Motivation: MRSI provides label-free mapping of brain metabolites, but its accuracy is compromised by B₁ inhomogeneity. Goal (s): To develop a learning-based B₁ correction method leveraging a unified subspace model derived from multiple centers for MRSI across scanners. Approach: Non-water-suppressed MRSI and VFA water data were acquired from 129 healthy subjects across three centers. A unified subspace model was built and learned from pre-scanned multi-center training data. The B₁ fields for new data were estimated using a maximum-a-posterior Bayesian approach. Results: In both healthy subjects and stroke patients, the proposed method achieved robust high-quality B₁ maps and produced significantly improved neurometabolite maps. Impact: The proposed B₁ correction method will enhance the quantitative accuracy of metabolite measurements and thus enhance the robustness and practical usefulness of MRSI in clinical applications.
Meng et al. (Tue,) studied this question.