listmode DL notebooks for PSMR 2024

SyneRBI / SIRF-Exercises

SIRF Training and demonstration material

http://www.ccpsynerbi.ac.uk

Apache License 2.0

17 stars 21 forks source link

listmode DL notebooks for PSMR 2024 #225

Closed gschramm closed 2 months ago

gschramm commented 2 months ago

Don't merge yet, because:

first needs merging of SIRF PR https://github.com/SyneRBI/SIRF/pull/1253
clean-up files + final conversion of ??_*.py notebooks to jupyter notebook

gschramm commented 2 months ago

@KrisThielemans for adding the LM caching:

here is where I setup the ListModeData object
here is where I setup the listmode logL and reconstruct

KrisThielemans commented 2 months ago

Some quick comments:

Is your random prefix time-frame dependent? They will need to have different scale, as well as ExamInfo. (We could come up with a function to create one, and then scale it down, but it's going to be ugly, and we have no time.)
can speed-up things with set_max_segment_num_to_process in the reconstructor (possibly also in ListmodeToSinograms)
https://github.com/gschramm/SIRF-Exercises/blob/2cdda25975855a4e4c96be8f5507b2712c73e2df/notebooks/Deep_Learning_listmode_PET/01_SIRF_listmode_recon.py#L199C1-L199C44 is wrong comment
try lm_obj_fun.set_max_cache_size(large_number). probably need to set cache_path for reliable operation

gschramm commented 2 months ago

Isn't it the easiest / most efficient if we only unlist + reconstruct the complete (1m and 60min) list mode files. Should avoid all kind of problems ...
is the "large number" in bytes / MB / GB?
Just starting reconstructing the 60min data after downloading. But I immediately get this error when reading the mumap file:
```
attn_image = sirf.STIR.ImageData("../../data/PET/mMR/NEMA_IQ/20170809_NEMA_MUMAP_UCL.v.hdr")
```

error: ??? '\'\nERROR: Error parsing: expected a vectorised key as in "image duration (sec)[1]", but no bracket found\n\' exception caught at line 417 of /users/nexuz/gschra2/nge/SIRF/build_lm/sources/SIRF/src/xSTIR/cSTIR/cstir.cpp; the reconstruction engine output may provide more information'


The header looks like this:

!IMAGE DATA DESCRIPTION:= !total number of data sets:=1 !image duration (sec):=0 !image relative start time (sec):=0

KrisThielemans commented 2 months ago

sure
bytes
yes, the full notebook contains extra stuff to handle this, but easiest is to use the same attenuation image and norm (or indeed acq_sens_model) as for the 1 min data. Note that there will be scale factor of 60 (or nearly, due to decay) between the images, as that's currently not yet taken into account.

gschramm commented 2 months ago

Thanks Kris! Will update the notebooks accoordingly tmr. To take care of the scale difference for the 60min ref. image, I will probably just rescale based on the mean intensity.

KrisThielemans commented 2 months ago

I'll merge this in a few minutes

gschramm commented 2 months ago

@KrisThielemans ready to be merged

KrisThielemans commented 2 months ago

squash or not? As you prefer

gschramm commented 2 months ago

please squash, but let me fix one typo before

gschramm commented 2 months ago

@KrisThielemans : typos fixed, ready to be squashed and merged