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

Abstract

Purpose: Removing water residual signals from MRS spectra is crucial for accurate metabolite quantification. However, currently available algorithms are computationally intensive and time-consuming, limiting their clinical applicability. This work aims to propose and validate two novel pipelines for fast water residual removal in MRS.

Methods: Two methods for water residual removal are proposed and evaluated: DeepWatR and WaterFit. DeepWatR uses a U-Net-like architecture with an attention mechanism for skip connections. WaterFit uses Torch auto-differentiation to estimate parameters for a 7-pool Lorentzian model. The accuracy and time-efficiency of these methods were assessed using simulated and in vivo 1H brain datasets and compared with the gold standard, Hankel Lanczos singular value decomposition method (HLSVD)Pro.

Results: In the simulated dataset, the average percentage quantification error was $\left(8.54\pm 17.49\right)$ % for DeepWatR and $\left(7.86\pm 16.69\right)$ % for WaterFit, both comparable to $\left(7.68\pm 15.6\right)$ % for HLSVDPro in the main metabolites (Cr, Cho, and NAA). DeepWatR was 51 times faster and WaterFit was 22.7 times faster than HLSVDPro for a dataset of 10 000 voxels when using a low-end graphics processing unit. For 100 voxels, the speed-up is 6.5 and 7.5 times faster for DeepWatR and WaterFit, respectively. WaterFit showed higher metabolite fitting accuracy after water removal compared to DeepWatR.

Conclusion: WaterFit showed a superior balance of accuracy and processing speed in removing water residual from MRS data compared to DeepWatR. The proposed WaterFit implementation significantly reduces preprocessing time while maintaining metabolite fitting accuracy comparable to gold standard methods. This advancement addresses the need for efficient processing methods that can facilitate analysis and enhance the clinical utility of MRS.