SmileGuard — Dental Detection Training Pipeline

YOLOv8-based dental detection model that identifies teeth, caries, cavities, cracks, and other intraoral conditions from dental images.

Features

• 9-class detection: tooth, caries, cavity, crack, calculus, gingivitis, hypodontia, mouth ulcer, tooth discoloration
• Dual dataset pipeline: Merges DentalAI (DatasetNinja) annotations with the Kaggle 6-Class Intraoral Disease Dataset
• Automated class balancing: Over/under-sampling with augmentation
• Patient reports: Human-readable detection summaries
• Flask API: REST endpoint for integration

Prerequisites

• Python 3.9+
• CUDA-capable GPU recommended (CPU works but is slow)
• Kaggle API credentials configured for dataset download

Setup

1. Clone the repo and install dependencies:

   pip install -r requirements.txt

2. Download the DentalAI dataset from DatasetNinja and place the dentalai-DatasetNinja/ folder in the project root:

3. Download the Roboflow dataset too https://universe.roboflow.com/ds/VIfxZmemET?key=IZUynBzy09

   

4. Open the notebook and run cells sequentially:

   dental_detection_complete.ipynb
   

Pipeline Overview

Step	Description
1	Download Kaggle 6-Class Intraoral Disease Dataset
2	Convert DatasetNinja annotations → YOLO format
3	Convert Kaggle classification images → YOLO format
4	Merge both datasets into a unified set
5	Validate merged dataset
6	Analyze & balance class distribution
7	Train YOLOv8 model
8	Evaluate model performance
9	Run inference with DentalDetector class
10	Export Flask API (app.py)

Training Configuration

The training cell auto-detects GPU/CPU. Edit the CONFIG dict in Step 7 to adjust:

CONFIG = {
    "model": "yolov8s.pt",   # or yolov8m.pt / yolov8l.pt
    "epochs": 100,
    "batch_size": 16,         # reduce to 8 or 4 if out-of-memory
    "img_size": 640,
    "device": 0,              # 0 = GPU, "cpu" = CPU
}

Note: Set device to 0 if you have a CUDA GPU. The notebook defaults to CPU if no GPU is detected.

API Usage

After training, a Flask API is generated at app.py:

python app.py

Endpoint	Method	Description
/detect	POST	Send a base64-encoded image, receive detections
/health	GET	Health check & loaded class list

Project Structure

SmileGuard-Train/
├── dental_detection_complete.ipynb   # Main training notebook
├── requirements.txt                  # Python dependencies
├── app.py                            # Flask API (generated by notebook)
├── yolov8s.pt                        # Pre-trained YOLOv8 weights
├── dentalai-DatasetNinja/            # DatasetNinja source dataset
├── yolo_dataset/                     # Converted YOLO labels (DatasetNinja)
├── kaggle_yolo/                      # Converted YOLO labels (Kaggle)
├── yolo_dataset_merged/              # Combined dataset
├── yolo_dataset_balanced/            # Balanced dataset (used for training)
└── Dental_Detection/                 # Training outputs & weights

Troubleshooting

Issue	Fix
Low accuracy	Add more data or increase epochs to 150–200
False positives	Raise confidence threshold to 0.6–0.7
Out of memory	Reduce batch_size to 8 or 4
Slow training	Use yolov8n.pt or reduce img_size to 416

License

See dentalai-DatasetNinja/LICENSE.md for dataset license terms.