Cerebrovascular Reactivity Measurements using Simultaneous 15O-Water PET and Deep-Learning Synthesized Multi-PLD PCASL

Motivation: Arterial spin labeling offers a less invasive alternative to measure cerebrovascular reactivity (CVR) compared to 15O-water PET. Goal(s): To provide validation of deep learning methods to estimate CVR from a single PLD of multi-PLD data. Approach: Using 1 PWI and M0 image from multi-PLD PCASL data as input, a 3D H-CNN estimated CBF maps at baseline and post-acetazolamide vasodilation. CBF and CVR measures by 15O-water PET, single-PLD, multi-PLD, and synthesized multi-PLD ASL were compared. Results: Synthesized CBF and CVR were similar to the ground truth multi-PLD ASL estimates and significantly closer to the gold standard 15O-water PET in comparison to single-PLD ASL. Impact: Though single-PLD ASL offers shorter scan durations, it largely underestimates CVR. Using deep learning models to synthesize multi-PLD ASL even from a single PLD may allow for more robust CVR estimates while potentially matching the shorter scan times.