We released preprocessing FS, V-Net, and UNETR codes.
Collaborators: Chaeyeon Lim, Juyoung Hahm, Jisoo Lee, Kyungsu Kim
Detailed instructions for testing the 3D images are as follows.
TesnorFlow and PyTorch implementation are based on original V-Net & UNETR code.
The original code of V-Net, UNETR are the following:
V-net: https://github.com/faustomilletari/VNet]
UNETR: https://github.com/Project-MONAI/research-contributions/tree/main/UNETR/BTCV
matplotlib
monai
tqdm
shutil
pdb
pandas
numpy
time
pytorch
tensorflow
Please request an email (kskim.doc@gmail.com) for the inference sample data (as this work is under review, it is open to reviewers only).
Instructions on how to run FreeSurfer are uploaded in FS_readme.txt.
Set segmentation file into 'folder_number'_seg.
Please read VNET_readme.txt to run the V-Net code.
Using skull-striped volume from FreeSurfer as input for V-Net, we obtained the labels and GT in pickle format through FreeSurfer.py, nii_to_pkl.py.
With the command 'python main_vnet.py,' the deep learning (DL) models' training and evaluation described in the research paper are conducted.
V-Net performs the evaluation using GPU and CPU and presents the processing time for each frameworks.
Please read UNETR_readme.txt to run the UNETR code.
Create 'Image_Tr' and 'Image_Ts' folder and obtain the labels and GT.
Train the model with the command 'python main_unetr.py' and evaluate with the command 'python eval_unetr.py'.
UNETR performs the evaluation using GPU and CPU and presents the processing time for each frameworks.
Segmentation results of a) CNN-based V-Net and b) ViT-based UNETR (left 3D images in first column and red-highlighted areas in second column) and FS (right 3D images in first column and blue-highlighted areas in second column) for each brain structure.
V-Net [https://github.com/faustomilletari/VNet] (Thanks to Fausto Milletari and Sagar Hukkire)
UNETR [https://github.com/Project-MONAI/research-contributions/tree/main/UNETR/BTCV] (Thanks to Ali Hatamizadeh and other contributors)