A Two-Stage Multi-loss Super-Resolution Network for Arterial Spin Labeling Magnetic Resonance Imaging

Abstract

Arterial spin labeling (ASL) perfusion magnetic resonance imaging (MRI) is a non-invasive technique for quantifying cerebral blood flow (CBF). Limited by the T1 decay rate of the labeled spins, very short time is available for data acquisition after one spin labeling cycle, resulting in a low spatial resolution. The traditional strategy to achieve high spatial resolution in ASL MRI is to add more labeling cycles. However, the total acquisition time is exponentially prolonged, making it highly sensitive to motions. Moreover, signal-to-noise-ratio (SNR) drops as spatial resolution increases. There needs an alternative approach to improve spatial resolution and SNR for ASL MRI without increasing scan time. Therefore, we propose a novel two-stage multi-loss super-resolution (SR) network (TSMLSRNet) for reconstruction of high resolution ASL images. Specifically, the first stage network uses the mean squared error (MSE) loss function to produce a first SR estimate, while the second stage network adopts the gradient sensitive (GS) loss function to further improve high-frequency details for the output SR image. The multi-loss joint training strategy is finally used to preserve both the low-frequency and high-frequency information of the ASL images. Moreover, the noise in ASL images is simultaneously reduced. Validation results using in-vivo data clearly show the effectiveness of the proposed ASL SR algorithm that outperforms state-of-the-art image reconstruction algorithms.