Skip to content

DATASET.md

Latest commit

 

History

History
91 lines (70 loc) · 3.06 KB

DATASET.md

File metadata and controls

91 lines (70 loc) · 3.06 KB

Prepare the dataset

The dataset split information is available in the configs/data_split folder.

CMRxRecon 2023 & 2024

We need to convert the original MATLAB training dataset to H5 format for faster slice data reading during training. Run the following command:

python prepare_h5_dataset_cmrxrecon.py \
    --input_matlab_folder /path/to/MICCAIChallenge2024/ChallengeData/MultiCoil \
    --output_h5_folder /path/to/cmrxrecon2024/h5_dataset \
    --split_json configs/data_split/cmr24-cardiac.json \
    --year 2024

The script first converts the MATLAB dataset into H5 files, which are saved under the h5_dataset folder.

For example, /path/to/MICCAIChallenge2024/ChallengeData/MultiCoil/Cine/TrainingSet/FullSample/P001/cine_sax.mat will be converted to /path/to/cmrxrecon2024/h5_dataset/P001_cine_sax.h5

After conversion, the script splits the dataset based on a predefined JSON file located in the configs/data_split folder using symbolic links.

The saved H5 file structure is as follows:

/path/to/cmrxrecon2024
│   ├── h5_dataset
│   │   ├── P001_T1map.h5
│   │   ├── P001_T2map.h5
│   │   ├── P001_cine_lvot.h5
│   │   ├── P001_cine_sax.h5
│   │   ├── P001_cine_lax.h5
│   │   └── ...
│   ├── train
│   │   ├── P001_T1map.h5 (symbolic link)
│   │   ├── P001_T2map.h5 (symbolic link)
│   │   ├── P001_cine_lvot.h5 (symbolic link)
│   │   ├── P001_cine_sax.h5 (symbolic link)
│   │   └── ...
│   ├── val
│   │   ├── P001_cine_lax.h5 (symbolic link)
│   │   └── ...

The mask_radial.h5 file required for training on the CMRxRecon2024 dataset is available here.

FastMRI-knee

To split the dataset, follow the instruction in the PromptMR repo.

FastMRI-brain

Download the dataset directly from the fastMRI website.

CC-brain

Original coil data Fixed coil data
Proposed Method Baseline Method

The original H5 dataset has an issue where the phase infomation is not processed correctly when using the challenge official script. To fix this, we need to preprocess the dataset using the following command:

python prepare_h5_dataset_cc_brain.py \
    --input_folder /path/to/calgary-campinas_version-1.0/CC359/Raw-data/Multi-channel/12-channel \
    --output_folder /path/to/cc-brain \
    --split_json configs/data_split/cc-brain.json

The saved folder structure is as follows:

/path/to/cc-brain
│   ├── poisson_sampling # Poisson sampling masks
│   ├── train
│   ├── val
│   ├── test_full
│   ├── test_acc05
│   └── test_acc10