Torch-Based Fitting for Accelerated Water Residual Removal in MRSI Data.

Motivation: This study addresses the time-consuming water residual removal process in MRSI, aiming to expedite it, given the lack of GPU implementation for methods like HLSVD-PRO. Goal(s): We aim to increase water residual removal speed using Torch for GPU-parallelized linear combination model fitting. Approach: Our method utilizes Torch to model and optimize water residuals fitting with 7-Lorentzian profiles. Performance is evaluated with in a large in-vivo dataset, comparing our method to HLSVD-PRO both in efficiency and accuracy. Results: Our approach accelerates water residual removal, outperforming HLSVD-PRO in processing speed by a factor 14x, while mantaining equivalent quantification accuracy, offering promise for MRSI applications. Impact: This study's accelerated water residual removal method in MRSI can benefit scientists and clinicians by reducing processing time. And it opens avenues for more extensive research in the topic.