2025 SNU FastMRI challenge

Team Name: SNUNet
Leaderboard SSIM Score: 0.9756 1st
Final score: 1st place out of 154 teams

1. Model Explanation

Classifier + PromptMRPlus MoE Model

Training consists of two steps:

• Train Classifier for 43 epochs
• Train PromptMRPlus
  • Train knee 4x/8x, brain 4x/8x model for 50 epochs
  • stop brain 4x at 37 epochs
  • stop brain 8x at 35 epochs

2. How to set?

This document assumes a Python 3.12.9 environment.

Install the required packages using the requirements.txt file.

pip3 install -r requirements.txt

Move the image and kspace data in the val folder into the train folder.

python move.py

3. How to train?

First, train the classifier.

sh train_classifier.sh

When training finishes, the model is saved at the following path.

• classifier_model : /root/result/test_BrainKnee/checkpoints/model.pt

Run the .sh files below in order to train the models.

sh train_knee4.sh
sh train_brain4.sh # train up to epoch 36 (0-based index)
sh train_knee8.sh
sh train_brain8.sh # train up to epoch 34 (0-based index)

When training finishes, the models are saved at the following paths.

• knee4x_model : /root/result/submission_knee4/checkpoints/model.pt
• brain4x_model : /root/result/submission_brain4/checkpoints/model.pt
• knee8x_model : /root/result/submission_knee8/checkpoints/model.pt
• brain8x_model : /root/result/submission_brain8/checkpoints/model.pt

4. How to reconstruct?

Place the trained models at the following locations with the following names.

• classifier_model : /root/result/weights/classification_model.pt
• knee4x_model : /root/result/weights/knee4_model.pt
• brain4x_model : /root/result/weights/brain4_model.pt
• knee8x_model : /root/result/weights/knee8_model.pt
• brain8x_model : /root/result/weights/brain8_model.pt

Run reconstruct.sh to perform the reconstruction.

sh reconstruct.sh

5. How to leaderboard_eval?

Set yp in the leaderboard_eval.sh file to '/root/result/weights/reconstructions_leaderboard'. (It is already set in the .sh file.)

sh leaderboard_eval.sh

If a dependency problem occurs with cv2, run the commands below.

apt-get update
apt-get install -y libgl1-mesa-glx