Dental Decay Object Detection

This project trains a Faster R-CNN model to detect dental decay from bitewing images used in dentistry.

Features

• Uses Pytorch Lightning
• COCO dataset style standardized data input
• Faster R-CNN implementation
• Metrics
• Calculation of mAP on test data implemented in model class
• Manually calculating precision, recall, accuracy, specificity
• True Positive, False Negative, ... values calculated for each class

Purpose

• main.py
  • All in one file
  • In later larger projects, code will be separated into several files in order to be easier to read.
  • Contains dataset class, dataloader, model, training, etc.
  • Visualize function is used to visualize the prediction and ground truth
• convert_data_to_coco.py
  • Converts out custom data from .nrrd (3dslicer output) to coco style JPEG and JSON.

Usage

Web GUI

1. Clone this repository
2. Run web.py

cd dental_decay_detection
python web.py

3. Go to http://127.0.0.1:33517
4. You can use the sample.jpg in images directory for testing, otherwise upload your image to see the prediction.

• The web interface supports NRRD, JPG, and PNG file formats.

Training

1. Prepare a JSON and a directory for each of train, val, test. The directory contains images, and each JSON has COCO style annotation data of the images.
2. Use the main.py to train a model on your data. Checkpoint of the model will be saved in lightning_logs directory.
3. In the end, infer.py can be used for inference using the trained model checkpoint.

---

Results

Currently trained on 847 images (697 train, 75 val, 75 test), the resulting mAP value on test dataset is 0.8251.

Bugs

Precision, recall, accuracy, specificity calculation code is corrupt I think. Results are unusual. Fixing in the future...

Demo