FALCON for Motion Correction in PET Imaging - Your Tracer Experiences?
LalithShiyam opened this issue · 13 comments
Hey GitHub enthusiasts! Have you explored the incredible potential of FALCON for motion correction in PET imaging studies using various tracers? Or have you gone beyond and utilized it for other modalities as well? Share your experiences and let's geek out together!
- [18F]PSMA-1007
- [18F]flubatine
- [18F]PI-2620
@rullator Man, you are going at it :D hope FALCON is serving you well :D
Absolutely!
New scanner - new questions - new challenges :)
I´m currently testing FALCON with 13N-Ammonia PET cardiac imaging. So far, I would say that rigid and affine modes don´t work at all. The optimization gets stuck at no motion at all maybe due to the limited axial FOV of the scanner (16 cm). Deformable registration seems to work good for late frames, but no so good for early frames (bolus).
Hi there @mnamias, thanks for your feedback.
I am not surprised that it doesn't work on the very early frames. Nor do I know any algorithms that can handle such high varying tracer activity at the start. We tried to make this work by using cGANs to convert earlier frames to late frames - but this involves quite a bit of work. Anyhow, currently this is a work that we are doing with Copenhagen @DrLyngby, he can tell you more about it, may be @Keyn34 can also update!
Just curious - why would rigid/affine work with cardiac images, as they undergo heavy contractions? Did you expect it to work?
Hi @LalithShiyam , thanks for your prompt reply!
Just curious - why would rigid/affine work with cardiac images, as they undergo heavy contractions? Did you expect it to work?
We are performing dynamic scans for blood flow quantification. We don´t want to compensante heart contraction, we want to compensate patient motion which is very frequent during pharmacologic stress. Does the similarity metric (NCC) exclude voxels with zero values? Since we have signal all around the FOV, trying to move one frame would generate low NCC values due to overlapping zero values with non-zero values at the edge of the FOV, wouldn´t it?
@mnamias I would just go for deformable registration. Since it's around the cardiac region. And I am quite positive it's not going to work for early phase with fast changing tracer kinetics. It needs a different approach, where n+1th frame is registered with nth frame.
And NCC is your best bet for deformable registration as NMI is not really suited as the cost function can be reduced with unrealistic deformations.
Happy to look into it deeper if you can share a sample dataset (anonymized of course).
Cheers,
Lalith
@DrLyngby do you have any ammonia datasets we can test?
Hi all,
I'm happy to tip in on the experiences I've had with FALCON, for 82Rb MPI assessments.
So far, I've focused the testing of FALCON on static (perfusion) images in attempts to correct for cardiorespiratory motion, with the preliminary results for corrections for cardiac contractions described in the recent JNM paper. In my experience, all correction methods worked well, with the best results obtained using deformable registrations.
(https://jnm.snmjournals.org/content/jnumed/64/7/1145.full.pdf)
Re: inter-frame motion correction for dynamic reconstructions: Unfortunately, I haven't obtained reliable results for the MoCo throughout the dynamic series. However, I've also found that late-frame motion correction works quite well. On this topic: Piotr's Slomka's group, however, has recently reported a method that works quite well. @LalithShiyam: Is it possible to adapt FALCON to provide similar corrections?
(https://jnm.snmjournals.org/content/early/2023/11/30/jnumed.123.266208.abstract)
@LalithShiyam : Unfortunately, I don't have access to any NH3 data at the moment.
Cheers,
Martin
Dear @DrLyngby , thanks for your input on this topic!
@LalithShiyam Here you can find one dynamic 13N-Ammonia PET from our institution:
https://drive.google.com/file/d/1J3rrzrRLUYmTI7R6mmMW0kpMgc9mPOw2/view?usp=sharing
Best,
Mauro