Skip to content

mdciri/teeth-segmentation

Repository files navigation

Teeth Segmentation

Teeth segmentation using Pytorch and Monai. The data used belong to the Tufts Dental Database.

Data split

The Thfts dataset is composed by 1000 images in total which were randomly divided as 70-20-10% in training, validation, and test set. The split is done by:

pyhton data_split.py

which save the dataset lists in the data.json file. It stores also the class names and the class weights for the segmentation.

Model

Simple UNet network created using monai.netowrks.

Training

pthon train.py -md "<model_directory>" -d "mps" -g 0 -bs 16 -lr 1.e-4 -ne 100

This scripts runs the training scrits. It:

  • saves the best model at the specified "<model_directory>",
  • uses the device "mps" (but you can use "cuda" or "cpu")
  • GPU at position 0,
  • batch size 16,
  • learning rate 1.e-4,
  • 100 number of maximum epochs.

Evaluation

pthon evaluation.py -md "<model_directory>" -d "mps" -g 0 -bs 1 

This script computes evaluation over the training, validation, and test dataset saving the dice score per each image in the file evaluation_results.csv.

GUI

pthon gui.py

This scripts runs a simple GUI where the user can upload an image, visualize the model's segmentation prediction, and, in case, save it.