tatasauce4life/single_coil_knee

fastMRI Single-Coil Knee (Preprocessed for Deep Learning)

This dataset contains preprocessed slices from the fastMRI Knee Single-Coil Validation set. The original data was converted from .h5 format to normalized .npy slices for faster loading and training.

Dataset Details

• Source: NYU fastMRI Knee Dataset
• Format: .npy (NumPy arrays)
• Shape: (2, 320, 320) representing the Real and Imaginary components of the k-space.
• Total Slices: 5,769 slices
• Split:
  • train/: 180 volumes (approx. 90%)
  • val/: 19 volumes (approx. 10%)

Preprocessing Pipeline

1. Slice Selection: Removed the first 20% of slices from each volume to exclude non-informative data.
2. Inverse FFT (IFFT): Converted raw k-space to image space.
3. Cropping: Central crop to (320, 320) in the image domain.
4. Normalization: Volume-wise max-normalization based on image intensity.
5. FFT: Converted back to k-space after cropping and normalization.
6. Storage: Saved as float32 complex-valued components (Real, Imag).

Usage

import numpy as np
import os

# Load a sample slice
data = np.load('s0000_000.npy')
print(data.shape) # Output: (2, 320, 320)

Citation

If you use this dataset, please cite the original fastMRI paper: @article{zbontar2018fastmri, title={fastMRI: An Open Dataset and Benchmarks for Accelerated MRI, author={Zbontar, Jure and others}, journal={arXiv preprint arXiv:1811.08839}, year={2018} }