Accelerated water residual removal in MRS: Exploring deep learning versus fitting-based approaches.

Area of Science:

• Magnetic Resonance Spectroscopy (MRS)
• Metabolite Quantification
• Signal Processing

Background:

• Water residual signals in MRS spectra complicate accurate metabolite quantification.
• Existing water removal algorithms are computationally intensive, limiting clinical application.

Purpose of the Study:

• To propose and validate two novel, fast pipelines for water residual removal in MRS.
• To improve the clinical applicability of MRS through efficient data processing.

Main Methods:

• Developed DeepWatR (U-Net architecture with attention) and WaterFit (Torch auto-differentiation with Lorentzian model).
• Evaluated accuracy and time-efficiency using simulated and in vivo 1H brain MRS data.
• Compared performance against the gold standard Hankel Lanczos singular value decomposition method (HLSVD)Pro.

Main Results:

• DeepWatR and WaterFit showed quantification errors comparable to HLSVDPro on simulated data.
• WaterFit was 22.7x faster and DeepWatR was 51x faster than HLSVDPro on a large dataset.
• WaterFit demonstrated higher metabolite fitting accuracy post-water removal compared to DeepWatR.

Conclusions:

• WaterFit provides an optimal balance of accuracy and speed for water residual removal in MRS.
• The WaterFit pipeline significantly reduces preprocessing time while maintaining metabolite quantification accuracy.
• This advancement enhances the clinical utility of MRS by enabling faster, more accurate analysis.