Table 3.

Broad level comparison of core approaches to using AI in PET image reconstruction.

	Direct AI methods	Unrolled AI methods	AI methods without training data
Interpretability	Low	Good	Good
Use of physics model	No	Yes	Yes
Generalisation (outside training distribution)	Can be challenging	Better than direct methods	Good
Suitability for fully 3D image reconstruction	Challenging	Yes	Yes
Typical number of training pairs	>10,000 (2D reconstruction)	~50 (3D reconstruction)	1
Inference time/time to reconstruct from new data	Fast	Potentially the slowest	Faster than unrolled methods, depending on implementation
Expected image quality if within training distribution	Potentially the best	Good	Does not benefit from population data

2D, two-dimensional; 3D, three-dimensional; AI, artificial intelligence; PET, positron emission tomography.