An adaptive deep learning reconstruction for both hyperpolarized 13C magnetic resonance spectroscopic imaging and deuterium metabolic imaging

Motivation: Heteronuclear magnetic resonance spectroscopic imaging (MRSI) can assess tumor aggressiveness and response to treatments. Regarding its slow acquisition and starving signal-to-noise ratio, efficient deep-learning reconstruction adaptive for different applications is required. Goal(s): To propose an adaptive deep learning method for reconstructing high-quality MRSI. Approach: A deep learning prior was trained using singular maps extracted from hyperpolarized 13C MRSI and deuterium metabolic imaging (DMI), generated through multi-pool exchange and free induction decay. The prior was incorporated with SPICE and date fidelity terms for MRSI reconstruction. Results: The model was evaluated on various datasets, demonstrating its generalizability in reconstructing high-quality MRSI using high acceleration rates. Impact: The generalizability of the proposed pipeline for high-quality MRSI reconstruction has been demonstrated in various applications, including HP 13C MRSI and DMI, suggesting its feasibility as a molecular imaging tool for both scientific and clinical applications.