Improved Nuisance Signal Removal for 3D 1H-MRSI Using Physics-Driven Subspace Learning

Motivation: Accurate subspace estimation in the UoSS model is essential for removing intensive lipid signals in 1H-MRSI without lipid suppression. Goal(s): Our goal was to learn the subspace such that UoSS model provided better lipid signal estimation in 3D 1H-MRSI without lipid suppression. Approach: A novel physics-based subspace learning incorporating all the lipid spectral components enhanced the UoSS-based removal of unsuppressed lipid signals in 3D 1H-MRSI and was tested on in-vivo MRSI data. Results: Our proposed method demonstrated the capability of estimating signals from the lipid components (0.9-2.77 ppm) while preserving the metabolite of interest in MRSI (4.4 × 4.4 × 6.4 mm3 resolution). Impact: The proposed method can potentially accelerate the data acquisition time and improve the nuisance removal outcome in 3D 1H-MRSI without lipid suppression.