Deep Learning-Based MR Vascular Fingerprinting for Dynamic Quantification of Cerebral Hemodynamics During Gas Challenges

Motivation: Dynamic contrast enhancement MRI and fMRI provide perfusion and oxygenation information but cannot quantify multiparametric estimations. Magnetic Resonance Vascular Fingerprinting (MRvF) decodes multiple physiological parameters in one acquisition. Deep learning enhances MRvF for accurate estimations. Goal(s): Using deep learning-based MRvF, we quantify cerebral oxygen saturation, blood volume (CBV), vessel radius (R), and T2 across normoxia, hyperoxia, and hypoxia. Approach: We train GESFIDE-based (gradient-echo sampling of free induction decay and echo) model on synthetic data to estimate parameters during gas challenge. Results: We show increased T2 in GM during hyperoxia and elevated CBV and R in GM and WM during hypoxia, aligning with literature. Impact: We demonstrate that deep learning-enhanced MRvF offers improved accuracy in quantifying cerebral oxygenation dynamics compared to dictionary matching. By testing differences during controlled gas challenges, we show expected range and dynamics of vascular parameter quantifications for understanding underlying brain health.