HAI Book 2025 - Flipbook - Page 463
Schwarz, Adam
Robustness of the centiloid scale across research and commercial software
using [18F]flutemetamol PET images
Adam Schwarz3, Ariane Bollack1,2, Chris Buckley1, Mark Battle1, Oskar Hansson4, Pierrick
Bourgeat5, Vincent Dore5,6, Jurgen Fripp5, Lennart Thurfjell7, Will Balhorn8, Mike Haas8, Chris
Page1, Gill Farrar1
1
GE HealthCare, Chalfont St. Giles, GB
UCL, London, GB
3
GE HealthCare, Holliston, MA, US
4
Lund University, Lund, SE
5
CSIRO, Brisbane, AU
6
Austin Health, Melbourne, AU
7
Combinostics, Tampere, FI
8
MIM Software Inc., Cleveland, OH, US
2
Background: Amyloid PET quantification using the Centiloid (CL) metric is becoming more prevalent and may
have a role in clinical decision-making. To evaluate the robustness and interchangeability of CL values obtained
from different software, we systematically assessed repeatability, reproducibility, and reliability across six CL
quantification pipelines applied to 210 [18F]flutemetamol scans.
Methods: Three datasets were used: (1) An AD test-retest cohort (N=10x2); (2) an aMCI cohort (N=80); (3) cases
from the BioFINDER-1 cohort enriched for amyloid loads around common positivity thresholds (0-50CL) (N=110).
Two regulatory-approved software (cPET, MIMneuro) and four research pipelines (CapAIBL, rPOP, Amypype,
SPM8) were evaluated. Dataset 1 was used to assess repeatability. Dataset 2 was used to assess reproducibility in
terms of absolute agreement in continuous CL values. Datasets 2 and 3 were used to assess reliability for
dichotomizing amyloid scans into positive and negative. The latter was performed for three thresholds of 11, 25
and 37 CL.
Results: The test-retest analysis revealed an absolute bias of
