Deep learning enables quantitative kinetic modeling from low-dose dynamic [$$^{18}$$F]-MK6240 Tau PET

Abstract Dynamic positron emission tomography (PET) imaging with the tau tracer [ $$^{18}$$ F]-MK6240 is widely used in Alzheimer’s disease research, but the time frames of a dynamic acquisition are short and contain few detected events. This makes the images noisy, and leads to inaccurate estimates of kinetic parameters, especially when estimating kinetic parameters at the voxel level. Reducing the injected dose would lower radiation exposure and make repeated scanning more practical, but only if kinetic accuracy is preserved. In this work we evaluated two deep learning denoising methods, U-Net and Restormer, on low-dose (10% of full dose) dynamic [ $$^{18}$$ F]-MK6240 PET data from 59 subjects, covering cognitively normal individuals and patients with mild cognitive impairment and Alzheimer’s disease. We assessed performance through time-activity curve fidelity, parametric map quality, and regional bias and variance analysis for two clinically critical biomarkers, namely the relative tracer delivery rate R 1 and the distribution volume ratio DVR. Both methods clearly reduced bias and standard deviation of kinetic parameters compared to unprocessed low-dose images. Restormer showed lower bias than U-Net across most brain regions and time frames, with better preserved TAC shape particularly in the late frames critical for DVR estimation. The results support that a 90% dose reduction is possible without losing the quantitative accuracy needed for tau burden assessment in clinical and research settings.