Stanford Knee MRI Multi-Task Evaluation (SKM-TEA) 2021 Dataset

This project folder contains the configuration files, preprocessing, and visualization scripts for the Stanford Knee MRI Multi-Task Evaluation (SKM-TEA) 2021 dataset.

Related papers:

• https://openreview.net/forum?id=YDMFgD_qJuA.

Visualization An example notebook for visualizing the data is provided in the visualize notebook. You just need to set the path where the dataset is downloaded. The original notebook is copied from the StanfordMIMI/skm-tea repository and provided by the SKMTEA authors.

Preprocessing No preprocessing is needed for the SKMTEA dataset. You just need to generate train, val, and test sets depending on the task you use the dataset for. For example, for the reconstruction task, you need to run the generate_sets.sh script.

Training/Testing

Important

The SKM-TEA dataset is natively supported in atommic. Therefore, you do not need to create a custom dataset class. You just need to set the dataset_format argument in the configuration file to the desired SKM-TEA dataset version. Also the FFT needs to be centered. For example:

model:
    fft_centered: true
    fft_normalization: ortho

    train_ds:
        dataset_format: skm-tea-echo1
        fft_centered: true
        fft_normalization: ortho

    validation_ds:
        dataset_format: skm-tea-echo1+echo2
        fft_centered: true
        fft_normalization: ortho

    test_ds:
        dataset_format: skm-tea-echo1+echo2-mc
        fft_centered: true
        fft_normalization: ortho

The skm-tea-echo1 dataset contains only the first echo of the multi-echo data. The skm-tea-echo2 dataset contains only the second echo of the multi-echo data. The skm-tea-echo1+echo2 dataset sums the first and second echoes of the multi-echo data. The skm-tea-echo1+echo2-mc dataset stacks the first and second echoes of the multi-echo data as channels.

For training a model, you just need to set up the data and export paths to the configuration file of the model you want to train. In train_ds and validation_ds please set the data_path to the generated json files. In exp_manager please set the exp_dir to the path where you want to save the model checkpoints and tensorboard or wandb logs.

You can train a model with the following command:

atommic run -c /projects/REC/SKMTEA/conf/train/{model}.yaml

For testing a model, you just need to set up the data and export paths to the configuration file of the model you want to test. In checkpoint (line 2) set the path the trained model checkpoint and in test_ds please set the data_path. In exp_manager please set the exp_dir to the path where the predictions and logs will be saved.

You can test a model with the following command:

atommic run -c /projects/REC/SKMTEA/conf/test/{model}.yaml

Note: The default logger is tensorboard.