AC CT output seems to be faulty

[Keyn34]

Hey everyone.

We previously used your tool successfully to generate AC-CTs from different dose level PET images. Please take a look at the publication if you are interested.

However, when we reran the tool on other subjects, we noticed that the AC-CT results were suboptimal.

This is an example:

The OPTOF-NAC image is on the right, and the generated AC-CT is on the left. As you can see, the images are flipped and mostly empty.

I used only DICOM as input, since I want to use the generated AC-CTs for further reconstructions.
I tested it on a Windows 10 PC and a Ubuntu 22.04 PC, both running Python 3.11 and 3.10 with NVIDIA RTX 3090s.

Do you have any idea what the culprit here is? Thanks a bunch in advance.

[claesnl]

Thank you for posting the issue! I think the issue could be that the dcm2nii-converter reads the direction from the DICOM data wrong, and therefore flips the input, which causes the model to perform worse (although it was trained with flip as part of the augmentation) and I am also seeing a relatively hypointense bladder, which could offset the normalization. In any case, I am confused with the distribution of predicted values, as they are not CT-like at all.

Which version of CTlessPET did you use? I recently updated the model, including preprocessing and model weights, to version 0.2. Can you check which version you use? (e.g. cat ~/CTlessPET_params/FDG_version).

Would it be possible to share at least a subset of the dicom slices so I can try to reproduce the error if it persists in the newest version?

[Keyn34]

Thank you for the response!

I have noticed other errors related to dicom2nifti, but they originate from the package, not your code.
For example, this one, which I fixed by adding the suggested line of code in the site-packages folder of my virtual environment.

I was also confused by that. Because we have seen your tool perform well during our previous study.

I installed and tested this run with the latest version, 0.2.3.
cat ~/CTlessPET_params/FDG_version also gives 0.2

I am not allowed to share DICOM images unfortunately, but I will confirm this again. Would the NIFTI variant help you? I did not try to run it with NIFTI input only.

[claesnl]

Excellent, then you are on the latest version for sure.
Thank you for sharing the dicom2nifti bug - this confirms that something might be off with the normalization.
If you are able to share the NiFTI version I would happily try to debug - it should be sufficient to reproduce the error on my side.

[Keyn34]

Please find a sample set here. I have included the original CT and the OPTOF-NAC PET. Since I already had them in NIFTI, I just ran CTlessPET with the FDG model on the NIFTI files to check if that changed anything. Unfortunately, I got the same result.
I have also included this resulting synthetic CT in the upload.

This was the output in the terminal:

########################
If you are using CTlessPET, please cite the following paper:

Montgomery ME, Andersen FL, d’Este SH, Overbeck N, Cramon PK, Law I, Fischer BM, Ladefoged CN. Attenuation Correction of Long Axial Field-of-View Positron Emission Tomography Using Synthetic Computed Tomography Derived from the Emission Data: Application to Low-Count Studies and Multiple Tracers. Diagnostics. 2023; 13(24):3661. https://doi.org/10.3390/diagnostics13243661
########################

100%|███████████████████████████████████████████████████████████████████████████████| 2814/2814 [01:04<00:00, 43.47it/s]
100%|███████████████████████████████████████████████████████████████████████████████| 2814/2814 [01:05<00:00, 43.18it/s]
100%|███████████████████████████████████████████████████████████████████████████████| 2814/2814 [01:05<00:00, 43.12it/s]
Image saved successfully!

-i was the OPTOF-NAC PET
--CT was the original CT
-o was the path to the synthetic CT to be generated
-m FDG

[claesnl]

Thank you for sharing the data. I could reproduce the error, and as I though, it has to do with the normalization.
The max of your data is <300 counts/sec, where my training data was usually between 5.000 and 10.000. My normalization uses a mask before calculating the percentiles, masking the foreground of the NAC image above 50 counts/sec - which for your data constitutes ~20% of the voxels.

I have added a temporary fix where you can overwrite the threshold. See the most recent commit. Ideally this should of course be calculated automatically based on the data, but that is for a future version.
I have pushed v0.2.4 to pip if you want to try it out.

I was able to produce an sCT using your data:
CTlessPET -i OPTOF-NAC-PT.nii --CT AC_CT_reference.nii -o AC_sCT_v024_gt1.nii -m FDG --normalization_masking_threshold 1

The produced synthetic CT now looks more like a CT, albeit with suboptimal performance. I think the low number of counts in the image (low-dose or low-acq?) is significantly lower than what the model has seen during training.

[Keyn34]

Thank you very much!
Very Interesting observation. Any clue why we never encountered this before with the older version?

We simulated very low counts from ca. 100 MBq injected activity in our previous analysis - the same as here, where the PET only contains 5% of the original injected activity.

The CT now looks exactly as I would have expected.

I will give the updated version a go next week!