HAI Book 2025 - Flipbook - Page 302
Jones, David
59
Interpretable prediction of amyloid PET positivity using non-negative
matrix factorization
Gemeng Zhang1, Leland R. Barnard1, Venkatsampath Gogineni1, Nick Corriveau-Lecavalier1,2,
Hugo Botha1, Brian J. Burkett3, Derek R. Johnson3, Val J. Lowe3, Stuart J. McCarter1, Vijay K.
Ramanan1, Jonathan Graff-Radford3, David S. Knopman1, Ronald C. Petersen1, David T. Jones1
1
Department of Neurology, Mayo Clinic, ROCHESTER, MN, US
Department of Psychiatry and Psychology, ROCHESTER, MN, US
3
Department of Radiology, ROCHESTER, MN, US
2
Background: ยด-amyloid accumulation is a key biomarker in Alzheimer's disease (AD). Standardized uptake value
ratio (SUVR) images and the derived Centiloid (CL) values are commonly used to evaluate amyloid PET
positivity(A+) alongside visual reading. We propose an interpretable non-negative matrix factorization (NMF)
method to assist the visual reading process by decomposing the amyloid PET scans into representative
components, facilitating accurate A+ detection while ensuring interpretability for clinical validation.
Method: PET images from the 18F-Florbetapir tracer were analyzed. A total of 2300 amyloid-PET images from
Imaging Dementia 3 Evidence for Amyloid Scanning (IDEAS) study was used for NMF model training ( and
classification models training (200 A+, 200 A-). The remaining 1700 images (1121 A+, 579 A-) were used as a test set.
Two external datasets, the Mayo Clinic Alzheimer9s Disease Therapy Clinic (ADTC) (84 A+, 11 A-) and the ADNI4
cohort data (31 A+, 35 A-) are used for validation. Amyloid-PET images were registered to an amyloid-PET
template, and whole brain SUVR images were generated using the cerebellum as the reference region. The NMF
model was fitted using SUVR images, and logistic regression (LR) model was used for A+/A- prediction. LR
weights were further used to interpret the relationship between NMF features and A+/A- classification.
Results: The LR model achieved an AUC of 0.94 (95% CI: 0.93-0.95) on IDEAS test set, 0.98 (95% CI: 0.94-1.00) on
Mayo ADTC and 0.97 (95% CI: 0.94-1.00) on ADNI4 (See Fig.1). NMF components C2 (whole-brain SUVR average)
and C8 (gray-over-white contrast) were the top influencers for A+ prediction, while C7 (white-over-gray contrast)
was most influential for A- prediction (See Fig.2).
Conclusion: The strong prediction performance across two external datasets demonstrates that NMF-based
SUVR patterns generalize well and provide interpretable, accurate A+ predictions. Visualizations of key
components align with established radiological reading protocols.
HAI2025 - 302
