BEN

A generalized Brain Extraction Net for multimodal MRI data from rodents, nonhuman primates, and humans

See also: | Github project link

Z. Yu, X. Han, S. Zhang, J. Feng, T. Peng and X. -Y. Zhang, "MouseGAN++: Unsupervised Disentanglement and Contrastive Representation for Multiple MRI Modalities Synthesis and Structural Segmentation of Mouse Brain," in IEEE Transactions on Medical Imaging, 2022, doi: 10.1109/TMI.2022.3225528.

We release BEN (version 0.2) during November based on the reviewers' suggestions and our experiences accumulated in clinical practice.

(Old version is moved to 'doc' branch)

Please refer to list.

Coincidentally, the motivation of BEN's training strategy is somewhat consistent with Cellpose v2. BEN tries to quickly develop a customized model specific to each user application with the help of the AdaBN module.

🚀 Quick start to use BEN or replicate our experiments in 5 minutes!

Feature	Description	Colab link
Transferability & flexibility	BEN outperforms traditional SOTA methods and advantageously adapts to datasets from diverse domains across multiple species [1], modalities [2], and MR scanners with different field strengths [3].	[1] [2] [3]
Quality assessment	Unlike traditional toolboxes, which mainly rely on manual inspection to assess the brain extraction quality, BEN incorporates a quality assessment module to automatically evaluate its brain extraction performance.
Speed 🚀	Inference: Less than 1 sec/scan; Adaptation to a new imaging center/application: About 10 min for deployment.	-

Name	Description	Related Fig. in paper	Snapshot	Colab link
1. BEN workflow & architecture	BEN renovates the brain extraction workflow to adapt to multiple species, modalities and platforms	Fig. 1 & 2		This repository
2. ✨ Cross species	Deploying BEN to new species requires minimal or even no labels.	Fig. 3		Cross species
3. Cross modalities / MR scanners	Deploying BEN to new modalities/MR platforms requires minimal or even no labels.	Fig. S1 & S2		Cross modalities Cross MR scanners
4. Compare with other toolbox	BEN outperforms traditional SOTA methods and advantageously adapts to datasets from various domains across multiple species, modalities, and field strengths.	Fig. 4
5. Uncertainty & inter-rater variations	Compared to traditional toolboxes, BEN provides a measure of uncertainty that potentially reflects rater disagreement	Fig. 5		Disagree map Uncertainty
6. ✨ Volumetric quantification for longitudinal studies	BEN improves the accuracy of atlas registration and benefits brain volumetric quantification compared with using other toolbox.	Fig. 6 & 7		Registration comparision Volumetric quantification

The details can be found in this folder.

Visit our documentation or video tutorials for installation, tutorials and more.

• Installation
• Quick Start / Tutorial
  • Cross modalities
  • Cross MR scanner with different field strengths
  • Cross species
  • Try your data
  • Video tutorials
• Resources
  • Interface for neruoimaging toolboxes
  • Data release
  • Pretained weight
  • Contributing to BEN

An Nvidia GPU is needed for faster inference (less than 1 sec/scan on 1080ti gpu).

Requirements:

• tensorflow-gpu == 1.15.4
• Keras == 2.2.4
• numpy == 1.16
• SimpleITK == 2.0
• opencv-python == 4.1
• scikit-image == 0.16.2

Install dependencies:

git clone https://github.com/yu02019/BEN.git
cd BEN
pip install -r requirement.txt

The target domain data folder looks like this: (Download data from this repository/Colab or put your data here.)

• All the undermentioned results can be repeated via our tutorial Notebook.
• New weight will be saved independently for further customized application.

1. Modality: T2WI -> EPI

2. For this exemplar domain adaptation (DA) task, No label is used (zero-shot).

3. From top row to the third row: Raw image, Baseline result, BEN's result.

   

1. MR scanner with different field strengths: 11.7 T -> 7 T

2. For this exemplar domain adaptation (DA) task, No label is used (zero-shot).

3. From top row to the third row: Raw image, Baseline result, BEN's result.

   

1. Species: Mouse -> Rat

2. For this exemplar domain adaptation (DA) task, only ONE label is used.

3. The segmentation results are shown in red, the ground truth are shown in orange.

4. From top row to the fifth row: Raw image, Zero-shot (0 label used), finetune (1 label used), BEN's result (1 label used), Ground truth.

   

5. (Optional) Just do some simple postprocessing here, e.g., only save the top-K largest connected regions.

6. Compared with other methods, it further shows BEN's advantages

   

Feel free to try your data or deploy BEN to your preprocessing pipeline. Details can be found in notebook and video tutorials. Pretrained weights can be found in dataset_release.

# Update BEN (domain adaptation)
python BEN_DA.py -t train_folder -l label_folder -r raw_image_folder -weight pretrained_weight_path -prefix new_model_name -check check_orientation
# Run inference
python BEN_infer.py -i input_folder -o output_folder -weight model_weight_path -check check_orientation

To further validate BEN’s generalization, we have evaluated BEN on two new external public ex-vivo MRI datasets (rTg4510 mouse: 25 ex-vivo scans, and C57BL/6 mouse: 15 ex-vivo scans). When only one label is used for BEN adaptation/retraining, impressive performance is achieved on both datasets, despite the fact that BEN was originally designed for in-vivo MRI data.

Dataset	Used label	Description	Automatically generated reports	Video link
rTg4510 mouse	1	Ex-vivo scans with obvious distortion and different orientations.		youtube
C57BL/6 mouse	1	Ex-vivo scans. There is no obvious gap between the brain and the skull borderlines, making the task difficult.		youtube
C57BL/6 mouse	0 (zero-shot)	Ex-vivo scans. The domain gap exists in ex-vivo MRI data and in-vivo images in our training images could be so large that it compromises the performance. In this case, we suggest users add several labels to update BEN.		youtube

Pretrained weight used in tutorials can download from Google Drive

The usages and details can be found in this folder.

The details can be found in this folder.

If you find our work / datasets / pretrained models useful for your research, please consider citing:

@article{yu2022generalizable,
  title={A generalizable brain extraction net (BEN) for multimodal MRI data from rodents, nonhuman primates, and humans},
  author={Yu, Ziqi and Han, Xiaoyang and Xu, Wenjing and Zhang, Jie and Marr, Carsten and Shen, Dinggang and Peng, Tingying and Zhang, Xiao-Yong and Feng, Jianfeng},
  journal={Elife},
  volume={11},
  pages={e81217},
  year={2022},
  publisher={eLife Sciences Publications Limited}
}

@dataset{yu_ziqi_2022_6844489,
  author       = {Yu Ziqi and
                  Xu Wenjing and
                  Zhang Xiao-Yong},
  title        = {{A longitudinal MRI dataset of young adult C57BL6J 
                   mouse brain}},
  month        = jul,
  year         = 2022,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.6844489},
  url          = {https://doi.org/10.5281/zenodo.6844489}
}

Disclaimer: This toolkit is only for research purpose.