Thread by @pezlarson

Synthesizing medical images is now easy to do with deep learning and good datasets, but what is it good for? I've viewed this with a lot of skepticism, but have now started to accumulate some potential use cases, specifically for synthetic #PET in #PETMR
1/n

Maybe I'll start with what I think it is probably NOT good for - diagnosis!

Fundamentally, the information content in synthetic medical images is derived from the distribution and patterns in the data that the model was trained on.

2/n

For an individual patient, this means that any synthetic images created will just contain the information already in their individual (non-synthetic) images combined with population-level distributions and patterns from the training data.

3/n

Synthetic images are not adding any specific information about the individual that was not already present in the input. If synthetic images are expected to **add anything for diagnosis**, this seems like a dangerous use case to me
4/n

But there are use cases for non-diagnostic applications, here's a few I can think of
- synthetic CT for radiation therapy planning
- synthetic/pseudo CT for MR-based attenuation correction in #PETMR
- (new ideas) synthetic #PET for improved PET reconstruction in #PETMR
5/n

Synthetic PET for MR-based Attenuation Correction algorithm development/evaluation #PETMR ieeexplore.ieee.org/document/99553…
Generate realistic PET based on #MRI, so can compare PET reconstructions objectives with different MRI approaches without PET scan
@absudabsu @thomashopemd
6/n

Synthetic PET as a Prior to improve PET reconstruction using #MRI data doi.org/10.1117/12.258…
Since the synthetic PET based on #MRI has closer correspondence to actual PET scan, synthetic PET provides a better prior when constraining PET reconstruction
@ndwork @absudabsu
7/n

Maybe I'll add more examples of other areas onto this thread too (requests?)
(n-epsilon)/n