Transferring Adult-like Phase Images for Robust Multi-view Isointense Infant Brain Segmentation

Authors

Huabing Liu*, Jiawei Huang, Dengqiang Jia, Qian Wang, Jun Xu, and Dinggang Shen

Citation

@article{liu2024transferring, title={Transferring Adult-like Phase Images for Robust Multi-view Isointense Infant Brain Segmentation}, author={Liu, Huabing and Huang, Jiawei and Jia, Dengqiang and Wang, Qian and Xu, Jun and Shen, Dinggang}, journal={IEEE Transactions on Medical Imaging}, year={2024}, publisher={IEEE} }

Introduction

This repo includes the source codes and pretrained models for our latest work on isointense infant brain segmentation. The two major components are 1) disentangled cycle-consistent adversarial network (dcan) for style transfer between isointense and adult-like phase images; 2) the segmentation network coseg that implements multi-view learning to incorporate adult-like phase images in isointense infant brain segmentation. If you find this repo useful, please give it a star ⭐ and consider citing our paper in your research. Thank you.

1. Create Environment:

• Python 3 (Recommend to use Anaconda)

• NVIDIA GPU + Pytorch

• Python packages:

pip install -r requirements.txt

2. Preparation

Prepare code

Build up the workspace, so that everything can be correctly stored:

sh install.sh

Prapare dataset

For your own dataset, format each data as:

|--<name_of_the_data>
    |-- t1.nii.gz
    |-- t2.nii.gz
    |-- seg.nii.gz

for T1-weighted images, T2-weighted images, and segmentation (if exists), respectively.

Then put formatted data into correct folders:

• for isointense phase images, put them into <pwd>/dcan/data/raw/6m
• for adult-like phase images, put them into <pwd>/dcan/data/raw/12m

Suppose <pwd> is the directory of this repo

Prapare pretrained models

For test-only purpose of this repo, we have shared all the pretrained models:

Method	Model Zoo
dcan	Google Drive
coseg	Google Drive

Put downloaded *.pth into Results folders

3. Run DCAN

Preprocess

Run proc.ipynb

Modify the proc.ipynb:

• for processing isointense phase images:

data_path = 'data/raw/6m'
out_path = 'data/processed/6m'

• for processing adult-like phase images:

data_path = 'data/raw/12m'
out_path = 'data/processed/12m'

Train

Python3 train.py

Synthesize data for downstream segmentation task

Run syn.ipynb

Modify the syn.ipynb:

• for transferring isointense phase images to adult-like contrast

config.dataset.src_dir = 'data/processed/6m'
config.dataset.dst_dir = 'data/processed/12m'

• for transferring adult-like phase images to isointense contrast:

config.dataset.src_dir = 'data/processed/12m'
config.dataset.dst_dir = 'data/processed/6m'

4. Run COSEG

Preprocess

Run proc.ipynb

Modify the proc.ipynb

• for processing source isointense phase images:

data_path = '../dcan/data/processed/6m'
out_path = 'data/processed/6m'

• for processing synthetic isointense phase images:

data_path = '../dcan/data/syn/6m'
out_path = 'data/syn/6m'

• for processing source adult-like phase images:

data_path = '../dcan/data/processed/12m'
out_path = 'data/processed/12m'

• for processing synthetic adult-like phase images:

data_path = '../dcan/data/syn/12m'
out_path = 'data/syn/12m'

Train

sh train.sh

Test

python3 test.py

P.S.

If you find any bugs upon running this repo, please raise an issue in the github page or send me an email (hbliu98@gmail.com).