Swin-Fetal-Brain-Segmentation

This Swin-UNETR has been developed to automatically segment fetal brain from 3D fetal functional MR images and codes are based on this repository. Here you can find:

• Python code to Train or fine tune your Swin-UNETR on your dataset using weight random inizialization or our pre-traind weigth from a rs-fMRI fetal brain segmentation task (reccomanded option);
• Python code to Test our model on new rs-fMRI fetal scans;
• A folder logdir where the output are saved during training;
• A folder output where the output are saved during testing;
• Images folders: imagesTr and labelsTr (Train and validation images and label as defined in json file), imagesTs and labelsTs (Test images and label)
• A json folder which contains two jsons file for training and test;
• Images and Swin pretrain weight can be downloadeed from here.

Results of Swin-UNETR model and ground truth with fetal rs-fMRI scan.

Before to start

• We reccomand to use Conda - see here.
• Create a new conda environment with python 3.10.8 and install the pytorch.yml file.
• Download the images and pretrain weights from here and:
  • Place the pretrain weight on the 'weight' folder.
  • If you want to try our images just download them and place it on the correct folder;
  • If you want to use your own images place it on the correct folder and create jsons file with the new images;
• Change the main path (data_dir - 'Insert/your/path') on line 59 of Train.py and line 62 of Test.py
• You can also install a CNN we reccomand the nnUNET. Please see here.

How to use it

Be sure to work on a visible GPU.

• Open the terminal;
• Activate the new conda enviroment;
• Enter 'python path/to/your/Train.py' or 'python path/to/your/Test.py'

Citation

Pecco, N., Della Rosa, P. A., Canini, M., Nocera, G., Scifo, P., Cavoretto, P. I., Candiani, M., Falini, A., Castellano, A., & Baldoli, C. (2024). Optimizing performance of transformer-based models for fetal brain MR image segmentation. Radiology: Artificial Intelligence, 0(ja), e230229. https://doi.org/10.1148/ryai.230229