WALRUS: Robust and Transferable Deep Learning for Water and Lipid Removal in 1H-MRSI

WALRUS improves 7 T 1H-MRSI water and lipid nuisance removal and metabolite quantification by empirically calibrating broadened simulation distributions, outperforming heuristic training.

• Why it matters: Reliable brain metabolic mapping with 1H-MRSI is limited by residual water and extracranial lipid contamination, and prior deep learning approaches required subject-specific L2-derived lipid inputs and masks.
• What they did: WALRUS uses a fitting-consistent metabolite basis with deliberately broadened simulations calibrated from in vivo measurements, then evaluates simulation realism, architectural priors, and network optimization; it also tests removing the L2-derived prior and uses on-the-fly simulation.
• The result: In controlled 7 T experiments, empirical calibration substantially improves nuisance removal and quantification versus heuristic distributions, and a capacity-matched comparison finds no clear added benefit from the L2-derived prior, enabling inference without subject-specific lipid segmentation or L2 input; at 3 T, WALRUS outperforms conventional HLSVD+L2 on held-out data.

AI summary of the abstract · may contain errors · not medical advice · read the paper.