Premium
Amyloid Pattern Similarity Score (AMPSS): A reference region free measure of amyloid PET deposition in Alzheimer’s disease
Author(s) -
Prosser Lloyd,
Veale Thomas,
Malone Ian B,
Coath William,
Fox Nick C,
Cash David M
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.042673
Subject(s) - support vector machine , amyloid (mycology) , artificial intelligence , voxel , pattern recognition (psychology) , concordance , neuroimaging , alzheimer's disease , grey matter , alzheimer's disease neuroimaging initiative , medicine , pathology , nuclear medicine , psychology , white matter , magnetic resonance imaging , neuroscience , disease , computer science , radiology
Background Amyloid‐PET has high sensitivity for cerebral amyloid, a hallmark of Alzheimer’s disease (AD). Typically, a global measure of amyloid burden in used to classify individuals as amyloid positive or negative. Measures like the Standarised Uptake Value ratio (SUVr) rely on a reference region to normalise amyloid measures, but are non‐specific, sensitive to delineation errors, and reference region choice is debated. This study evaluates a summary measure based on the multivariate pattern of amyloid deposition without a reference region, named Amyloid Pattern Similarity Score (AMPSS). Method Alzheimer’s Disease Neuroimaging initiative (ADNI) individual and volumetric T1 MRI data were co‐registered. MRI data was then spatially normalised using DARTEL. Combined, these two transforms then map PET data into common space. A support vector machine (SVM) was trained using a) age‐matched normal controls with higher CSF Aβ 42 , and b) individuals with clinical diagnosis of AD and lower CSF Aβ 42 (Table 1). Normalised grey matter voxels were used as the SVM features. Rather than conventional binary classification, the SVM generated probabilistic scores based on logistic regression (i.e. AMPSS). Three evaluations were performed: (1) a validation set (Table 1), (2) comparison with CSF determined amyloid status, and (3) initial longitudinal trajectory evaluation (Table 2). Result The validation set AMPSS perfectly classified AD and controls. AMPSS results did not differ depending on specific reference region or global signal normalisation (between subjects ANOVA F(2,26) = 0.65, p = 0.526, r 2 .99). AMPSS showed high concordance with CSF defined amyloid status. Figure 1 highlights the strong AMPSS and CSF relationship. There was a nominal, non‐significant improvement in accuracy for this method compared to SUVr (p=.176), Figure 2 . Longitudinal data highlighted that individuals with baseline AMPSS >50% tended to show increased AMPSS at follow‐up, while most individuals <50% baseline AMPSS remained stable. However, there is evidence of some individuals with subthreshold amyloid accumulation. (Table 3, Figure 3 .). Conclusion The Amyloid Pattern Similarity Score is a reference‐free summary metric of amyloid deposition that performs comparably with conventional SUVr measures. Initial validation shows high agreement with CSF and good sensitivity to increasing amyloid accumulation over time, even in early disease.