Retrospective quantitative harmonization in PET using deconvolution and optimal filtering

The reliability of longitudinal quantitative PET image analysis suffers if scans are acquired on different PET scanners. Here, we describe a post-reconstruction harmonization method that can be implemented to enable quantitative PET analysis across scanners.
First, images are unfiltered by Wiener deconvolution. Then, the differences in contrast recovery coefficients (CRC) from NEMA phantom scans are minimized by finding isotropic 3D Gaussian filters for each scanner/reconstruction setting combination with a downhill-simplex optimizer. After optimal filters are established, PET images to be harmonized are unfiltered using Wiener deconvolution, and then re-filtered with the determined optimal settings.
We demonstrate that the method minimizes differences in CRC values for a set of 7 heterogeneous reconstruction protocols acquired on 3 different PET/CT scanners. We apply our method retrospectively to 18F-FDG PET/CT scans of a cohort of metastatic melanoma patients to show that performing harmonization has a significant impact in assigning lesion response as measured by change in PET SUV metrics.
In conclusion, the described method for harmonization enables accurate quantitative PET image analysis for retrospective datasets, and has clinical impact in response assessment settings.