MRISeqClassifier

MRISeqClassifier: A Deep Learning Toolkit for Precise MRI Sequence Classification

MRISeqClassifier is an open-source toolkit designed for classifying sequences (series) of MRI scans.

News

[2024 1127] Paper is accepted by AMIA 2025 Informatics Summit

[2024 0919] Paper submit to medRxiv

[2024 0917] Paper submit to AMIA 2025 Informatics Summit

[2024 0912] Github available

[2024 0806] The project started

Install

Create a new conda environment and install the required packages:

conda create -n nacc_img python=3.9
conda activate nacc_img

git clone git@github.com:JinqianPan/MRISeqClassifier.git
cd MRISeqClassifier
pip install -r requirements.txt

Important

In the requirements.txt, the PyTorch installation is not included.

Please use nvidia-smi to verify your CUDA version and then download the appropriate version of PyTorch from the official website.

Download Data And Best Models

If you want to use another data and train the model by yourself, you could skip this step.

Data

This toolkit is training by NACC MRI dataset. Please request data from https://naccdata.org/.

Tip

For how to download the image data from NACC AWS S3 bucket, please browse my guide: NACC Image Data Download Experience.

For how to preprocess NACC MRI data, please browse my guide: NACC MRI Data Processing.

After preprocessing the data, please change the IMAGE_PATH in the yaml file 00_config.yml.

Best Models

Please download best models from Google Drive, and put them into folder 02_models/best_model.

Note

If you are using your own dataset, please construct the training and testing sets as follows: (we use torchvision.datasets.ImageFolder to build data loader)

01_data
    |- ImageFolder
           |- train
                |- T1WI (label name)
                    |- MRI.jpg
                |- T2WI
                |- ...
           |- test
                |- ...

Results for the Remaining MRI Sequence Types in the NACC Dataset

We offer the result for the rest NACC MRI Sequence type in Google Drive.

Quick start

If you just want to use this tool to classify MRI images, please use the code below:

$PATH = 'your image path'
python 05_toolkit.py --path $PATH

Important

Just remember: Before training, this toolkit can only classify 6 types mentioned above.

If you want to use your own data, please go through the step training.

Training

For training data, you could use code below, or use slurm to run the code.

nohup python -u 03_training.py --proximal middle --model densenet121 --epoch 100 --fold 10 > ./output/10-Fold/denseNet121/output_mid.log 2>&1 &

After training all models, you could use code below to get the result of voting:

nohup python -u 04_training_vote.py --proximal middle --fold 10 > ./output/10-Fold/vote/output_mid.log 2>&1 &

Citation

If you use MRISeqClassifier in your research, please cite our paper:

@article{pan2024mriseqclassifier,
  title={MRISeqClassifier: A Deep Learning Toolkit for Precise MRI Sequence Classification},
  author={Pan, Jinqian and Chen, Qi and Sun, Chengkun and Liang, Renjie and Bian, Jiang and Xu, Jie},
  journal={medRxiv},
  pages={2024--09},
  year={2024},
  publisher={Cold Spring Harbor Laboratory Press}
}