Summary of warnings:
Two issues exist with Siemens HRRT PET images.

1. After analysis of the ADNI normal control PET subjects, it has become apparent that the baseline pattern of FDG metabolism is different for subjects scanned on HRRT PET scanners than on other scanners. This is not a resolution effect, since the differences are apparent on images smoothed to a common resolution (8 mm). In general, regions in the lower transaxial slices and in particular regions near the middle of these slices, show relatively lower metabolic activity in HRRT scans than in scans acquired on other scanners. For example, the value of the pons relative to the global mean is 12% lower for HRRT scans on average than for all other scans. While the cause of this is not certain, my best guess is that it is a related to scatter correction (or possibly attenuation correction or a combination of the two). The differences are consistent with the lower transaxial slices (base of the brain down into the neck) having greater scatter subtraction for HRRT scanners than other scanners. One possible reason would be that the axial FOV of HRRT scanners is much larger than most all other scanner models, and hence the correction can account for scatter which comes from the levels of the body (e.g. below the brain) imaged by the HRRT that are out of the axial field-of-view of other scanners. Thus, scatter that can be attributed to these “below-brain” slices imaged with the HRRT is being successfully removed by the scatter correction routine, resulting in relatively lower PET values in these inferior slices. Scatter from these areas of the body would not be accounted for and hence subtracted with other scanners since they are out of the axial FOV and not imaged.

This apparent difference between scanner types was not seen in the phantom studies, where the relative values in “pons” or “cerebellar” regions of the Hoffman phantom did not differ significantly between HRRT and other scanners. This result is also consistent with the idea of the difference being due to scatter correction. The Hoffman phantom scans did not have activity in these lower slices and hence there was not additional scatter in the field-of-view that could be removed. There was a torso phantom used with the Hoffman phantom scans (simulating out-of-field scatter), but this was outside the FOV for all scanners including the HRRT.

2. Some HRRT images have reconstruction (attenuation and/or scatter correction induced) artifacts. Some of these are quite subtle that were only noticed or confirmed when the 6 month or 12 month PET scans were obtained and overlaid with the baseline scans. This has resulted in some of the HRRT scans that had originally passed “QC” (and hence available) now being “failed”, and no longer available for download. LONI’s advanced database search can be used to find the current HRRT scans that have passed QC. In some cases these artifacts have caused even greater differences in the relative metabolic values in lower image slices, compounding the issue raised above, where for example pons values are relatively lower than in other scanners.

Current Advice (May 8, 2007):
1. At present, scans from HRRT scanners should be considered to produce a different pattern of FDG uptake than other scanners (see below). Analyses that assume comparability across all scanners should consider excluding HRRT scans for the time being (e.g. comparison of individual scans to a normal database).

2. Check to see if any HRRT PET scans that you have downloaded are no longer available. Do not use any HRRT images that are now failed and not available through LONI.

Plan:
1. Now that all the baseline PET scans on normal controls have been acquired and uploaded to LONI, we are working on an empirical correction for scanner-specific differences in PET scans such as those described above. Notices will be provided here concerning progress on this front.

2. The HRRT sites have figured out how to deal with most of the artifacts mentioned above. All failed scans from HRRT sites are being reprocessed in an attempt to reduce or remove these artifacts. Re-reconstructed images will be uploaded and QC’d. Notices will be provided here for all re-worked HRRT images that pass QC. After re-reconstruction and re-processing of these “failed”, other HRRT scans that are currently available will be re-reconstructed with the new processing stream, to make sure all subjects have optimal images over time.

– Bob Koeppe – 08 May 2007

Summary of warning:

There exists a software bug in OSEM reconstruction for Siemens BioGraph HiRez PET/CT scanners. Depending upon the circumstances, this artifact can be quite minor or can be quite substantial. Siemens is aware of this, has developed a patch for the problem, and is in the process of distributing the patch to the BioGraph HiRez sites. We will be asking all HiRez sites to re-reconstruct all of their PET images once the site has the software patch installed and tested.

The problem involves an error in the Siemens image reconstruction algorithm that may produce an artifactual asymmetry in the PET images. The actual asymmetry depends on how the patient is positioned in the scanner. These asymmetries may produce data analytic results that appear as if brain activity is higher in one hemisphere than another, potentially seriously affecting the ADNI results. The actual error for a particular scan depends upon a variety of physical factors and the specific software version the site used for that scan reconstruction.

Technical details of the problem:The problem is in the Siemens OSEM reconstruction, where the ordered subsets of the PET and CT data sets used by the EM algorithm are not correctly matched between emission (PET) and transmission (CT) data. In other words, the emission projections are being attenuation corrected with the wrong CT projections. The artifact appears as an asymmetry in the image that does not really exist. If the subject is low in the field-of-view, an image left > image right asymmetry occurs (mostly in posterior regions). If the subject is high in the FOV, the opposite occurs, but more in the frontal regions. The magnitude of the artifacts depends upon many factors, including physical factors and the particular software version and software parameters being used. These are summarized as follows. Physical factors: If the object being imaged is cylindrical and perfectly centered in the field-of-view, there is no artifact. As the object becomes more elliptical (such as a head), and as the object is located further from the center of the scanner’s FOV, the magnitude of the asymmetry increases. This error was not seen in the phantom studies that were done at the start of ADNI since the Hoffman phantom is cylindrical, and each site scanned the phantom very near to the center of the FOV.

Software version/parameter factors: The error is both software version and OSEM parameter dependent. The error is less noticeable for software versions 3.x and more noticeable in the newer software releases, 4.x. (see below for explanation). The ADNI protocol specifies 4 iterations and 16 subsets with reconstruction into a 168×168 grid using “TRIM=2”. In OSEM reconstruction, the number of projections has to be divisible by the number of subsets so that the number of projections per subset is constant. The number of angular projections for the BioGraph HiRez is 336. This is divisible by 16 (336/16=21 projections per subset). HOWEVER, a feature of Siemens software used to speed up reconstruction is to interpolate the number of rays per projection and the number of projections to the output image grid size. Thus, if the 336 projections are interpolated to 168 projections prior to reconstruction, 16 subsets cannot be used (168/16=10.5, and you can’t have half a projection). The older software either forced the user to select 14 subsets (168/14=12), or actually used 14 subsets even when 16 subsets was selected. When the version 4.x software was released, the option for 14 subsets disappeared, and when the user selects 16 subsets the software rounds this UP, not DOWN to the nearest number of acceptable subsets, which happens to be 21 (168/21=8). Thus, even though the software says 16 subsets, 21 are being used with version 4.x software. This has two effects. The minor one is that the images are noisier than before since 50% more total sub-iterations are being used (4 x 21, instead of 4×14), but more importantly, the error due to the mismatch of PET and CT projections for attenuation correction has become considerably more pronounced with 21 subsets than 14 subjects, due to the particulars of the actual software bug in Siemens code.

Current Advice (May 8, 2007):

All BioGraph HiRez PET scans need to be interpreted with caution. While the majority of artifact-induced asymmetries are probably only a few percent or less, in some cases these will be greater. If during longitudinal analysis, one finds a change in the degree of asymmetry between scans, this finding in particular should be questioned. Final conclusions should wait until the re-reconstructions are uploaded to LONI. One can make an educated guess about the magnitude of the artifact for a given scan by considering the above information (i.e. version 3 vs. version 4, location within the FOV), but also by looking at the raw images to see whether the asymmetries are seen in the scalp. While scalp uptake of FDG may also be asymmetric, the artifact will be more apparent in regions farther from the center of the object (such as scalp). Hence, in many of the ADNI BioGraph HiRez images, one can see a noticeable scalp asymmetry. In my review of all BioGraph HiRez scans, most subjects are scanned a little below the center of the field-of-view, and hence in many images one scan see an image left (brain right) posterior increase in scalp activity.

Plan:

When the software patch is available and re-reconstructions are performed by the sites, and uploaded to LONI, we will provide the updates here concerning the availability of these re-processed BioGraph HiRez scans.

– Bob Koeppe – 08 May 2007