Optimizing arterial spin labeling denoising with deep learning – effects of averaging and training strategies

Motivation: Optimizing deep-learning based denoising of ASL is of great interest, but has not been systematically studied. Goal(s): To investigate the effects of different averaging methods on deep-learning based denoising for optimized performance. Approach: ASL datasets of different SNR levels were created using different levels of averaging and methods (windowed vs. interleaved). Various full-3D Unet-based models were trained and compared. The effect of including the M0 scan was also studied. Results: Optimal denoising performance was observed when the SNR levels were matched in training and testing, with windowed averaging. Including M0 can provide additional improvement, especially at low SNR. Impact: This is the first study to systematically investigate the effects of different averaging and training strategies in DL-based ASL denoising. The results will help guide ASL scan and post-processing designs for significant reduction of scan time while achieving optimal performance.