In order to explore the ability of GP-based sampling and reconstruction to effectively subsample MR images, we first needed to establish a training set of data. In particular, we selected the IXI dataset, which contains 600 3D MRI datasets of healthy brains, some taken at 1.5 T and some taken at 3.0 T
36 . To partition the dataset into testing and training portions, we selected 7391 2D 256
pixel axial images
from T1-weighted 3D MR datasets in NIFTI format, reshaped them into
pixel images, where
and
, and then zero-padded the left and right edges to make the images into
squares with each pixel representing 1.2
1.2 mm
2. For each image, the intensity
was normalized by its
. These images were then converted to
-space data
using a FFT and cropped to retain the central
(
) region for computational efficiency. These
-space data were randomly divided into
,
, and
partitions for training and testing. Additional diffusion weighted images were collected as 2D MR images using a Philips
Achieva 1.5 T system with a gradient
s/mm
2 (b1000) and the apparent diffusion coefficient map (ADC) was calculated.
Xu Y., Farris C.W., Anderson S.W., Zhang X, & Brown K.A. (2023). Bayesian reconstruction of magnetic resonance images using Gaussian processes. Scientific Reports, 13, 12527.