This example shows forward projection and reconstruction of DDPM-generated objects (SOMs) using the rSOS method to create a test dataset. It saves the reconstructions in HDF5 format.

Command-line Options:

acceleration_factor (int)       : Acceleration factor for sparse sampling (2, 4, 6, or 8).
object_hdf5_path (str, optional): Path to the DDPM-generated SOMs with added signals from demo 2.

Usage:

python rsos_ddpm_test.py [acceleration_factor] [object_hdf5_path]

Examples: Run with acceleration factor 4:

python rsos_ddpm_test.py 4

Input files are .hdf5 files obtained from Demo 2. The reconstructed images are saved in HDF5 format in the ./rsos_rec/ folder. Each HDF5 file contains the following datasets: H_s for singlet image reconstructions, H_d for doublet image reconstructions, and L_list for the signal lengths corresponding to each reconstructed image.

A couple of MR SOMs with the doublet and singlet signals as inputs, $f(r)$ , to this code.

The FFT data collected at each coil is modeled as

$$ g_i = \Phi \mathcal{F} S_i f(r) + n_i, $$

where $n_i$ denotes zero-mean Gaussian noise at each coil with a standard deviation set as 15. Our forward model uses 8 coils ($S_i$) and a Poisson disk–based subsampling pattern ($\Phi$), as shown below:

Then reconstruction at each coil is combined using iFFT in the follow manner:

$$ \begin{align} \hat{f}_i = \mathcal{F}^{-1} g_i, \\\ \hat{f}_{\text{rSOS}} = \sqrt{\sum_{i=1}^{N_c}|\hat{f}_i|^2}. \end{align} $$

Both the accelerated and fully sampled rSOS reconstructions are saved in the final HDF5 files using this script.