ADPrep ‐ A fully automated Alzheimer’s Disease Neuroimaging Preprocessing Pipeline for MRI and multi‐tracer PET data

Abstract

AbstractBackgroundState‐of‐the‐art preprocessing of MRI and PET data is crucial for Alzheimer’s disease (AD) neuroimaging research. Therefore, standardization and harmonization of neuroimaging preprocessing across sites is key to generate comparable and sharable datasets and to reduce potential bias introduced by different preprocessing strategies. Further, neuroimaging preprocessing requires strong expertise in different programming languages (R/MATLAB/Python/Bash) and software packages (SPM/FSL/AFNI/FreeSurfer), i.e., a major barrier for non‐expert users. Therefore, we developed a docker‐based, state‐of‐the‐art, fully automated, neuroimaging toolbox that integrates robust preprocessing of structural/functional MRI, and multi‐tracer PET (amyloid/tau/FDG/TSPO), generating easy to use spreadsheet and nifti outputs. The toolbox requires no programming expertise and can facilitate neuroimaging analyses, data harmonization and sharing across the AD neuroimaging community.MethodThe pipeline works on standardized bids‐formatted data and was fully developed in nipype (see Figure 1). Preprocessing for structural MRI includes volumetric and cortical thickness assessments for widely used brain atlases (Desikan‐Killiany/Schaefer/LPBA/Hammers/Neuromorphometrics/Cobra/Destrieux), plus spatially normalized and smoothed grey/white‐matter segments for voxel‐based morphometry and generation of subject‐specific PET‐reference regions. Functional MRI processing includes slice‐timing and motion correction, nuisance regression, spatial normalization, and functional connectivity assessments for the above‐mentioned atlases. PET processing includes generation of spatially normalized SUVR images for different tracer‐specific reference regions, as well as extraction of atlas‐based SUVRs (PVE and non‐PVE‐corrected). Standardized imaging outputs in FreeSurfer space were benchmarked against available ADNI data to illustrate comparability with existing pipelines (see Figure 2). Cluster implementation is provided to ensure large‐scale data processing.ResultThe toolbox was tested successfully on large‐scale multimodal datasets, including several thousand scans from ADNI, ADNI‐DOD, and A4 with an overall processing failure rate of <4%. Using ADNI data as a benchmark, our data closely reproduce global amyloid‐PET (r = 0.99, p<0.001, Fig.2A) and temporal meta tau‐PET SUVRs (r = 0.98, p<0.001, Fig.2B), suggesting that PET‐thresholds can be applied. Runtime is ∼1h for a structural MRI and ∼5hrs for a PET including PVE‐correction.ConclusionWe propose a user‐friendly multimodal robust neuroimaging preprocessing toolbox, that can be applied to different neuroimaging datasets by non‐expert users. The toolbox provides outputs that can be directly used for statistical analyses. Future work involves implementing other imaging modalities, e.g., diffusion tensor imaging.