HAI Book 2025 - Flipbook - Page 408
Bourgeat, Pierrick
101
DeepSUVR: Towards redefining the Centiloid masks using a datadriven approach
Pierrick Bourgeat1, Jurgen Fripp1, Leo Lebrat1, Ashley Gillman1, Timothy Cox1, Manu Goyal5,
Duygu Tosun-Turgut4, Pamela LaMontagne5, Tammie Benzinger3, Michael Weiner4, John
Morris3, Victor L Villemagne6, Colin Masters7, Christopher Rowe2,7, Vincent Dore1,2
1
CSIRO Health and Biosecurity, Brisbane, AU
Department of Molecular Imaging & Therapy, Austin Health, Melbourne, AU
3
Knight Alzheimer Disease Research Center, St. Louis, MO, US
4
San Francisco Veterans Affairs Medical Center, San Francisco, CA, US
5
Mallinckrodt Institute of Radiology, Washington University School of Medicine, St. Louis, MO, US
6
Department of Psychiatry, University of Pittsburgh, Pittsburgh, PA, US
7
The Florey Institute of Neuroscience and Mental Health, University of Melbourne, Melbourne, AU
2
Background: PET quantification using the Standardised Uptake Value Ratio (SUVR) is hindered by noise, spill in,
and specific binding in the reference region. We evaluated a novel deep learning method which learns from noise
in longitudinal trends to correct SUVR quantification and use it to derive new optimized Centiloid masks.
Method: 2281 participants with 2+ visits in AIBL/ADNI (7380 scans) had their Amyloid PET images spatially
normalised and quantified usinon-g the Centiloid SPM8 pipeline. A deep learning network (DeepSUVR) was trained
to predict a SUVR correction factor (CF) for each spatially normalised image. For each iteration, the prediction
was run on 2 random timepoints from the same participant, resulting in 2 Adjusted Centiloids computed using
SUVR*CF and each tracer9s standard Centiloid transform. A loss function was defined to penalise unexpected
temporal changes: Centiloid decreasing over time, Centiloid deviating from the Centiloid/Year vs mean Centiloid
curve. The model was trained using 5-fold cross-validation on ADNI+AIBL. For each tracer, new reference and
target masks were generated through optimisation to maximise the correlation with the corrected SUVRs in
ADNI+AIBL. The resulting masks were evaluated on the OASIS and GAAIN calibration datasets.
Results: The new target masks were similar to the original Centiloid masks, but the new reference masks showed
greater variability across tracers and included more white matter. Both DeepSUVR and the masks increased the
inter-tracer correlation in the GAAIN dataset while reducing the variance in the young controls. They also
increased the correlation between Centiloid/Year vs mean Centiloid in AIBL+ADNI (from ρ=0.41 to ρ=0.52) and
OASIS (ρ=0.49 to ρ=0.58).
Conclusions: We evaluated a new technique to correct SUVR estimation based on noise in longitudinal trends,
which was used to define new masks. These masks provide an insight into the inner-workings of DeepSUVR and
could inform the development of new Centiloid masks.
HAI2025 - 408
