AI Insight
WALRUS is a deep learning framework that removes water and lipid contamination from proton magnetic resonance spectroscopic imaging (1H-MRSI) brain scans more effectively than conventional methods. The system uses empirically calibrated simulated training data matched to the same metabolite basis used in downstream analysis, and unlike previous approaches, does not require subject-specific lipid segmentation or auxiliary inputs during use. Testing at both 7 Tesla and 3 Tesla magnetic field strengths demonstrated improved metabolite quantification accuracy and better generalization across different acquisition settings.
Why it matters
This framework could improve the reliability of non-invasive brain metabolism measurements used in neurological research and clinical diagnostics. The simplified inference process and systematic adaptability to different scanner settings make it more practical for widespread clinical implementation than previous deep learning approaches.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Proton magnetic resonance spectroscopic imaging (1H-MRSI) enables non-invasive mapping of brain metabolism, but reliable quantification is challenged by residual water and extracranial lipid contamination. Recent work demonstrated that a deep learning model trained on simulated metabolite spectra combined with in vivo-derived water and lipid signals can outperform conventional nuisance-removal methods. However, this approach relied on heuristic simulation parameter distributions and a subject-specific L2-derived lipid estimate as an auxiliary network input, requiring subject-specific lipid masks and L2 processing during inference. Here, we introduce WALRUS (Water And Lipid Removal Using Spectral Learning), a fitting-consistent and empirically calibrated framework developed through a systematic evaluation of simulation realism, architectural priors, and network optimization. WALRUS uses the same metabolite basis as downstream spectral fitting and calibrates deliberately broadened simulation distributions from in vivo measurements. Controlled 7 T experiments demonstrate that empirical simulator calibration substantially improves nuisance removal and metabolite quantification compared with heuristic training distributions. With the optimized simulator, a capacity-matched comparison reveals no clear additional benefit from the L2-derived prior, allowing its removal and substantially simplifying inference. Further improvements are achieved with the final higher-capacity U-Net, while on-the-fly simulation improves generalization compared with training on a fixed synthetic dataset. The same framework was subsequently recalibrated and retrained at 3 T, where WALRUS outperformed conventional HLSVD+L2 processing on held-out data. Together, WALRUS combines these advances into an optimized framework that improves nuisance removal and downstream metabolite fitting, requires no subject-specific lipid segmentation or L2-derived auxiliary input during inference, and can be systematically adapted to different acquisition settings.
Source: WALRUS: Robust and Transferable Deep Learning for Water and Lipid Removal in 1H-MRSI