Deep learning project for automated tooth decay detection and classification from smartphone microphotography images using YOLOv5. The system identifies and classifies different stages of tooth tissue destruction, demonstrating how computer vision can support early dental diagnosis and screening.
- Developed an end-to-end deep learning pipeline for dental image analysis.
- Applied YOLOv5 for multi-class tooth decay detection and classification.
- Classified three stages of tooth tissue destruction from smartphone microphotography images.
- Achieved up to 78.3% mAP@0.5 on the evaluation dataset.
- Investigated the impact of data augmentation on model performance.
Dental caries (tooth decay) is one of the most common oral health conditions worldwide. Early detection plays a critical role in preventing disease progression and reducing the need for invasive treatment.
Traditional diagnosis relies on visual examination by dental professionals. However, advances in computer vision create opportunities for automated screening systems that can assist clinicians and improve accessibility to dental assessment.
This project explores whether deep learning models can accurately detect and classify different stages of tooth tissue destruction directly from smartphone microphotography images.
Automated dental image analysis presents several challenges:
- Small and subtle visual differences between decay stages.
- Variations in lighting, image quality, and camera positioning.
- Limited dataset size compared to large-scale computer vision benchmarks.
- Similar visual characteristics between adjacent decay categories.
- Class imbalance across disease stages.
To improve model robustness, image augmentation techniques were incorporated during training and the performance impact was evaluated.
The project followed the following workflow:
- Collect smartphone microphotography images of teeth.
- Annotate tooth decay regions and assign disease-stage labels.
- Apply image preprocessing and augmentation techniques.
- Train a YOLOv5 object detection model.
- Evaluate performance using mAP, Precision, Recall, F1 Score, and Confusion Matrices.
- Compare model performance under different training configurations.
- Analyze strengths and limitations of the final system.
The dataset contains labeled dental microphotography images representing three stages of tooth tissue destruction.
- Visible Change Without Cavitation
- Visible Change With Microcavitation
- Visible Change With Cavitation
- Bounding Box Detection
- Multi-Class Classification
The dataset is not included in this repository due to availability restrictions.
YOLOv5 was selected due to its strong balance between detection accuracy and computational efficiency.
The model was trained to simultaneously:
- Localize regions affected by tooth decay.
- Classify decay severity into one of three categories.
- Generalize across varying image conditions through augmentation.
| Metric | Value |
|---|---|
| mAP@0.5 | 78.3% |
| Classes | 3 |
| Model | YOLOv5 |
The model successfully distinguished between different stages of tooth decay while maintaining strong detection performance across the evaluation dataset.
Training and validation metrics demonstrate consistent convergence and performance improvement throughout the training process.
The confusion matrix illustrates class-level prediction performance and highlights the model's ability to distinguish between different decay stages.
The Precision-Recall curve demonstrates strong overall detection performance, with particularly strong results for advanced decay classes.
The F1 curve was used to identify optimal confidence thresholds and evaluate the balance between precision and recall.
- Deep learning can effectively assist in automated dental image analysis.
- YOLOv5 achieved strong performance for multi-class tooth decay detection.
- Data augmentation improved model robustness and generalization.
- Advanced decay stages were identified with particularly high accuracy.
- The project demonstrates the potential of AI-assisted dental screening systems.
tooth-decay-detection-using-deep-learning/
│
├── README.md
│
├── notebooks/
│ └── yolov5_training.ipynb
│
├── reports/
│ └── tooth_decay_project_report.pdf
│
└── assets/
├── results_with_augmentation.png
├── confusion_matrix_with_augmentation.png
├── PR_curve_with_augmentation.png
└── F1_curve_with_augmentation.png
- Python
- YOLOv5
- PyTorch
- OpenCV
- NumPy
- Matplotlib
- Jupyter Notebook
Potential applications include:
- AI-Assisted Dental Screening
- Clinical Decision Support
- Early Caries Detection
- Oral Health Monitoring
- Medical Image Analysis
- Computer-Aided Diagnosis
A detailed technical report describing the methodology, experiments, and findings is available in:
reports/tooth_decay_project_report.pdf