Denoising of arterial spin labeling data: wavelet-domain filtering compared with Gaussian smoothing

To investigate a wavelet-based filtering scheme for denoising of arterial spin labeling (ASL) data, potentially enabling reduction of the required number of averages and the acquisition time. ASL magnetic resonance imaging (MRI) provides quantitative perfusion maps by using arterial water as an endogenous tracer. The signal difference between a labeled image, where inflowing arterial spins are inverted, and a control image is proportional to blood perfusion. ASL perfusion maps suffer from low SNR, and the experiment must be repeated a number of times (typically more than 40) to achieve adequate image quality. In this study, systematic errors introduced by the proposed wavelet-domain filtering approach were investigated in simulated and experimental image datasets and compared with conventional Gaussian smoothing. Application of the proposed method enabled a reduction of the number of averages and the acquisition time by at least 50% with retained standard deviation, but with effects on absolute CBF values close to borders and edges. When the ASL perfusion maps showed moderate-to-high SNRs, wavelet-domain filtering was superior to Gaussian smoothing in the vicinity of borders between gray and white matter, while Gaussian smoothing was a better choice for larger homogeneous areas, irrespective of SNR.