Accuracy and Longitudinal Consistency of PET/MR Attenuation Correction in Amyloid PET Imaging amid Software and Hardware Upgrades

BACKGROUND AND PURPOSE: Integrated PET/MR allows the simultaneous acquisition of PET biomarkers and structural and functional MRI to study Alzheimer disease (AD). Attenuation correction (AC), crucial for PET quantification, can be performed by using a deep learning approach, DL-Dixon, based on standard Dixon images. Longitudinal amyloid PET imaging, which provides important information about disease progression or treatment responses in AD, is usually acquired over several years. Hardware and software upgrades often occur during a multiple-year study period, resulting in data variability. This study aims to harmonize PET/MR DL-Dixon AC amid software and head coil updates and evaluate its accuracy and longitudinal consistency. MATERIALS AND METHODS: Tri-modality PET/MR and CT images were obtained from 329 participants, with a subset of 38 undergoing tri-modality scans twice within approximately 3 years. Transfer learning was used to fine-tune DL-Dixon models on images from 2 scanner software versions (VB20P and VE11P) and 2 head coils (16-channel and 32-channel coils). The accuracy and longitudinal consistency of the DL-Dixon AC were evaluated. Power analyses were performed to estimate the sample size needed to detect various levels of longitudinal changes in the PET standardized uptake value ratio (SUVR). RESULTS: The DL-Dixon method demonstrated high accuracy across all data, irrespective of scanner software versions and head coils. More than 95.6% of brain voxels showed less than 10% PET relative absolute error in all participants. The median [interquartile range] PET mean relative absolute error was 1.10% [0.93%, 1.26%], 1.24% [1.03%, 1.54%], 0.99% [0.86%, 1.13%] in the cortical summary region, and 1.04% [0.83%, 1.36%], 1.08% [0.84%, 1.34%], 1.05% [0.72%, 1.32%] in cerebellum by using the DL-Dixon models for the VB20P 16-channel coil, VE11P 16-channel coil, and VE11P 32-channel coil data, respectively. The within-subject coefficient of variation and intraclass correlation coefficient of PET SUVR in the cortical regions were comparable between the DL-Dixon and CT AC. Power analysis indicated that similar numbers of participants would be needed to detect the same level of PET changes by using DL-Dixon and CT AC. CONCLUSIONS: DL-Dixon exhibited excellent accuracy and longitudinal consistency across the 2 software versions and head coils, demonstrating its robustness for longitudinal PET/MR neuroimaging studies in AD.