Vertexwise discriminant analysis of ADNI [18F]florbetapir PET images for classification of amyloid-positive and amyloid-negative subjects

Abstract

Classification of subjects as Amyloid-Positive (Aβ+) or Amyloid-Negative (Aβ-) based on Amyloid positron emission tomography (PET) scans is increasingly utilized in research studies and clinical practice. While qualitative, visual assessment is currently the gold-standard approach, automated classification techniques possess the inherent advantages of being objective, reproducible, and efficient. The goal of this work was to develop a statistically-driven approach that will facilitate automated classification of subjects based on conventional [18F]florbetapir PET scans. [18F]florbetapir PET and 3D T1-weighted MR images were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) normal control (NC) mild cognitive impairment (MCI) subjects. The PET volumes were registered to a customized MRI template in MNI stereotaxic space, and standardized uptake value ratio (SUVR) images were generated and projected onto each subject's cortical surface using Biospective's fully-automated PIANO TM image processing software. Subjects were initially classified as Aβ+ or Aβ- based on a visual evaluation of the [18F]florbetapir PET scans. A vertexwise discriminant analysis was performed in order to generate cortical surface parametric maps of accuracy, sensitivity, and specificity. A non-parametric conjunction analysis with multiple comparisons-controlled thresholding generated a Statistical Region-of-Interest (statROI). Receiver operating characteristic (ROC) analysis, based on individual statROI measurements, was used to obtain a threshold value for the separation of Aβ+ and Aβ- subjects. The vertexwise discriminant analysis resulted in parametric maps with maximum accuracy, sensitivity, and specificity values of 0.91 0.94, 0.95, respectively. The statROI included sub-regions of the precuneus, anterior/posterior cingulate cortex, medial/lateral frontal cortex, lateral temporal lobe, and inferior parietal lobule. ROCs based on this composite statROI yielded an optimal threshold for classification of Aβ+ and Aβ- subjects. This novel, vertexwise discrimant analysis approach allowed for determination of an objective statROI and associated threshold that effectively discriminated Aβ+ or Aβ- subjects. We anticipate that the statROI defined in this study will replace existing “composite ROIs”, and allow for standardized analysis across studies. Ultimately, automated classification of subjects based on Amyloid PET scans may complement or supplant visual reads in routine clinical practice and for eligibility assessment in clinical trials.