Improved comparability between measurements of mean cortical amyloid plaque burden derived from different PET tracers using multiple regions‐of‐interest and machine learning

• Published in: Alzheimer's & dementia Vol. 17; no. S1
• Main Authors: Chen, Kewei, Ghisays, Valentina, Luo, Ji, Chen, Yinghua, Lee, Wendy, Benzinger, Tammie L.S., Wu, Teresa, Reiman, Eric M., Su, Yi
• Format: Journal Article
• Language: English
• Published: 01.12.2021
• ISSN: 1552-5260; 1552-5279
• DOI: 10.1002/alz.051419

Background The conversion of standard uptake value ratios (SUVRs) to Centiloid units using previously described methods has partially but not completely improved the ability to compare measurements of mean cortical amyloid plaque burden, classify positive versus negative amyloid PET scans, assess changes over time, and evaluate amyloid‐modifying treatment effects using different radiotracers. Here, we demonstrate improved harmonization between PiB and florbetapir (FBP) PET measurements of mean cortical SUVRs (mcSUVRs) using multiple regions‐of‐interest (ROIs) and machine learning algorithms. Methods PiB and FBP PET image pairs from 91 subjects in the Open Access Series of Imaging Studies (https://www.oasis‐brains.org/) were used as the training set to optimize the algorithms to generate each subject’s pseudo‐PiB mcSUVR measurements from his or her FBP multi‐ROI SUVRs . PiB and FBP image pairs from 46 subjects in the Avid‐related Centiloid Project (www.gaain.org/centiloid‐project) were then used as the test set to compare Pearson’s correlation coefficients (r) between the pseudo and actual PiB mcSUVR measurements using our approach to that used in the Centiloid method. SUVRs with cerebellar reference were extracted from multi‐ROIs using FreeSurfer with partial volume correction. Pseudo mcSUVRs were generated using ensemble regression, partial least square regression (PLSR) and artificial neural network (ANN) separately, and their correlations with the actual PiB mcSUVR were compared to the PiB‐FBP correlation on Centiloid in both training and test sets. Results In the training set, the Centiloid based PiB‐FBP correlation was r=0.904. An ANN with 7/6 neurons in the 1st/2nd hidden layers improved it to 0.987 (p≤8.3e‐12) and PLSR to 0.973 (p≤3.8e‐5). In the independent test set, ANN improved r to 0.981 from 0.927 (p≤6.6e‐4) and PLSR to 0.964 (p=0.011). Ensemble regression did not improve r. Conclusion The ANN and PLSR algorithms mapping multi‐ROI FBP SUVRs to PiB mcSUVR appear to increase the comparability between measurements of mean cortical amyloid plaque burden using different PET tracers. Additional studies are needed to demonstrate the generalizability to other amyloid PET tracers, clarify its comparability in distinguishing between positive and negative amyloid PET scans, demonstrate its value in longitudinal studies and clinical trials, and extend our approaches to different tau PET ligands.