Skip to content

Latest commit

 

History

History
executable file
·
103 lines (78 loc) · 7.31 KB

README.md

File metadata and controls

executable file
·
103 lines (78 loc) · 7.31 KB
Raw

Reproducing PromptMR results on fastMRI MultiCoil Knee dataset

All the following commands should be run under the promptmr_examples/fastmri folder.

Download the fastMRI MultiCoil Knee dataset

Download knee multicoil_train and multicoil_val set from fastMRI website: https://fastmri.med.nyu.edu/.

Split the multicoil_val dataset

Since the online evaluation platform for original multicoil_test set is no longer available (see here), we can run the following command to automatically split original 199 cases in multicoil_val set into 99 validation cases and 100 test cases. Before running, please first rename the original multicoil_val folder to multicoil_val_origin. (If you have already downloaded the multicoil_test, rename that folder to multicoil_test_origin as well.) Then, download the sampling mask files I used for 100 splited test cases from here (fastMRI random type sampling mask, 46 files for acc=4, 54 files for acc=8) . Finally, executing the following command will automatically generate our split multicoil_val and multicoil_test folders. Please modify the data_path and mask_path to reflect your own dataset and sampling mask file locations.

python split_val_test.py \
--data_path /research/cbim/datasets/fastMRI/knee_multicoil/multicoil_val_origin \
--mask_path /research/cbim/datasets/fastMRI/knee_multicoil/multicoil_test_mask

Pretrained models

We provide Google Drive links for downloading our models trained on the fastMRI MultiCoil Knee train set.

Model Trained on Training Acceleration # of Params NMSE/PSNR/SSIM 4x NMSE/PSNR/SSIM 8x Link
E2E-VarNet train+val 4x and 8x 30M 0.0053/39.37/0.9236 0.0087/37.30/0.8936 official repo 1
HUMUS-Net train 8x only 109M - 0.0090/37.20/0.8946 official repo 2
HUMUS-Net-L train 8x only 228M - 0.0086/37.45/0.8955 official repo 2
PromptMR train 8x only 80M - 0.0080/37.78/0.8983 my link1
PromptMR train 4x and 8x 80M 0.0051/39.71/0.9264 0.0080/37.78/0.8984 my link2

Note: In this table, train and val means the original train set and validation set, respectively. Evaluations were performed on the split test set (100 cases), which is from original validation set (199 cases). E2E-VarNet and HUMUS-Net(-L) were assessed using their respective official pretrained models. However, E2E-VarNet has only made available the model that was trained on the combined training and original validation sets with acceleration factors of 4x and 8x. This implies that our split test set was already exposed to E2E-VarNet during its training phase. Consequently, the results of E2E-VarNet should be considered for reference purposes only and not subjected to rigorous comparison with other models.

You can directly use the following command to download our PromptMR models to the pretrained_models folder:

mkdir pretrained_models
cd pretrained_models
# model trained with acc=8 only
gdown 1HBlwrmOaMQycohznYrg-uu4GmiVfWjkL
# model trained with acc=4 and 8
gdown 1afLCO3C_S4e-q7QCt04Ksmv34jETrFQ7

Inference

The following command will reproduce the results of the pretrained PromptMR model on the fastMRI MultiCoil Knee Test Set. Please modify data_path and output_path to reflect the locations of your dataset and output reconstruction, respectively. The state_dict_file specifies the path of the downloaded pretrained model.

CUDA_VISIBLE_DEVICES=0 python run_pretrained_promptmr_fastmri_knee_inference.py --challenge varnet_knee_mc \
--state_dict_file pretrained_models/promptmr-12cascades-epoch=43-step=764324.ckpt \
--data_path /research/cbim/datasets/fastMRI/knee_multicoil/multicoil_test \
--output_path /research/cbim/vast/bx64/PycharmProjects/fastmri_results/reproduce_promptmr_knee

Evaluate

The following commands will evaluate the results of the pretrained PromptMR models on the fastMRI MultiCoil Knee x8 Test Set. Please modify the target-path, predictions-path and test-path to your own path correspondingly. acceleration filters the test cases with the specific acceleration factor.

python evaluate.py \
--target-path /research/cbim/datasets/fastMRI/knee_multicoil/multicoil_test_full \
--predictions-path /research/cbim/vast/bx64/PycharmProjects/fastmri_results/reproduce_promptmr_knee/reconstructions \
--test-path /research/cbim/datasets/fastMRI/knee_multicoil/multicoil_test \
--challenge multicoil \
--acceleration 8

Training

The following command will train the PromptMR model on the fastMRI MultiCoil Knee Training Set. Please modify the data_path to reflect your own dataset location. center_fractions specifies the low frequency fractions, while accelerations defines sampling rate of the mask. The mask_type is the type of the sampling mask (for the fastMRI knee dataset, we use the random type). Use no_use_ca to disable channel attention . The checkpoints and log files will be saved in the folder specified by exp_name . The use_checkpoint enables a technique that trades compute for memory, which is useful when GPU memory is limited.

# train with acc=8 only
CUDA_VISIBLE_DEVICES=0,1 python train_promptmr_fastmri.py \
--challenge multicoil \
--center_fractions 0.04 \
--accelerations 8 \
--mask_type random \
--data_path /research/cbim/datasets/fastMRI/knee_multicoil \
--exp_name promptmr_train \
--num_gpus 2 \
--no_use_ca \
--use_checkpoint

# train with acc=4 and 8
CUDA_VISIBLE_DEVICES=0,1 python train_promptmr_fastmri.py \
--challenge multicoil \
--center_fractions 0.08 0.04 \
--accelerations 4 8 \
--mask_type random \
--data_path /research/cbim/datasets/fastMRI/knee_multicoil \
--exp_name promptmr_train \
--num_gpus 2 \
--no_use_ca \
--use_checkpoint

FAQ

Q: How is multicoil_test_mask generated?

A: The masks were generated randomly using the same code that used for applying random masks to fully-sampled data during training. generate_test_masks.py contains the code for generating the masks we provided in google drive.