Arterial spin labeling cerebral blood flow quantification from quantitative transport mapping based on multiscale fluid mechanics simulation and deep learning.

OBJECTIVE: The traditional kinetic model (TKM) for perfusion quantification from arterial spin labeling (ASL) assumes a global arterial input function (AIF) and uniform plug flow of spin label transport. This AIF assumption causes systematic errors of overestimating the local arterial input to a voxel, mainly due to the lack of flow velocity dispersion across and along vessels, resulting in underestimation of cerebral blood flow (CBF). Recently developed quantitative transport mapping (QTM) is free of and overcoming AIF problems. We present QTM deep neural network (QTMnet) for quantifying CBF. METHODS: QTMnet was trained on synthesized multi-delay ASL data from multiscale vascular fluid mechanics simulations for CBF quantification from ASL data. QTMnet was validated against in-silico truth using simulated brain ASL data and against phase-contrast (PC) measured total brain flow using in vivo brain ASL data; QTMnet was also compared to TKM for CBF quantification from ASL. RESULTS: QTMnet reduced the CBF underestimation in TKM by approximately 3-4X: Against the in-silico ground truth, the TKM underestimation was reduced from TKM single/multiple delay biases of -12.80/-9.73 mL/100g/min to QTMnet bias of -3.96 mL/100g/min; Against average whole brain PC flow, the TKM underestimation was reduced from single/multiple delay biases of -16.0±3.1/-13.5±3.3 mL/100g/min to QTMnet bias of -3.31±2.7 mL/100g/min. CONCLUSION: QTMnet is a promising approach for CBF quantification from ASL. QTMnet demonstrates higher accuracy in both in-silico and in-vivo ASL data compared to TKM. SIGNIFICANCE: Fluid-mechanics-based QTMnet improves CBF quantification from ASL data.