Brain-PET image reconstruction methods affect software-aided diagnosis in patients with neurodegenerative diseases

Objectives: The introduction of new developments in neuro-PET, such as digital silicon photomultiplier detectors, time-of-flight (TOF), point-spread-function (PSF) modelling and penalized reconstruction methods has resulted in improved spatial resolution and signal to noise ratio in reconstructed images. The objective of this study was to investigate how the improved PET image reconstruction affects quantitative outcome measures and software-aided assessing of patient pathologies. Methods: Twenty-four subjects with 18F-FDG PET examinations, 15 neurodegenerative disease subjects and 9 melanoma subjects without brain involvement (normal controls), scanned on a digital TOF PET/CT scanner, were included in the study. All subjects received 3 MBq/kg, patients were scanned for 10 min starting 45 min post injection, whereas controls were scanned circa 70 min post injection during 2 min. Two reconstruction methods were used; ordered subsets expectation maximization (OSEM; 3 iterations, 16/34 subsets, 3/5 mm Gaussian postfilter, +/- TOF, +/- PSF modelling) and block-sequential regularized expectation maximization (BSREM) (TOF, PSF, penalty regularization parameter β of 75, 150, 225, and 300). A 256 by 256 matrix over a 25 cm FOV was used. Each reconstructed image was anatomically normalized into template space using Cortex ID Suite software (GE Healthcare). Automated analysis of tracer uptake in 26 regions of interest and comparison with the corresponding tracer uptake in normal subjects in terms of z-scores was performed using the whole brain as reference area. Results: TOF, PSF modelling and BSREM gradually increased the relative uptake difference to the normal subjects' database within the software, i.e. resulting in decreasing z-scores. The controls of the study yielded similar results to the normal database when using OSEM 3/16 5 mm reconstruction, without TOF and PSF, with a mean z-score of -0.3. BSREM with β 150, OSEM-TOF-PSF 3/34 3 mm and 5 mm resulted in average z-scores of -1.2, -0.99, and -0.62, respectively. Reducing the filter from 5 to 3 mm decreased z-scores between 19% and 33%, while increasing the number of subsets from 16 to 34 decreased z-scores by 6-29%. Use of TOF with OSEM resulted in slightly decreased z-scores of 1-12%. PSF modelling also decreased z-scores by 9% to 28%. BSREM reconstruction resulted in reduced z-scores compared to OSEM reconstructions regardless of β, the z-score difference to OSEM 3/16 5 mm was -0.96, -0.71, -0.59, and -0.4