This script inserts doublet and singlet signals into DDPM generated MR images from demo 1. The singlet versus doublet signals for different signal lengths are determined using the 2AFC-based detection table provided below. This, in turn, means that singlet and doublet signals are set based on the acceleration factor, signal contrast, and signal length (in pixels). This code saves the objects with signals in HDF5 format.

Command-line input options:

  acceleration (int)              : Acceleration factor for sparse sampling (2, 4, 6, or 8).
  contrast (float)                : Signal amplitude/contrast value.
  signal_lengths (str)            : Comma-separated signal separation lengths, e.g. "4,5,6,7,8".
  object_hdf5_path (str, optional): Path to the DDPM-generated objects from demo 1.

Output The output HDF5 files are saved to ./objects/. Each file contains datasets H_s (singlet reconstructions), H_d (doublet reconstructions), and L_list (signal lengths).

Usage:

python signal_insertion_test.py [acceleration factor] [contrast] [signal_lengths] [object_npz_path]

Examples: Run with acceleration factor 4 corresponding to the 7th row in the 2-AFC table below (also employed for testing in our DLMO paper):

python signal_insertion_test.py 4 0.7 '4,5,6,7,8'

A couple of MR images with the doublet signal corresponding to the demo run.

A couple of MR images with the siglet signal corresponding to the demo run.

Note that the limiting conditions for different acceleration factor using iFFT-based reconstruction was determined across a series of 2AFC studies performed by a trained non-physician reader as shown below. Each cell corresponds to one 2AFC study defined by a specific combination of signal intensity and signal length. Colored cells indicate conditions under which the reader achieved $100%$ accuracy on the Rayleigh discrimination task. Note that whenever $100%$ accuracy was achieved at a higher acceleration factor, the same signal condition also yielded $100%$ accuracy at lower acceleration factors. For example, a signal with intensity 1.3 and length 8 mm resulted in perfect accuracy at acceleration factors of $8\times$, then the same $100%$ accuracy is expected at $1\times$, and $4\times$ accelerations. Also, for a given signal intensity, note that whenever $100%$ accuracy was achieved at a particular signal length, all longer signal lengths yielded $100%$ accuracy as well. For example, for the acceleration factor $1\times$, $100%$ accuracy was observed for a signal intensity of $1.3$ at a length of $4$ mm. The same held true for all lengths greater than $4$ mm at the same $1.3$ intensity for $1\times$. Here 1px = 1mm.

The shaded box indicates combinations of intensity and signal length used to generate testing images that encompass limiting conditions (for acceleration factors 4 and 8), to evaluate whether AI-based reconstruction provides better discriminatory capability than conventional iFFT-based reconstruction.

Furthermore, we want our imaging system and its reconstruction method to perform as well as possible under limiting conditions, (i.e., allow users to detect and discriminate the most challenging combinations of signal length and signal intensity.)

The relevance of this 2AFC study can also be inferred from its close alignment with the smallest signal length (~4 mm) and the lowest signal intensity (~0.3), which are similar to those observed in non-specific white lesion (NSWL) distributions from the fastMRI+ dataset. For this analysis, we include only NSWLs with height and width < 20 px.

From these plots, we observe that NSWLs can be as small as 4–5 px in both width and height, with lower-end intensity values ranging from approximately 50/256 to 75/256. These limiting sizes and intensity values are similar to those obtained in our 2AFC study when MR acquisition is performed at a fully sampled rate.