Skip to content

Latest commit

 

History

History
81 lines (55 loc) · 3.69 KB

README.md

File metadata and controls

81 lines (55 loc) · 3.69 KB

JoB-VS

This repository provides a PyTorch implementation of JoB-VS: Joint Brain-Vessel Segmentation in TOF-MRA Images presented in ISBI 2023. JoB-VS performs a joint-task learning for brain and vessel segmentation (JoB-VS) in Time-of-Flight Magnetic Resonance images, being an end-to-end vessel segmentation framework. Unlike other vessel segmentation methods, our approach avoids the pre-processing step of implementing a model to extract the brain from the volumetric input data. Our method builds upon Towards Robust General Medical Image Segmentation with a segmentation head that allows the simultaneous prediction of the brain and vessel mask.

Paper

JoB-VS: Joint Brain-Vessel Segmentation in TOF-MRA Images
Natalia Valderrama1, Ioannis Pitsiorlas2, Luisa Vargas1, Pablo Arbeláez1*, Maria A. Zuluaga2
ISBI 2023.

1 Center for Research and Formation in Artificial Intelligence (CINFONIA), Universidad de Los Andes.
2 Data Science Department, EURECOM, Sophia Antipolis, France

Installation

Cloning the repository

$ git clone https://github.com/BCV-Uniandes/JoB-VS.git
$ cd JoB-VS
$ python setup.py install

Dataset Preparation

  1. Download your data and create a json file in the OASIS3 format. Here you can find an example of how the data must be organized. Specify the root to your data in the json file.

  2. Set the data_root, out_directory and num_workers variables in the file data_preprocessing.py and run the command:

python data_preprocessing.py

Your data will be organized in the following way:


Fold_X
|_ imagesTr
|_ |_ *.nii.gz
|_ imagesTs
|_ |_ *.nii.gz
|_ labelsTr
|_ |_ *.nii.gz
|_ dataset.json
|_ dataset_stats.json

Our benchmark is setup for 2 folds.

  1. (optional) If data doesn't have any labels, as in the IXI dataset, please use this file data_preprocessing.py.

Training and evaluating the models

We train JoB-VS on the original images, without using brain masks, and then we fine-tune the models using Free AT, as done in ROG:

# For training on original images
python main.py --gpu GPU_IDs --batch BATCH_SIZE --fold FOLD --data_ver OUT_DIRECTORY --name OUTPUT_DIR

# For the Free AT fine tuning
python main.py --gpu GPU_IDs --batch BATCH_SIZE --fold FOLD --data_ver OUT_DIRECTORY --name OUTPUT_DIR_FREE_AT --ft --pretrained OUTPUT_DIR --AT

For evaluating the models, modify the EXPS_PATH, PATH_ANNS and PATH_PREDS in the file run_evaluations.py:

python run_evaluations.py

If you want to make inference with our models, please download our weights in this link and run:

# For the training on original images
python main.py --gpu GPU_IDs --batch BATCH_SIZE --data_ver YOUR_DATA --name OUTPUT_DIR --load_weights WEIGHTS_PATH --test

If you are using the ixi dataset, please add the (--ixi) flag.

MONAILabel APP

Please find all the information for the MONAILabel app in the branch monai.