https://github.com/chenliu-1996/deepcontrast_demo

This is the repository that accompanies the submission "Deep Learning Substitutes Gadolinium in Detecting Functional and Structural Brain Lesions with MRI", under review at Nature Biomedical Engineering.

https://github.com/chenliu-1996/deepcontrast_demo

Repository

This is the repository that accompanies the submission "Deep Learning Substitutes Gadolinium in Detecting Functional and Structural Brain Lesions with MRI", under review at Nature Biomedical Engineering.

https://github.com/ChenLiu-1996/DeepContrast_Demo/blob/master/

# DeepContrast_Demo
This is the repository that accompanies the submission **Deep Learning Substitutes Gadolinium in Detecting Functional and Structural Brain Lesions with MRI**, *under review at Nature Biomedical Engineering*.

Due to file size limits of GitHub (nothing > 100 MB allowed), the NifTI scans and model weights cannot be uploaded to this repository. Instead, we stored them in [Google Drive](https://drive.google.com/drive/folders/1l5GU6E0iCHbs24ZNzN6uIQgbQud1DZ3e?usp=sharing) and currently these data and files shall be publicly available. Once missing files and model weights are downloaded from Google Drive and placed at the correct locations, an experienced deep learning researcher shall be able to replicate the results we reported in the test-retest reliability study.

Due to the Google Drive sharing settings in our institution, any new incoming request to access the shared folder will need our permissions to proceed. We aim to process the request within the same day. 

Please note that the code was developed on Linux (Ubuntu 16.04 LTS), and it may require some adjustments if you intend to run it on a different operating system. Future updates may provide cross-platform versions or OS-specific versions upon request.


## Overview of this repository
```
DeepContrast_Demo
     (#) Test_retest_data_complete
    |    (#) NatureBME_share_nonContrast
    |    (#) NatureBME_share_GBCAuptake
    |    (#) NatureBME_share_GBCApredicted
    |    (#) NatureBME_share_brainMask
    |    (#) NatureBME_share_tissueLabel
    |
     Healthy_Human_Brain_Model
    |    deep_learning_model
    |   |    data_loader.py
    |   |    network.py
    |   |    solver.py
    |   |
    |    (#) saved_model_weights
    |        (#) ResAttU_Net-SGD-0.1000-CVPR_Adaptive_loss-4-epoch18.pkl
    |
     Demo_scripts
    |    generate_new_predictions.py
    |    verify_old_new_predictions_identical.py
    |    visual_inspection.py
    |    test_retest_evaluation.py
    |
     Newly_generated_prediction
    |
     Environment_setup
         DeepContrast.yml

(#): Download from Google Drive.

```
### Explanations of the folders:
**Test_retest_data_complete** is supposed to contain all data from the test-retest reliability dataset. However, as GitHub has a strict data upload limit of 100 MB, we have to omit the actual data files from this repository, and instead make these files available on [Google Drive](https://drive.google.com/drive/folders/1l5GU6E0iCHbs24ZNzN6uIQgbQud1DZ3e?usp=sharing).

**Healthy_Human_Brain_Model** contains the backbone of the Healthy Human Brain Model (both the architecture and the trained model weights) introduced in our manuscript. Currently only the customized testing code is made available. Again, since the model weights (543 MB) exceeds the file size limit, we have to keep the "saved_model_weights" folder empty and only share that over [Google Drive](https://drive.google.com/drive/folders/1l5GU6E0iCHbs24ZNzN6uIQgbQud1DZ3e?usp=sharing).

**Demo_scripts** contain the four sample scripts to demonstrate the model.

**Newly_generated_prediction** is an empty folder and will be filled with new predictions once "./Demo_scripts/generate_new_predictions.py" is executed.

**Environment_setup** contains the anaconda configuration file "DeepContrast.yml" with which one can quickly configure an environment suitable to run our scripts. If it doesn't work on your machine, you would probably need to manually install the required packages. More details can be found [here](https://github.com/SAIL-GuoLab/DeepContrast_Demo/tree/master/Environment_setup).

*More detailed descriptions can be found in the respective folders.*

## Preparing the Anaconda Environment to Execute the Code.
You can either create an anaconda environment from the ./Environment_setup/DeepContrast.yml file or use the following commands to prepare the environment. If you don't have anaconda installed, you can refer to detailed instructions at "./Environment_setup/".
```
conda create -n DeepContrast

conda activate DeepContrast
conda install python=3.7 numpy scipy scikit-image scikit-learn seaborn -c anaconda
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
conda install nibabel tqdm -c conda-forge
conda install cudatoolkit=10.2 -c pytorch
```

## Authors
Chen Liu*, Nanyan Zhu*, Dipika Sikka, Xinyang Feng, Haoran Sun, Xueqing Liu, Sabrina Gjerswold-Selleck, Hong-Jian Wei, Pavan S. Upadhyayula, Angeliki Mela, Peter D. Canoll, Cheng-Chia Wu, Andrew F. Laine, Jeffrey A. Lieberman, Frank A. Provenzano, Scott A. Small, Jia Guo, for the Alzheimers Disease Neuroimaging Initiative.

**Chen Liu and Nanyan Zhu contributed equally to this work and are joint first authors.**

**Correspondance: Jia Guo (jg3400@columbia.edu).**

## Link to Manuscript
Liu C. & Zhu N., et al. Deep Learning Substitutes Gadolinium in Detecting Functional and Structural Brain Lesions with MRI. 26 August 2020, PREPRINT (Version 1) available at Research Square. Under Review at Nature Research. https://doi.org/10.21203/rs.3.rs-56518/v1

Please note that the version above is not the most up-to-date version yet. There are minor differences between the one above and the latest version submitted to the chief editor via the manuscript tracking system.

## Code availability
The trained Healthy Human Brain Model, alongside the test-retest reliability dataset (n = 11, each with two test-retest acquisitions) with both non-contrast scans and ground truth GBCA-uptake maps, is available on GitHub (link to be announced). The scripts that predict GBCA-uptake maps from non-contrast scans, as well as the script performing quantitative evaluations, are included. All code and data (except for those from public datasets) are proprietary and managed by the Columbia Technology Ventures Office of Intellectual Property. The custom training code and large-scale datasets are not publicly available.

## Data availability
The authors declare that all data supporting the results in this study are available from the corresponding author J.G. upon reasonable request, after permission from the Columbia Technology Ventures Office of Intellectual Property.

Owner

GitHub Events