metadata

license: afl-3.0
tags:
  - medical
size_categories:
  - 100M<n<1B

Dataset for AVS-Net Pre-training

The dataset utilized in the pre-training of the AVS-Net: Attention-based Variable Splitting Network for P-MRI Acceleration model, developed by Y Zhang, J Li, Z Wang, J Duan, and J Li, incorporates data from five distinct protocol sequences. These are:

(coronal_pd)Coronal Spin Density-weighted without Fat Suppression
(coronal_pd_fs)Coronal Spin Density-weighted with Fat Suppression
(sagittal_pd)Sagittal Spin Density-weighted
(sagittal_t2)Sagittal T2-weighted with Fat Suppression
(axial_t2)Axial T2-weighted with Fat Suppression

The dataset is structured on a slice-by-slice basis, with each slice containing 20 cases. Each case is comprised of two files: rawdata*.mat and espirit*.mat. The dataset's structure can be outlined as follows:

Dataset architecture:

name: /rds/projects/d/duanj-ai-in-medical-imaging/knee_fast_mri
Protocol: [coronal_pd, coronal_pd_fs, sagittal_pd, sagittal_t2, axial_t2]

Approximately 40 slices per protocol, each slice containing 15 channels, with a height and width (HW) of (640, 368)

```
knee_nyu
    - axial_t2  coronal_pd(X)  coronal_pd_fs  sagittal_pd  sagittal_t2
        |           |           |               |           |
    - [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], [11, 12, 13, 14, 15, 16, 17, 18, 19, 20] masks
                    |                                       |
    -            [train]                                  [val]
        |                                                            |
    - espirit*.mat(1-40), rawdata*.mat(1-40)                       *_masks.mat
```

In this structure, each protocol has approximately 40 slices, each consisting of 15 channels. The dimensions of the data are 640x368 (height x width). For each protocol, the slices are further divided into two groups: the training set ([train]) and the validation set ([val]). The training set includes the espirit*.mat and rawdata*.mat files for each slice, while the validation set contains *_masks.mat files.

Dataset Usage

For a standalone knee dataset download, use git lfs(https://git-lfs.com/) to download from the huggingface datasets(https://huggingface.co/datasets/AVS-Net/knee_fast_mri):

# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone -j8 git@hf.co:datasets/AVS-Net/knee_fast_mri

Known Issues and Resolutions

1. Network Connection Issue

For enhanced network connection quality, it is recommended to employ the ssh protocol instead of https.

# Rather than utilizing `https://huggingface.co/datasets/AVS-Net/knee_fast_mri`
# Clone the repository using `git@hf.co:datasets/AVS-Net/knee_fast_mri`
# As an example:
git clone -j8 git@hf.co:datasets/AVS-Net/knee_fast_mri

1. Interruptions During Download

Certain error messages may appear during the download process due to interruptions. These errors can include:

error: ... : cannot add to the index - missing --add option?
batch response: Post ... : read: connection reset by peer
error: failed to fetch some objects from 'https://hf.co/datasets/AVS-Net/knee_fast_mri.git/info/lfs'

Following the instructions below allows for the handling of these interruptions.

# Navigate (`cd`) to the directory containing the `lfs` folder
# Intead of using `git pull`,
# Use `git lfs pull` to resume the download progress for `lfs` projects

git lfs pull

Please note that this process will resume the download from where it was interrupted, thereby ensuring the integrity of your downloaded data.