Decomposition-based harmonization for quantitative PET imaging across scanners and radiotracers

Area of Science:

• Medical Imaging
• Neuroscience
• Oncology

Background:

• Quantitative positron emission tomography (PET) is crucial in oncology and neuroscience.
• Scanner, protocol, and radiotracer variability compromise PET quantitative accuracy.
• This variability limits the reliability of multicenter PET studies.

Purpose of the Study:

• To develop and validate PETHarmony, a novel voxel-level harmonization framework.
• To minimize inter-scanner and inter-tracer variability in quantitative PET data.

Main Methods:

• PETHarmony uses a linear neural network and singular value decomposition for voxel-level harmonization.
• Performance was assessed using paired PET/CT and PET/MR scans, phantom validation, cross-tracer comparisons, and multicenter Alzheimer's disease (AD) data.
• Harmonization procedures involved leave-one-out cross-validation and training on unpaired data.

Main Results:

• PETHarmony effectively eliminated discrepancies between PET/CT and PET/MR images.
• Phantom validation demonstrated improved quantitative accuracy.
• Cross-tracer consistency was markedly improved, with regression lines approaching the line of identity.
• Multicenter AD classification accuracy improved by up to 21.7%.

Conclusions:

• PETHarmony achieves robust voxel-level harmonization of multicenter PET data.
• It significantly improves cross-scanner and cross-tracer consistency.
• PETHarmony enhances diagnostic accuracy and offers a practical solution for standardizing quantitative PET in clinical trials.