Question
Question Posted 11/25/13:
Dear Experts,
With regard to "Updated UC Berkeley- AV45 Analysis Dataset is Now Available", I see that the data are not the same for those subjects represented in the May 16th data release (996 ea.). The overall bias is not bad (e.g., -0.0034 for SUMMARYSUVR_WHOLECEREBNORM), but there are some subjects that are quite different (see subject 2240 baseline: WHOLECEREBELLUM in May was 0.582, in Nov was 1.318). The test regions were also much larger, so the summary SUVR was close, but not equal (1.411 in May, 1.544 in Nov)
Is my assumption that these data were fully re-processed correct?
Dear Experts,
With regard to "Updated UC Berkeley- AV45 Analysis Dataset is Now Available", I see that the data are not the same for those subjects represented in the May 16th data release (996 ea.). The overall bias is not bad (e.g., -0.0034 for SUMMARYSUVR_WHOLECEREBNORM), but there are some subjects that are quite different (see subject 2240 baseline: WHOLECEREBELLUM in May was 0.582, in Nov was 1.318). The test regions were also much larger, so the summary SUVR was close, but not equal (1.411 in May, 1.544 in Nov)
Is my assumption that these data were fully re-processed correct?
Response posted 12/09/13 by Susan Landau:
There is a minor difference between the May and Nov AV45 datasets, which is that the May dataset used ROIs that were in PET resolution, whereas the Nov dataset used ROIs in the MRI resolution. The resampling of the ROIs shouldn't have any substantive influence on the SUVR values, which is what we found. Our comparison of the old and new pipelines (N=911) showed an R-squared value of 0.995 (slope of 0.97, intercept = 0.03). The most discrepant subject is the one you identified (RID: 2240) who had an SUVR change of one-tenth of an SUVR value, which is almost certainly due to the fact that that particular subject's cerebellum is cut off of the AV45 scan. However, we feel that the fact that the subject's SUVRs ended up being so close (within one-tenth of an SUVR) despite the missing data is a strong feature of the pipeline.
The bottom line is that the two datasets are in agreement to 2 sig digits (as you mentioned, a mean difference of 0.003), reflecting very minimal difference between the pipelines that shouldn't have any substantive impact on analyses.
The bottom line is that the two datasets are in agreement to 2 sig digits (as you mentioned, a mean difference of 0.003), reflecting very minimal difference between the pipelines that shouldn't have any substantive impact on analyses.
