A Generalizable Pipeline for MRI Sequence Design Using Bayesian Optimization

Motivation: Conventional methods for designing MRI sequences are effective in specific scenarios but lack generalizability across various contexts, highlighting the need for a more universally applicable approach. Goal(s): To establish a generalizable pipeline for MRI sequence design capable of meeting the broad spectrum of clinical needs. Approach: We propose a globally sample-efficient Bayesian optimization-based method for MRI sequence design, enhanced by an iterative, data-driven approach that aligns the optimal rotation scheme with the sequence parameters. Simulation experiments were performed across three scenarios. Results: Results demonstrated that the proposed method improved imaging quality in abdominal T1-weighted spoiled-GRE, brain MRF, and hyperpolarized carbon-13 phantom imaging. Impact: This novel, generalizable pipeline enables the development of optimized, tailored imaging protocols to address diverse clinical needs across various MR applications. It holds the potential to integrate advanced imaging techniques into routine practice, enhancing both diagnostic and research efficiency.