Ctrl+K

Deep MRI Reconstruction with Radial Subsampling (SPIE 2022)

Deep MRI Reconstruction with Radial Subsampling (SPIE 2022)#

This folder contains the training code specific for our experiments presented in our paper Deep MRI Reconstruction with Radial Subsampling accepted and presented at SPIE Medical Imaging Conference 2022 (Arxiv pre-print: https://arxiv.org/abs/2108.07619).

Dataset#

We used the Calgary-Campinas public brain multi-coil MRI dataset which was released as part of an accelerated MRI reconstruction challenge. The dataset is consisted of 67 3D raw k-space volumes collected on a Cartesian grid (equidistant). After cropping the 100 outer slices, these amount to 10,452 slices of fully sampled k-spaces which we randomly split into training (40 volumes), validation (14 volumes) and test (13 volumes) sets (see lists/).

Training#

The standard training command direct train can be used. Training model configurations can be found in the configs/ folder.

After downloading the data to <data_root> a command such as the one below is used to train the RIM models:

direct train <experiment_directory> \
            --cfg projects/spie2022_radial_subsampling/configs/base_<radial_OR_rectilinear>.yaml \
            --training-root <data_root>/Train/ \
            --validation-root <data_root>/Val/ \
            --num-gpus <number_of_gpus> \
            --num-workers <number_of_workers> \
            [--resume]

For further information about training see training.

The validation volumes can be computed using predict_val.py (see Validation).

During training, training loss, validation metrics and validation image predictions are logged. Additionally, Tensorboard allows for visualization of the above.

Note#

  • For the experiments with Radial Subsampling we used base_radial.yaml to train the RadialRIM

  • For the experiments with Rectilinear Subsampling we used base_rectilinear.yaml to train the RectRIM

Inference#

Validation#

To make predictions on the validation set (14 volumes, see lists/val/) a command such as the one below is used to perform inference on dataset with index <dataset_validation_index> as various datasets can be defined in the training configuration file.

cd projects/
python3 predict_val.py <output_directory> \
                --cfg projects/spie2022_radial_subsampling/configs/base_<radial_OR_rectilinear>.yaml \
                --checkpoint <path_or_url_to_checkpoint> \
                --data-root <data_directory_containing_test_samples_as_in_`lists/val/val.lst`>
                --filenames-filter spie2022_radial_subsampling/lists/val/val.lst
                --validation-index <dataset_validation_index> \
                --num-gpus <number_of_gpus> \
                --num-workers <number_of_workers> \

Test#

To make predictions on the test set (13 volumes, lists/test/) a command such as the one below is used to perform inference on the inference dataset as defined in the inference section of the inference/configs/<acceleration>x/base_<subsampling_type>.yaml configuration file.

direct predict  <output_directory> \
                --data-root <data_directory_containing_test_samples_as_in_`lists/test/test.lst`>
                --cfg projects/spie2022_radial_subsampling/configs/inference/<acceleration>x/base_<radial_OR_rectilinear>.yaml
                --checkpoint <path_or_url_to_checkpoint> \
                --filenames-filter projects/spie2022_radial_subsampling/lists/test/test.lst
                --num-gpus <number_of_gpus> \
                --num-workers <number_of_workers> \
Contents