DentalSeg: A Solution for ToothFairy2 Challenge

The proposed solution is based on nnU-Net v2 framework. The inference code has been rewritten to enable the algorithm to run easily and rapidly.

Usage

Following these steps to integrate DentalSeg with nn-UNet:

1. Download and install nnUNetv2 using the command

  git clone https://github.com/MIC-DKFZ/nnUNet.git
  cd nnUNet
  pip install -e .

2. Copy the loss functions and network training code files to the corresponding directories in nnUNet using the following commands:

cp DentalSeg/loss/* nnUNet/nnunetv2/training/loss/
cp DentalSeg/nnUNetTrainer/* nnUNet/nnunetv2/training/nnUNetTrainer/

3. Our solution includes two stages, which means we need to train two networks. Please follow the official commands of nnUNetv2.

   Network1 Experiment Planning and Preprocessing

   nnUNetv2_plan_and_preprocess -d [d1]  -c 3d_fullres -np 4
   

   Network1 Training

   nnUNetv2_train [d1]  3d_fullres [fold] 
   

   Network2 Experiment Planning and Preprocessing

   nUNetv2_plan_and_preprocess -d [d2]  -c 3d_fullres -np 4
   

   Network2 Training

   nnUNetv2_train [d2]  3d_fullres [fold] 
   

   Network2 Finetuning

   nnUNetv2_train [d2] 3d_fullres [fold] -tr nnUNetTrainer_Tversky_no_mirror
   

To inference by running the commond after modifying the file address in inference.py:

python inference.py