HAI Book 2025 - Flipbook - Page 243
Fajardo-Valdez, Alfonso
56
The impact of spatial smoothing on the accuracy of Centiloid
thresholds
Alfonso Fajardo-Valdez1,2, Ting Qiu1,2, Jonathan Gallego-Rudolf1,2, Mohammadali Javanray1,2,
Sylvia Villeneuve1,3,4
2
Douglas Mental Health University Institute, Montreal, Quebec, Canada, Montreal, QC, CA
Integrated Program in Neuroscience, Faculty of medicine, McGill University, Montreal, Canada, Montreal, QC, CA
3
Montreal Neurological Institute, McGill University, Montreal, Canada, Montreal, QC, CA
4
Department of Psychiatry, Faculty of medicine, McGill University, Montreal, Canada, Montreal, QC, CA
2
Introduction: The Centiloid (CL) scale was developed to harmonize amyloid-beta (A´) values across tracers,
preprocessing methods and scanners. CLs are useful to derive comparable cut-off thresholds to assess A´
positivity across sites. For CLs to be reliable, the linear model output evaluating ground truth vs obtained CLs
needs to respect some criteria. Here, we examined differences due to spatial smoothing (SS), a common
procedure in PET preprocessing, on SUVR-to-CL conversion, CL-based thresholds, and group-level SUVR
statistics.
Methods: We obtained CLs and 50-70 min 18F-NAV4694 PET scans from 55 subjects (10 young, 45 AD) from the
GAAIN website. Preprocessing was conducted via the VilleneuveLab pipeline (Vlpp; see Qiu et al., 2024) using 4
distinct Vlpp SS Spatial Smoothing (FWHM) kernels: 0mm, 4mm, 8mm and 12mm. We extracted A´ SUVRs from
the GAAIN cortical mask (whole-cerebellum as reference region) and we calculated CLs using the standard GAAIN
18
F-NAV4694 CL equation:
CL NAV4694 = 100 x (SUVRNAV4694 3 1.031) / 1.172
Throughout linear models, we evaluated the accuracy of CLs. We then classified older individuals as A´- or A´+
based on CL12 or CL20 thresholds. Finally, we calculated the effect size (Cohen9s D) when comparing SUVR means
of Desikan-Killiany (DK) regions or global amyloid index (average of 40 DK regions) between groups. Analyses were
reproduced in a sample of n=45 older adults from the PREVENT-AD cohort (fig 2).
Results: No SS resulted in higher accuracy of CLs (fig 1a / table 1). At C12 threshold, greater SS increased the
number A´+ individuals, whereas at CL20, classifications outcomes were highly similar. Despite these, a
significant effect of SS was found when comparing A´+ vs A´- SUVRs in both datasets.
Conclusions:
•
•
•
More reliable CLs were obtained with non-smoothed data.
SUVRs are influenced by SS.
Evaluation with other tracers and pipelines is needed.
Figures:
HAI2025 - 243
