An AI-Based Solution for Denoising Fast-Acquisition [F]FDG PET: Clinical Feasibility and Quantitative Assessment.

Benefits in patient comfort, efficiency, and sustainability can come from reducing positron emission tomography (PET) scan's acquisition duration. This study assesses the clinical adequacy of restoring fast-acquisition F-fluorodeoxyglucose ([F]FDG) PET to its standard-of-care image quality through deep-learning-based (DL) methods. Fast and standard whole-body [F]FDG PET acquisitions of 117 oncological patients were included in the training and testing of three convolutional neural networks. The best-performing network during training was chosen for clinical evaluation on the test set (N = 25). Visual assessment and lesion detectability of the fast acquisitions, of 20 and 30 seconds per axial field of view (s/AFOV), with and without DL-based denoising, and of the local standard of care, of 70 s/AFOV, were performed by three experienced nuclear medicine physicians. Quantification was conducted globally (voxel-wise), in healthy organs and the reported lesions. Optimised Gaussian and non-local means filters served as benchmarks. Visual assessment revealed 20 and 30 s/AFOV with DL-based denoising to have similar image quality to the standard of care. Average lesion-based sensitivity and positive predictive value were 74% and 72%, respectively, for 20 s/AFOV + DL and 72% and 80% for 30 s/AFOV + DL. DL-based denoising displayed the highest voxel-wise agreement with the standard-of-care (p < 0.001). Liver and lungs in the DL-denoised images exhibited a higher signal-to-noise ratio than the standard of care. The median absolute maximum standardised uptake value deviation in the lesions was as low as 0.39 for 20 s/AFOV + DL and 0.30 for 30 s/AFOV + DL. The proposed DL-based method proved to be suitable for the restoration of fast-acquisition whole-body [F]FDG PET, having resulted in images similar to the standard-of-care acquisitions. DL-based denoising outperformed standard benchmark methods.