Table 2:

A summary of the types of PET images used for validation.

Type of images	Associated ground-truth or surrogate of truth	Realism of image characteristics	Realism of tumors	Computational time	Convenience
Synthetic images (no simulation of physics beyond addition of blur and noise to the ground-truth)	Perfect (voxel-by-voxel)	Low	Low to high. Depends on the digital phantom used.	Low	Easy to produce in large numbers.
Simulated images (e.g. with GATE (Le Maitre et al., 2009; Papadimitroulas et al., 2013) or SIMSET (Aristophanous et al., 2008))	Perfect (voxel-by-voxel)	Medium to High	Low to high. Depends on the digital phantom used.	High	Implementation is not straightforward. Time consuming. A proprietary reconstruction algorithm is not easily available.
Physical phantom acquisitions	Imperfect (relies on known geometrical properties + associated high resolution CT).	High (real)	Usually simplified objects. Depends on the physical phantom used.	N/A	Requires access to a real scanner and phantom. Can be time consuming.
Clinical images	Approximate	High (real)	High (real)	N/A	Rare datasets, difficult to generate. Digitized histopathology measurements are full of potential errors.
	Approximate (Consensus of manual delineations by several experts).	High (real).	High (real).	N/A	At least three manual contours are recommended. Time consuming.