For each subject, all scans were collected from ADNI's image and data archive using a specific advanced search (“AV45 Coreg, Avg, Std Img and Vox Siz, Uniform Resolution”). The scans from this search were coregistered PET-MR and intensity normalized images that used Statistical Parametric Mapping (SPM8), a medical imaging process which allows SUV comparisons within select regions to be made in a given subject (Smith et al., 2022 (link)). Coregistering is important because MR has fine anatomical detail and PET cannot delineate anatomic structures (Robertson et al., 2016 (link)). PET transfers radiotracer information to MR throughout the coregistering process. Over the 20-min acquisition time, each image was resized to a uniform voxel size and each uniform size was 160 × 160 in-plane, along with 96 axial slices (Reith et al., 2020 (link); Landau et al., 2021 ). All images were normalized and rescaled to 224 × 224 to accommodate the ImageNet pretraining.
We obtained the 18F-florbetapir cortical summary SUVR (“SUMMARYSUVR_WHOLECEREBNORM”) for each scan from the UC Berkeley AV45 Analysis. This calculation required FreeSurfer processing which included skull-stripping, segmentation, and delineation of cortical and subcortical regions in MRI scans which were co-registered to PET scans using SPM8. The cortical summary region (“COMPOSITE_SUVR”) was calculated by taking the mean uptake of all SUVR values from the subregions. These SUVR (“COMPOSITE_SUVR”) values were calculated with respect to the reference region (“WHOLECEREBELLUM_SUVR”) to derive the summary SUVR value for the whole cerebellum (“SUMMARYSUVR_WHOLECEREBNORM”) for each scan (Landau et al., 2021 ).
SUV(t) represents the radioactivity concentration in the subcortical and cortical regions (ROI) averaged during a period of time over the quantity of the injected dose (kBq/mL) divided by the weight (kg). This value is then calculated with respect to the reference region which determines SUVR.