Analysis of a deep learning-based method for generation of SPECT projections based on a large Monte Carlo simulated dataset

EJNMMI Physics

Table 1 Composition of the activity mask dataset (total of 10,000 masks)

Name of mask	Description	Sample size
Random shapes	Voxelized random shapes	7500
Hot Jaszczak with cold random shapes	Inverted version of Random shapes masks. Here, the entire Jaszczak cylinder with the exception of the random shapes is filled with activity	2000
Chessboard pattern	Randomly sized squares placed on a randomly sized isotropic 3D grid	100
Rod pattern	Rods with randomly large cross sections placed perpendicular to the transverse plane on a randomly sized isotropic 2D grid	100
Cross pattern	Randomly sized cross extended along the axial direction	100
Stripe pattern	Seven stripes with a thickness varying between 2.4 and 16.8 mm placed perpendicular to the transverse plane	100
NEMA phantom spheres	Voxelized activity distribution of the NEMA phantom (six fillable spheres with inner diameters 10/13/17/22/28/37 mm)	100

Before inclusion in the mask, all objects were randomly rotated in the three spatial dimensions (exception: NEMA phantom, random two-dimensional rotation in the transversal plane)