Impact of deep learning denoising on kinetic modelling for low-dose dynamic PET: application to single- and dual-tracer imaging protocols
Florence M. Muller, Elizabeth Li, Margaret E. Daube-Witherspoon, Austin R. Pantel, Corinde E. Wiers, Jacob G. Dubroff, Christian Vanhove, Stefaan Vandenberghe, Joel S. Karp
Abstract
Abstract Purpose Long-axial field-of-view PET scanners capture multi-organ tracer distribution with high sensitivity, enabling lower dose dynamic protocols and dual-tracer imaging for comprehensive disease characterization. However, reducing dose may compromise data quality and time-activity curve (TAC) fitting, leading to higher bias in kinetic parameters. Parametric imaging poses further challenges due to noise amplification in voxel-based modelling. We explore the potential of deep learning denoising (DL-DN) to improve quantification for low-dose dynamic PET. Methods Using 16 [ 18 F]FDG PET studies from the PennPET Explorer, we trained a DL framework on 10-min images from late-phase uptake (static data) that were sub-sampled from 1/2 to 1/300 of the counts. This model was used to denoise early-to-late dynamic frame images. Its impact on quantification was evaluated using compartmental modelling and voxel-based graphical analysis for parametric imaging for single- and dual-tracer dynamic studies with [ 18 F]FDG and [ 18 F]FGln at original (injected) and reduced (sub-sampled) doses. Quantification differences were evaluated for the area under the curve of TACs, K i for [ 18 F]FDG and V T for [ 18 F]FGln, and parametric images. Results DL-DN consistently improved image quality across all dynamic frames, systematically enhancing TAC consistency and reducing tissue-dependent bias and variability in K i and V T down to 40 MBq doses. DL-DN preserved tumor heterogeneity in Logan V T images and delineation of high-flux regions in Patlak K i maps. In a /[ 18 F]FDG dual-tracer study, bias trends aligned with single-tracer results but showed reduced accuracy for [¹⁸F]FGln in breast lesions at very low doses (4 MBq). Conclusion This study demonstrates that applying DL-DN trained on static [ 18 F]FDG PET images to dynamic [ 18 F]FDG and [ 18 F]FGln PET can permit significantly reduced doses, preserving accurate FDG K i and FGln V T measurements, and enhancing parametric image quality. DL-DN shows promise for improving dynamic PET quantification at reduced doses, including novel dual-tracer studies.