Generative Adversarial Network‐Enhanced Ultra‐low‐dose [18F]‐PI‐2620 Tau PET/MR Imaging

Abstract Background In neurodegenerative diseases such as Alzheimer’s disease, simultaneous PET/MRI provides a “one‐stop shop” to acquire biomarkers as outlined by the AT(N) criteria. However, to enable large‐scale imaging studies of dementia with frequent follow‐ups, a reduction of injected radiotracer dose and the enhancement of the degraded image quality after reconstruction needs to be addressed. With deep learning‐based methods, we have previously demonstrated the use of U‐Net‐based methods that incorporate spatially correlated PET/MRI information to produce high quality 18 F‐florbetaben amyloid PET images from scan protocols with markedly reduced injected radiotracer dose (as low as ∼1% dose). However, for tau radiotracers, where the uptake is focal, has greater image noise, and the signal is weaker than that of amyloid, we also aim to investigate whether a generative adversarial network (GAN) can enhance ultra‐low‐dose tau PET images and whether the generated images are of diagnostic quality. Method 44 participants (18 female, 70.2±7.5 years) were recruited to train the ultra‐low‐dose tau GAN; 221±61 MBq of the tau radiotracer 18 F‐PI‐2620 was injected. T1‐weighted and T2‐FLAIR MRI were acquired simultaneously. A subset of the list mode consisting of 1/20 of the events (simulating 5% dose) was reconstructed to produce an ultra‐low‐dose PET image and used as inputs along with the MR images to train the GAN. The image quality of the enhanced PET and the ultra‐low‐dose PET images were compared to the original full‐dose image using three image‐based metrics ; the mean and standard deviation of uptake in the medial temporal lobe and inferior temporal cortex as well as SUVRs were also calculated. Three clinicians were recruited to evaluate tracer uptake on various regions of interest. Result The enhanced images showed significant noise reduction compared to the low‐dose image (Figure); uptake patterns in the low‐dose and enhanced images were read accurately (Gwet’s AC1>0.84 for all relevant regions) compared to their full‐dose counterparts. Conclusion The deep learning‐enhanced images had marked noise reduction and were able to be read clinically for regional uptake patterns of tau accumulation similarly as the full‐dose images. This technique can potentially increase the utility of hybrid PET/MR imaging in clinical diagnoses and longitudinal studies.