🤖 AI Summary
Whole-brain ¹H-MRSI suffers from concurrent contamination by scalp lipid signals and residual water artifacts, severely limiting metabolite quantification accuracy. To address this, we propose WALINET—a novel end-to-end dual-branch convolutional neural network—marking the first supervised deep learning framework for joint water and lipid suppression in MRSI. WALINET integrates spectroscopic and spatial inputs, jointly encodes frequency-domain specificity and spatial consistency, and employs registration-free spectral-domain augmentation. Validated across multicenter datasets, it achieves residual water/lipid levels below 0.5%, enhances metabolite SNR by 3.2×, and processes up to 500 voxels per second. This enables real-time, high-resolution MRSI post-processing, establishing a robust, efficient signal separation paradigm for whole-brain noninvasive metabolic imaging.
📝 Abstract
Proton magnetic resonance spectroscopic imaging ( 1H$$ {}^1mathrm{H} $$ ‐MRSI) provides noninvasive spectral‐spatial mapping of metabolism. However, long‐standing problems in whole‐brain 1H$$ {}^1mathrm{H} $$ ‐MRSI are spectral overlap of metabolite peaks with large lipid signal from scalp, and overwhelming water signal that distorts spectra. Fast and effective methods are needed for high‐resolution 1H$$ {}^1mathrm{H} $$ ‐MRSI to accurately remove lipid and water signals while preserving the metabolite signal. The potential of supervised neural networks for this task remains unexplored, despite their success for other MRSI processing.