Towards harmonized spectral quantification in MRSI: Comparative analysis of Backward-Linear-Predicted and original ¹H-FID-MRSI dephased data

Object We hypothesized that the inherent acquisition delay (AD) in ¹H-FID-MRSI can introduce systematic LCModel quantification biases due to strong spectral dephasing, and that Backward-Linear-Prediction (BLP) reconstruction toward AD = 0 ms can harmonize metabolite estimates across acquisitions with various delays. Materials and Methods 2D ¹H-FID-MRSI were acquired in rats at 14.1T with three AD values (0.71, 0.94, 1.30 ms). Hippocampal metabolites were quantified using LCModel and AD-matched basis sets. Complementary Monte-Carlo simulations (n = 1000) replicated ¹H-FID-MRSI spectra at multiple ADs under realistic SNR conditions. BLP was applied to in vivo and simulated FIDs to back-predict missing points up to AD = 0 ms, enabling quantification within a unified basis set framework. Results In vivo and simulated data showed clear AD-dependent variations for several metabolites (Gln, tCho, tNAA, Ins, Tau), with discrepancies frequently >10% despite AD-specific basis sets. Simulations confirmed metabolite-specific biases increasing with AD. BLP reconstruction preserved quantification consistency up to ∼0.98 ms of recovered FIDs, reducing inter-AD mismatches in vivo —particularly for Tau, tNAA and tCho—lowering the mean discrepancy from 10.5% to ∼5%. Discussion These findings show that AD affects ¹H-FID-MRSI quantification in LCModel, whereas BLP reconstruction can harmonize spectra across delays by enabling a virtual AD = 0 ms quantification scheme. This supports BLP as a practical strategy to improve consistency and comparability in MRSI studies.