Deep Learning Image Denoising in PET: Quantitative Impact on Kinetic Modeling and Clinical Metrics

The increased sensitivity exhibited by long axial field-of-view (LAFOV) PET systems allows measurement of tracer kinetics across multiple organs and lesions at reduced doses and/or scan times. This flexibility has been utilized to perform dual-tracer imaging which in oncology offers complementary quantification about metabolic tumor processes. Yet to make dual-tracer studies more suitable for clinical practice, these scan protocols should be pushed towards lower injection doses and shorter scan times, but reduced count statistics inherently result in noisier PET images. Many have previously shown promising denoising performance of deep learning (DL) methods for low count PET images acquired at a single (static) late time point. However, only a few studies have investigated the use of DL to denoise dynamic PET data and the impact on kinetic modeling of parametric estimates. We first trained a DL algorithm (U-Net) with low-to-high count static ¹⁸ F-FDG datasets and evaluated its performance using noise and clinical metrics. We next assessed the transferability of this pre-trained DL model for frame-by-frame image denoising of dynamic ¹⁸ F-FGln and looked at the effect of such DL-based denoising on estimations of the volume of distribution of ¹⁸ F-FGln (kinetic parameter to quantify glutamine pool size). This work provides the foundation for our next investigations into the use of DL for optimizing scan protocols to make dual-tracer studies more practical, and more accurate.