MRI morphometric data as predictor of PET amyloid positivity: New insights using brain connectivity metrics and machine learning

Abstract

AbstractBackgroundThe presence of brain amyloid‐beta plaques is relevant for the diagnosis of Alzheimer’s Disease (AD) [1‐3]. However, positron emission tomography (PET), used to assess amyloid loading, is expensive and not widely available as other imaging modalities [2]. Therefore, the aim of this study was to assess if magnetic resonance imaging (MRI) morphometric data (volumes, areas and thicknesses) and derived brain connectivity metrics would predict PET amyloid positivity.MethodMRI and 18F‐florbetapir PET images, age, gender, Mini‐Mental State Examination (MMSE) scores, and corresponding diagnosis were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database at baseline (90 AD: 77 amyloid positive; 177 Mild Cognitive Impairment ‐ MCI: 130 amyloid positive; 122 Control: 72 amyloid positive). PET images were processed using PETSurfer to compute the relative Standardized Uptake Value (rSUV) values [5] to be used as labels to divide subjects into amyloid positive (rSUV > 1.1) and amyloid negative (rSUV < 1.1) regardless of diagnosis. Volumetric T1‐weighted MRI images were segmented and parcellated using FreeSurfer to obtain morphometric data of cortical and subcortical structures. These data were also used to derive graph theory connectivity (GTC) metrics, which translate relations between the morphometry of different brain regions. Then, all these data were used as features to obtain models for predicting amyloid positivity using various machine learning algorithms. A 10‐fold cross‐validation approach with a data split of 70% for training and 30% for testing was used.ResultThe combination of morphometric data and GTC metrics, demographic and MMSE score data was the model with the highest accuracy: 77.8%, AUC‐ROC=0.719 (CI95%=[0.633; 0.804]), NPV=76.9%, PPV=77.9% using a support vector machine algorithm. Using the morphometric data and GTC metrics only the model achieved an accuracy of 77.3%, AUC‐ROC=0.713 (CI95%=[0.627; 0.799]), NPV=66.7%, PPV=79.6%, whilst using morphometric data alone achieved an accuracy of 75.6%, AUC‐ROC=0.671 (CI95%=[0.583; 0.760]), NPV=64.7%, PPV=77.4%. .ConclusionWhile the relationship between brain atrophy and amyloid deposition has been reported previously [6], the results described here suggest that GTC increases the sensitivity in depicting the atrophy‐amyloid correlation. Additionally, results also suggest that volumetric T1‐weighted may provide additional information for predicting amyloid positivity and loading.