[Castor-users] Scatter rate estimation for LM-TOF-MLEM

[STUTE Simon]

Hello Jakub,

Thanks for the detailed email !

Do you take the scan duration into account in the datafile header ?

You say that you estimated the scatter rate for each TOF bin independently to be 1-2. Is it "1 or 2" (which would be insanely high) or "e-2" ?

How did you compute this estimated scatter rate ?

And how did you translate this information into your list-mode file ?

Best
Simon






________________________________________
De : Castor-users <castor-users-bounces at lists.castor-project.org> de la part de jbaran <jbaran at ifj.edu.pl>
Envoyé : lundi 6 décembre 2021 21:13
À : Castor Users
Objet : [Castor-users] Scatter rate estimation for LM-TOF-MLEM

Dear CASToR Users and Developers,

me and my colleagues found an issue with scatter correction implemented
in CASToR (based on simulated data). I will start from the beginning.

We performed a simulation with GATE (NEMA IEC phantom) for the long
axial FOV PET scanner and tried to reconstruct the results (true+scatter
events) with scatter correction information incorporated. We estimated
the scatter rate for each TOF bin independently (it is about 1-2 per TOF
bin). We also specify the size of the TOF bin in the header. For the
comparison, as the reference, we also reconstructed 2 other cases:
- only true coincidences
- true+scatter coincidences without any correction
For all three cases we use the same TOF resolution. All the images are
done with LM-MLEM and raytracing projector. All the reconstructed images
are attenuation corrected. You can find in the attached plot the
reconstructed images from 20th iteration (NEMA_IEC_factor1.png):
TOP LEFT: true + scatter coincidences scatter corrected
TOP RIGHT: true + scatter coincidences
BOTTOM LEFT: true coincidences
BOTTOM RIGHT: profiles along the orange line from previous plots

Unfortunately, the images reconstructed with true+scatter coincidences
with scatter correction do not look like those reconstructed only with
true. They are more like the one reconstructed with true+scatter events.
Even the values are at the same level as for the images reconstructed
with true+scatter events. Moreover, we also dig a little bit in the
CASToR source code and dump forward projection values and it seems that
the factor of the image estimation is three levels of magnitude greater
than the additive correction factor. It suggests that probably our
scatter correction does not work correctly.

So we started to play around with scatter correction rates. We did some
other exercises and multiplied our scatter rate by several factors
ranging from 1 to 10E6. I attached the results for the multiplicative
factor equal to 20000 (NEMA_IEC_factor20000.png). The legend is as for
the previous image. The results with the scatter correction applied
start to look like images reconstructed with true coincidences. However
the scatter effect (increased signal in the middle of NEMA IEC) inside
the phantom is visible. Image quality (factor_20000 images) at the
boundaries of the scanner is better. Additionally, the overall voxel
units are still at the level of the voxels of images reconstructed with
true+scatter coincidences. Study with other factors also does not affect
image in the expected way. So, we think that at some point our scatter
correction method might not work properly but we are also wondering if
the scatter correction in CASToR is also implemented correctly.

To check it we performed one more study - simulation of the cuboid
phantom with very high statistics. We performed the reconstruction
without TOF resolution for such cases:
- true coincidences
- true + scatter coincidences
- true + scatter coincidences with scatter correction performed by
CASToR internally.
The resulted reconstructed images are shown in cuboid.png:
TOP LEFT: true coincidences
TOP RIGHT: true + scatter coincidences
BOTTOM LEFT: true + scatter coincidences with scatter correction
performed by CASToR internally
BOTTOM RIGHT: profiles along the orange line from previous plots
The number of scatter coincidences registered per LOR is around 10-30
coincidences.
What we found is that the images reconstructed with true+scatter and
true+scatter with incorporated scatter correction look almost the same
in terms of image quality and voxel activity units, which is surprising.

In conclusion, we are wondering if you have ever found similar things
and could help us understand why the scatter correction does not work
(especially for the last case). The second thing is that maybe there is
a feature of the CASToR which we do not consider.

We will appreciate your help.

All the best,
Jakub