cases, and a user-friendly web interface for seamless integration into clinical workflows.

To further enhance model robustness, we implemented 5-fold cross-validation and conducted hyperparameter tuning using grid search. The optimal hyperparameters identified include a learning rate of 0.0001, batch size of 32, and Adam optimizer, which provided the best trade-off between stability and training speed. The mean accuracy obtained across folds was 87.6%, demonstrating strong reliability. Additionally, we integrated Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations to improve model interpretability, enabling practitioners to understand how the model makes predictions.

To ensure generalizability, we introduced additional evaluation metrics such as specificity and sensitivity, offering deeper insights into model performance across different dental conditions. Moreover, we incorporated an ensemble learning approach, combining predictions from CNN, ResNet, and ViT to further enhance classification accuracy. This fusion approach resulted in improved overall performance, particularly in distinguishing closely related dental abnormalities.

Another enhancement involves deploying the system in a cloud-based framework to facilitate real-time inference and seamless integration with telemedicine platforms. The cloud deployment ensures scalability, enabling widespread accessibility for dental professionals and researchers. Furthermore, the model is continuously updated with newly acquired clinical data to improve predictive accuracy and adaptability to evolving diagnostic standards.

The study focuses on utilizing deep learning to diagnose periodontitis and dental caries in dental X-ray images [1]. It highlights the transformative role of CNNs in detecting and classifying dental diseases, particularly through techniques like image enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE) and bilateral filtering. While traditional approaches often focus on a single condition, this work emphasizes the benefits of simultaneous multi-condition recognition to streamline diagnostic workflows and improve patient outcomes. Research by Kim et al. [2] addresses the challenges associated with dental implant system (DIS) classification. Traditional imaging techniques, such as two-dimensional radiographs, often struggle with the subtle visual differences in implant systems. By employing deep learning models trained on a multicenter dataset annotated by experts, this study demonstrates how robust datasets and advanced DL strategies can overcome these limitations, providing high classification accuracy and advancing the diagnostic capabilities for implant-related conditions.

The review further underscores the utility of CNNs in dental diagnostics, particularly in detecting anomalies, dental caries, and periodontal diseases [3]. It highlights the predominance of panoramic radiographs as the primary imaging modality and identifies the role of DL in tasks such as classification, object detection, and segmentation. The study emphasizes the need for high-quality, diverse datasets to enhance model reliability and draws attention to the existing gaps in research on dental anomalies due to their rarity.

Another systematic review explores the use of AI in detecting dental caries using oral photographs [4]. The authors examine various approaches, including both traditional and deep learning-based methods, and highlight the accessibility of smartphones as a cost-effective alternative for teledentistry. Despite the variability in methodologies, the findings indicate strong potential for AI in early caries detection, while advocating for further research to expand the applicability of smartphone-acquired images in clinical and public health settings.

Finally, Hussain et al. [5] review the broader impact of AI on dental diagnostics, emphasizing its ability to support comprehensive health assessments, including indications of systemic conditions like osteoporosis and sleep apnea visible on panoramic X-rays [6-8]. The paper highlights AI's role in streamlining clinical decision-making by generating prioritized differential diagnoses, thereby improving efficiency and patient care outcomes. Challenges such as dataset availability, validation across diverse populations, and integration into workflows are noted as areas needing attention to fully realize AI's potential in dentistry [9, 10].

Collectively, these studies demonstrate the transformative impact of deep learning and AI in dental diagnostics, providing a solid foundation for developing automated systems that enhance diagnostic precision, streamline workflows, and address accessibility gaps in underserved areas [11, 12]. The existing body of work aligns closely with the goals of our project, emphasizing the use of CNNs for reliable and efficient dental disease detection while identifying areas for further innovation and research [13, 14].

3.2 Convolutional layers

Each convolutional layer applied a filter bank to extract features [24]. Convolution operations followed the equation:

O_i,j,k=∑_m=0^h−1 ∑_n=0^w−1 I_i+m,j+n⋅K_m,n,k+b_k

where, O_i,j,k is the output feature map, I_i+m,j+n is the input pixel, K_m,n,k is the convolutional kernel, and b_k is the bias term. The filters expanded from 64 in the initial layer to 1,024 in the final layer to capture increasingly abstract features:

• First Layer (64 filters): Detected basic features like edges and textures.
• Second Layer (128 filters): Combined lower-level features to identify early-stage dental anomalies.
• Third Layer (256 filters): Captured intermediate patterns like cavity sizes and gum-line structures.
• Fourth Layer (512 filters): Identified complex relationships between teeth and disease patterns.
• Fifth Layer (1,024 filters): Integrated global features for final classification.

3.3 Pooling and dropout

Max-pooling layers (2×22 \times 22×2) reduced the spatial dimensions by a factor of 2, facilitating feature abstraction while preventing overfitting through dropout regularization. Dropout layers followed the equation:

y=Dropout(x)={x^{1−p,if active}_{0,if dropped}}

where, p is the dropout probability, set to 0.25 in convolutional layers and 0.5 in dense layers.

3.4 Fully connected layers

The dense layers transformed the high-dimensional feature maps into class probabilities. The final dense layer employed the Softmax activation function:

Softmax(zi)=e^zi/∑je^zj

This function ensured normalized class probabilities for multi-class classification.

To enhance the robustness and effectiveness of our dental disease classification model, we incorporated two additional state-of-the-art deep learning architectures: ResNet and Vision Transformer (ViT). ResNet (Residual Networks) is designed to mitigate the vanishing gradient problem, enabling deep feature extraction with residual connections. ViT, on the other hand, utilizes self-attention mechanisms to capture long-range dependencies, making it highly efficient in analyzing complex patterns within images. Additionally, EfficientNet was integrated into the comparison due to its ability to optimize both accuracy and computational efficiency through compound scaling, which balances depth, width, and resolution [25].

Comparative Analysis of Model Architectures To ensure a fair comparison, all models were trained on the same dataset using identical preprocessing techniques, training parameters, and evaluation metrics.

240944fa-4520-410b-bef9-985834338407.png

Figure 2. Model summary

Both the 5-layer CNN and ViT achieved the highest accuracy of 87.6%, followed closely by ResNet-50 and EfficientNet. ViT provided superior feature extraction capabilities due to its attention-based mechanism, whereas CNNs, particularly ResNet, effectively captured hierarchical spatial patterns. EfficientNet demonstrated strong performance with moderate computational cost, making it a suitable choice for scenarios requiring efficiency without significant trade-offs in accuracy [26, 27].

The model architecture effectively balances complexity and efficiency. The five-layer CNN allows for hierarchical learning, starting from basic features like edges to advanced dental patterns. Batch Normalization ensures stable gradient flow, and dropout minimizes overfitting. The integration of attention mechanisms and inception-style blocks further enhances the model's ability to extract diverse and meaningful features. The final classification head consolidates all the learned features into a robust prediction, achieving strong accuracy and generalization capabilities (Figure 2).

3.5 Cross-validation and model robustness

To validate the generalization capability of our models, we implemented 5-fold cross-validation. Cross-validation divides the dataset into multiple subsets (folds), where the model is iteratively trained on different combinations of data, ensuring stability and reducing overfitting.

Cross-Validation Results:

Table 3. Cross-validation results

These results indicate that the model performs consistently across different subsets of data, demonstrating strong robustness and reliability (Table 3).

3.6 Hyperparameter tuning

Hyperparameter tuning was performed using grid search to optimize model performance. Below is a summary of the best-selected hyperparameters:

Table 4. Hyperparameter tuning values

These values resulted in the best trade-off between training speed, stability, and generalization (Table 4).

The performance of the proposed 5-layer CNN was thoroughly evaluated against multiple deep learning architectures, including 3-layer CNN, EfficientNet, ResNet-50, and Vision Transformer (ViT). The 5-layer CNN demonstrated its effectiveness in dental disease classification with a validation accuracy of 87.61%, significantly improving from an initial training accuracy of 68.15% to 86.69% across 25 epochs. The validation loss also declined from 0.7729 to 0.5347, indicating stable convergence.

The learning rate dynamically adjusted during training, ultimately stabilizing at 1.25×10⁻⁴, which facilitated smooth optimization and prevented overfitting. The model was evaluated using key performance metrics, including accuracy, precision, recall, and F1-score, across different disease categories.

5.1 Performance metrics calculation

The performance of the models was assessed using standard classification metrics:

Accuracy=TP+TN / TP+TN+FP+FN

where:

• TP (True Positive) - Correctly identified cases
• TN (True Negative) - Correctly rejected cases
• FP (False Positive) - Incorrectly classified healthy cases as diseased
• FN (False Negative) - Missed diseased cases

Each model was trained using identical hyperparameters, and their results were systematically compared across multiple factors.

5.2 Comparative performance of different models

A comparative study was conducted on five different architectures to evaluate their strengths and weaknesses (Table 5).

Table 5. Comparative performance of different models

Model	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Inference Time (ms)
5-Layer CNN	87.6	86.2	85.4	85.8	120
EfficientNet	85.9	85.0	84.3	84.6	180
ResNet-50	86.2	85.5	84.9	85.2	200
ViT	87.6	87.0	86.2	86.6	250
3-Layer CNN	82.3	81.1	80.5	80.8	100

5.3 Key insights from model comparison

5-Layer CNN: Balanced accuracy and efficiency, making it ideal for clinical use with real-time inference needs.

• EfficientNet: Performed well even with limited data but required higher computational resources.
• ResNet-50: Achieved high accuracy due to residual connections, preventing vanishing gradient issues in deeper architectures.
• ViT (Vision Transformer): Performed exceptionally well in capturing complex patterns but had the highest inference time, making it less suitable for real-time applications.
• 3-Layer CNN: Faster training but struggled with complex dental conditions, leading to lower accuracy.

The 5-layer CNN & VIT model offered a balance between computational efficiency and diagnostic accuracy, making it suitable for real-world dental applications.

5.4 Grad-CAM for model interpretability

Convolutional Neural Networks (CNNs) are widely used in image classification tasks due to their high accuracy and feature extraction capabilities. However, their black-box nature makes it challenging to understand how they arrive at their predictions. Gradient-weighted Class Activation Mapping (Grad-CAM) is a powerful explainability technique that provides visual insights into which regions of an image contribute most to a model's decision [30].

The flowchart (Figure 3) in the image illustrates the process of generating Class Activation Maps (CAMs) and how Grad-CAM enhances interpretability. Initially, an input image passes through a CNN, where multiple layers extract hierarchical features. The last convolutional layer plays a key role in spatial feature representation. By applying Class Activation Map (CAM) techniques, Grad-CAM produces heatmaps that highlight the most important regions influencing the model’s prediction. These heatmaps provide class-discriminative localization, allowing researchers and practitioners to assess whether the model focuses on the relevant image features [31, 32].

To further enhance interpretability, we incorporated Grad-CAM visualization into our dental disease classification model. Grad-CAM heatmaps were generated for various dental diseases, such as Caries, Gingivitis, and Hypodontia, demonstrating where the model concentrated its attention while making predictions. These visual explanations aid dental professionals in understanding the model’s reasoning, thereby increasing trust in AI-driven diagnostics.

In cases where even more refined explanations are needed, techniques like Guided Backpropagation can be used alongside Grad-CAM to produce high-resolution and detailed visualizations. This approach proves valuable in fields such as medical image analysis, autonomous driving, and industrial defect detection, ensuring CNN-based models make trustworthy decisions.

By integrating Grad-CAM with other visualization techniques, we bridge the gap between deep learning models and human interpretability, making AI-based solutions more transparent, explainable, and reliable for real-world applications.

• The red-highlighted regions in the Grad-CAM heatmaps indicate areas where the model identified dental anomalies.
• For caries, the model effectively captured decay regions around the tooth enamel.
• In gingivitis cases, the model’s attention focused on gum inflammation areas.
• Hypodontia cases showed model attention on missing teeth regions.
• Mouth ulcers and tooth discoloration cases were clearly detected in the soft tissue regions of the mouth.

These heatmaps (Table 6) provide clinicians with visual justifications for model predictions, increasing the transparency and trustworthiness of AI-driven diagnoses.

1151c726-b0b6-4d83-ab0c-7e47e517712e.png

Figure 3. Working flow of Grad-CAM

Below is an example of Grad-CAM heatmaps applied to different dental disease classifications:

Table 6. Example interpretation using Grad-CAM

Disease Type	Original Image	Grad-CAM Visualization
Caries	31.png	37.png
Gingivitis	32.png	38.png
Hypodontia	33.png	39.png
Mouth Ulcers	34.png	40.png
Tooth Discoloration	36.png	41png.png

5.5 Performance on minority classes and inference time

A key focus of our study was evaluating model performance on underrepresented dental conditions, particularly hypodontia, which is less frequently encountered in datasets. After applying data augmentation techniques, class-wise performance was analyzed (Table 7) (Figure 4).

Table 7. Class-wise performance post-augmentation

5.png

Figure 4. Performance metrics for different disease types

• Data augmentation significantly improved the recall for hypodontia from 78.2% to 82.9%, reducing bias in the model’s predictions.
• Caries and mouth ulcers showed high F1-scores due to the large number of samples available for training.

Additionally, Table 8 shows the inference time measurements across different models:

Table 8. Inference time measurements

• ViT exhibited the longest inference time (250ms), making it less ideal for real-time clinical applications.
• The 5-layer CNN was the best trade-off between accuracy and real-time inference speed, making it the preferred model for deployment.

6.png

Figure 5. Graph comparing the accuracy of different models

The 5-CNN Layer Model & VIT is more efficient for applications requiring high feature extraction and performance on large datasets, despite the trade-offs in training speed and memory usage.

This architecture has been successfully integrated into a scalable web-based framework, ensuring accessibility for dental healthcare professionals. Future work may involve extending the model to additional dental conditions and integrating interpretability techniques to provide visual explanations for predictions (Figure 5).

5.6 Flask-based UI for dental disease detection

To provide an interactive and user-friendly experience, we developed a Flask-based web application that allows users to upload dental images for disease prediction. The UI is designed to be intuitive and efficient, ensuring smooth interaction between users and the deep learning model.

Key Functionalities:

Image Upload:

Users can select and upload an image of a dental condition from their device.

The system accepts common image formats such as JPG, PNG, and JPEG.

Real-Time Prediction:

• Once an image is uploaded, it is preprocessed and passed through the trained model.
• The system predicts the disease type, displaying probability scores for different categories.

Result Display:

• A heatmap generated using Grad-CAM highlights the most relevant areas in the image that influenced the model’s decision.
• The predicted disease type, along with precision, recall, and F1-score, is presented to the user.

User Feedback Mechanism:

• Users can provide feedback if the prediction is incorrect, helping improve model performance in future updates.

Responsive and Accessible Design:

• The UI is designed using HTML, CSS, and Bootstrap, ensuring accessibility across various devices.

Figure 6 shows the screenshots of the UI Workflow.

711.png

72.png

73.png

74.png

75.png

Figure 6. Screenshots of the UI workflow