HAI Book 2025 - Flipbook - Page 200
Bourgeat, Pierrick
41
DeepSUVR: Using temporal constraints to improve tau quantification
Pierrick Bourgeat1, Jurgen Fripp1, Ashley Gillman1, Azadeh Feizpour2, Christopher Schwarz4,
Clifford R. Jack4, Val Lowe4, Michael W. Weiner3, Colin L. Masters5, Victor L. Villemagne6,
Christopher C. Rowe2,5, Vincent Dore1,2
1
CSIRO Health and Biosecurity, Brisbane, AU
Department of Molecular Imaging & Therapy, Austin Health, Melbourbe, AU
4
San Francisco Veterans Affairs Medical Center, San Francisco, CA, US
5
Department of Radiology, Mayo Clinic and Foundation, Rochester, MN, US
5
The Florey Institute of Neuroscience and Mental Health, University of Melbourne, Parkville, AU
6
Department of Psychiatry, University of Pittsburgh, Pittsburgh, PA, US
3
Background: Tau PET quantification using SUVR is affected by noise, spill in, and non-specific binding in the
reference and target regions. We evaluated a novel deep learning method which learns from noise in longitudinal
trajectories to correct the SUVR and improve quantification.
Method: 730 participants imaged at two or more visits using Flortaucipir (FTP=474) or MK6240 (MK=597) in
AIBL+ADNI (N=1771 scans) had their tau PET images spatially normalised and quantified using the CenTauR (CTR)
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 visits from the same
participant, resulting in 2 Adjusted CTR computed using SUVR*CF and each tracer9s CTR transform. A loss
function was defined to penalise unexpected temporal changes like CTR decreasing over time, CTR deviating
from the CTR/Year vs mean CTR curve. The model was trained using 5-fold cross-validation on AIBL+ADNI. Once
trained, the inference is performed on each visit independently. The model was evaluated on the out-of-fold testsets, using the percentage of outliers (defined as changes
15CTR/Yr), and comparing the rate of
change in the CU-/CU+/MCI+/AD+ using a matched subset of MK and FTP. The model was further evaluated on an
independent longitudinal test set from the Mayo Clinic comparing the rate of change in the CU/AD.
Results: In AIBL+ADNI, DeepSUVR reduced outliers by 44% on average in the MTL, Temporo-Parietal and Frontal
regions (Fig1). It also increased the separation in CTR/Yr between clinical groups in both AIBL+ADNI (Fig2) and in
the Mayo dataset (Fig3).
Conclusions: We have developed a novel technique to improve CTR quantification. While a larger dataset is
essential to better optimise the model, the preliminary results are promising and show potential to reduce
sample-sizes required in anti-tau trials.
HAI2025 - 200
