Improving Fetal Single‐Voxel Magnetic Resonance Spectroscopy With Spectral Registration

ABSTRACT

Purpose

To evaluate whether spectral registration improves spectral quality and quantification reliability in in vivo single‐voxel fetal brain MRS:

Methods

Fourteen fetal PRESS datasets were acquired at 3 T at TE = 35 ms and TE = 135 ms. Raw data were processed with eddy‐current phase correction, coil combination, spectral registration, and outlier rejection; spectra processed with eddy‐current correction alone were treated as uncorrected. Corrected and uncorrected spectra were compared using LCModel quality metrics (SNR, FWHM, and CRLB), metabolite concentrations and ratios, and qualitative spectral assessment.

Results

Spectral registration improved spectral appearance, with sharper peaks and reduced line broadening. SNR increased and FWHM decreased significantly at both echo times. At TE = 35 ms, pooled CRLB values decreased significantly after correction, and more estimates met CRLB thresholds, most notably for CRLB ≤ 50; corresponding CRLB improvements at TE = 135 ms were limited. No metabolite concentration comparison remained significant after false‐discovery‐rate correction, and only Glx/tCho at TE = 35 ms remained significant among metabolite ratios.

Conclusion

Spectral registration improved fetal MRS spectral quality and modestly improved quantification reliability, with the clearest benefit at short TE. These findings support spectral registration as a practical retrospective processing step for motion‐prone fetal MRS.

1. Introduction

Magnetic resonance spectroscopy (MRS) provides a unique, noninvasive means of probing cerebral metabolism in vivo and has become an important adjunct to structural MRI for assessing brain development and pathology. In the prenatal and perinatal setting, proton MRS has been used to characterize normal metabolic maturation [1, 2, 3, 4, 5] and metabolic alterations associated with hypoxic–ischemic injury in intrauterine growth restriction [6, 7, 8, 9, 10, 11], cardiac dysfunction such as congenital heart disease [12], intrauterine infection [13], and inborn errors of metabolism [14, 15, 16]. Compared with postnatal or adult applications, however, fetal MRS remains technically challenging, and its broader clinical and research adoption has been limited by low signal‐to‐noise ratio (SNR), broad spectral linewidths, and motion‐induced artifacts [17, 18, 19, 20, 21].

Fetal MRS is especially susceptible to signal degradation due to uncontrolled fetal motion, maternal respiration, temporal instability of the main B₀ magnetic field, and spatial B₀ and B₁ inhomogeneities. Because single‐voxel MRS relies on signal averaging to improve SNR, frequency and phase variations between successive acquisitions can cause destructive interference during conventional averaging, reducing effective SNR and broadening spectral peaks. These effects are particularly problematic in the fetal brain, where voxel sizes are limited by anatomy and safety considerations, metabolite concentrations are low, and scan times must be minimized [17, 18, 22]. Consequently, reliable in utero detection and quantification of metabolites such as N‐acetylaspartate, choline, creatine, and lactate remain challenging.

Several retrospective post‐processing approaches have been proposed to mitigate signal loss in fetal MRS. These methods share the principle of constructive averaging, in which individual transients are corrected for relative phase and frequency offsets before summation to preserve signal coherence. In fetal MRS, Shetty et al. [23] showed that this approach improves spectral quality and SNR, and Evangelou et al. [24] applied a similar strategy specifically to reduce motion‐induced spectral artifacts. Together, these studies established the value of retrospective phase‐ and frequency‐corrected averaging in fetal MRS.

Spectral registration is a related implementation of this framework, in which relative frequency and phase inconsistencies are jointly estimated directly from the free induction decays (FIDs) using a time‐domain nonlinear least‐squares optimization, enabling robust correction of inter‐average misalignment [25, 26, 27]. Its time‐domain formulation can also emphasize the early, high‐SNR portion of the FID, further improving robustness of the parameter estimates. This approach has been applied successfully in adult [28, 29] and pediatric [30, 31] MRS studies, including in challenging regions such as the brainstem, where Sirucek et al. [32] showed improved metabolite quantification accuracy and partial mitigation of motion‐related error in single‐voxel MRS. In this context, the present study extends prior fetal constructive‐averaging work [23, 24] by applying a spectral‐registration‐based implementation of retrospective phase‐ and frequency‐corrected averaging to fetal MRS.

Despite its utility in other anatomically and physiologically challenging settings, the extent to which this technique improves fetal MRS data remains mostly unexplored. The purpose of this study was therefore to evaluate the impact of spectral registration on in vivo single‐voxel fetal MRS by assessing qualitative improvements in spectral appearance and quantitative changes in spectral quality including SNR, linewidth, metabolite detectability, and reliability of quantification after correction. By comparing fetal MRS data processed with and without spectral registration, this work aimed to clarify the extent to which this technique can improve the reliability and interpretability of fetal spectroscopic measurements and support a more robust metabolic assessment of the developing human brain.

2. Methods

2.1. MRS Acquisition

All fetal MRS data were acquired on a 3 T Siemens Vida MR system, using a single‐voxel PRESS sequence [33] with TR = 2000 ms and TEs of 35 ms and 135 ms, enabling evaluation of spectral registration under differing spectral conditions, including variation in lipid contamination, macromolecular baseline, and SNR. RF excitation was delivered via the built‐in body coil, and MR signals were received using an anterior phased‐array body coil (Body 18) operating in conjunction with the integrated spine matrix coil (BioMatrix Spine 32 RS). The volume of interest (VOI) was placed entirely within the fetal cranium and ranged from 10.0 to 18.8 mL (mean = 16.6, sd = 2.7), with slight adjustment for fetal head size (Figure 1A). Gestational age ranged from 20.6 to 35.1 weeks (mean = 28.5, sd = 4.6). For each TE, we analyzed 14 datasets from 13 fetal subjects, including two datasets from a single subject. The study was approved by the institutional review board and written informed consent was obtained from all pregnant women who participated in the study. Each acquisition included 96 water‐suppressed averages and an unsuppressed water reference scan. Spectral data were acquired with 2048 points, a receiver bandwidth of 2500 Hz, and a transmit frequency offset of −2.3 ppm relative to water. Water suppression used WET [34] (40 Hz bandwidth), without outer‐volume saturation bands, while shimming and water frequency adjustments relied on vendor‐provided automated pre‐scans.

FIGURE 1.

(A) Sagittal, coronal, and axial reformats of a maternal abdominal TRUE‐FISP T₂‐weighted localizer showing the fetal brain with the prescribed single‐voxel PRESS VOI overlaid. (B) Flowchart summarizing the reconstruction and preprocessing pipeline: Coil noise decorrelation/whitening, water‐reference–based eddy‐current correction and coil combination, data‐driven estimation of $t_{\max }$ (noise‐floor cutoff), spectral registration (cross‐correlation initialization followed by nonlinear least‐squares refinement), outlier rejection of corrupted averages, and final averaging with LCModel quantification.

2.2. Preprocessing

Raw data were extracted and processed using a custom MATLAB pipeline (R2024b; The MathWorks, Natick, MA) (Figure 1B). Noise covariance estimated from the vendor noise pre‐scan was used to whiten the multi‐coil data via eigen‐decomposition [35], with the same whitening operator applied to both the water‐suppressed and unsuppressed water‐reference data across all averages. Eddy‐current correction (ECC) was then applied [36] and coil data were combined by weighted summation using sensitivity weights derived from the water reference [35]. To define a robust fitting window for spectral registration, a data‐driven cutoff time, $t_{\max }$, corresponding to the onset of the noise floor, was estimated from the multi‐average FIDs using a combined metric based on signal amplitude and phase variation (see Supporting Information in Section 1.1). Spectra processed with ECC alone were termed uncorrected, whereas spectra additionally processed with spectral registration and outlier rejection were termed corrected. Thus, the “uncorrected” spectra were ECC‐corrected but not spectrally registered across averages.

2.3. Spectral Registration and Outlier Rejection

Alignment of individual averages was performed using spectral registration formulated as a nonlinear least squares (NLS) optimization problem [25, 26]:

$$\mathrm{\Delta }{f}_a,\mathrm{\Delta }{\phi }_a=\underset{\mathrm{\Delta }f,\mathrm{\Delta }\phi }{\text{argmin}}{\frac{1}{2}‖s_{\mathit{ref}}(t)-s_a(t)e^{i2\pi \Delta f·t}{e}^{i\mathrm{\Delta }\phi }‖}_2^2$$

in which the frequency and phase shift shifts ($\mathrm{\Delta }{f}_a$, $\mathrm{\Delta }{\phi }_a$) were estimated for each average $s_a(t)$ relative to a robust reference $s_{\mathrm{ref}}(t).$ Corrections were applied as $s_a^{\text{corr}}(t)=s_a(t)\exp \left(i2\pi \mathrm{\Delta }{f}_{a}t\right)\exp \left(i\mathrm{\Delta }{\phi }_a\right)$. To improve the initialization of the NLS algorithm, frequency‐domain cross correlation [37] was performed over a metabolite‐rich band (e.g., 1.6–5.4 ppm), including the relatively high‐SNR residual water peak ($\sim$ 4.65 ppm), using the pointwise median spectrum as the reference to obtain stable initial estimates $\mathrm{\Delta }{f}_a^{(0)}$ and $\mathrm{\Delta }{\phi }_a^{(0)}$ for generating an initial aligned dataset. To enhance robustness in low‐SNR conditions, low‐rank SVD denoising across the average dimension was applied prior to parameter estimation (e.g., retaining $r=20$ components). NLS spectral registration was then performed in the time domain over $t\le t_{\max }$ (noise‐floor–based cutoff) using the median average FID as the reference; a Gaussian/resolution‐enhancement filter was applied to down‐weight the later time points. The final resulting registration parameters were then applied to the unfiltered data prior to averaging.

Outlier rejection was performed after estimating registration parameters but before final averaging to prevent motion‐corrupted averages from degrading the spectrum [23, 32]. Outliers were detected iteratively in the frequency domain using a pointwise median reference spectrum and a composite score based on spectral correlation, normalized spectral distance, and residual drift (cross‐correlation lag). Averages with scores exceeding a threshold based on the mean absolute deviation (MAD), that is, median + 4 $\times$ MAD, were excluded prior to final averaging (see Supporting Information in Section 1.2).

2.4. Quantitation

After spectral registration and outlier removal, metabolite quantification was performed using LCModel [37] (version 6.3) with simulated basis sets appropriate for the TE = 35 ms and TE = 135 ms data, and a fitted spectral range from 0.6–4.3 ppm. The metabolites in the basis sets included N‐acetylaspartate (NAA), creatine (Cr), choline (Cho), myo‐inositol (mI), scyllo‐inositol, lactate (Lac), glutamate (Glu), glutamine (Gln), glutathione (GSH), glycerophosphocholine (GPC), N‐acetylaspartylglutamate (NAAG), γ‐aminobutyric acid (GABA), phosphoethanolamine (PE), ascorbate (Asc), phosphocholine (PCh), phosphocreatine (PCr), aspartate (Asp), and taurine (Tau).

Quantitative analyses focused on metabolite ratios relative to total Creatine (tCr = Cr + PCr), total NAA (tNAA = NAA + NAAG), and total Choline (tCho = Cho + PCh + GPC), the Cramér–Rao lower bounds (CRLBs, reported as percentages), LCModel‐reported SNR and estimates of the full width at half maximum (FWHM) linewidth based on the NAA singlet. Because VOI size varied across fetal datasets, LCModel‐reported SNR values were normalized by the subject‐specific voxel volume (mL). These metrics were compared between spectra processed with and without spectral registration. Metabolite ratios referenced to tCr, tNAA, and tCho were reported to provide a comprehensive and literature‐comparable summary of the quantification results; however, because these ratios are algebraically related, they were interpreted as complementary ratio‐based descriptions rather than as independent statistical endpoints. As an additional spectral‐quality analysis, SNR and FWHM were estimated directly from the tCho peak, which is more consistently prominent across the gestational‐age range studied (see Supporting Information in Section 1.3).

2.5. Statistical Analysis

2.5.1. SNR and Linewidth

LCModel‐reported SNR and FWHM were used to assess spectral quality, with SNR normalized by subject‐specific voxel volume before statistical analysis. Volume‐normalized SNR and FWHM derived from the Cho peak were also evaluated. Analyses were performed separately for TE = 35 ms and TE = 135 ms. Because the same subjects were evaluated before and after correction, uncorrected and corrected values were compared using within‐subject paired testing. The primary inferential test was a two‐sided Wilcoxon signed‐rank test for paired differences. Paired t‐tests were additionally reported as complementary parametric comparisons. To control family‐wise error across the four primary comparisons (2 metrics × 2 TEs), signed‐rank p‐values were adjusted using the Holm–Bonferroni procedure (α = 0.05). Subject‐level effects were visualized with paired dot plots (uncorrected vs. corrected) with median and interquartile range overlays.

2.5.2. Cramér‐Rao Lower Bounds

CRLB analysis was restricted to tNAA, tCr, tCho, Glx (Glu + Gln), mI, and Lac, and was performed separately for the TE35 and TE135 datasets. Because corrected and uncorrected spectra were obtained from the same subjects, all comparisons were treated as paired. Overall differences in CRLB were assessed using the Wilcoxon signed‐rank test applied across all selected metabolite‐subject pairs within each TE. In addition, metabolite‐wise paired comparisons were performed across subjects for each metabolite, with Benjamini‐Hochberg false discovery rate correction applied to account for multiple testing. To assess the practical effect of spectral registration on quantitation reliability, the numbers and proportions of estimates meeting CRLB thresholds of ≤ 15%, ≤ 30%, and ≤ 50% were compared between corrected and uncorrected data using McNemar's exact test for paired binary outcomes. Absolute changes in the number of estimates meeting each threshold were also reported.

2.5.3. Metabolite Concentrations and Ratios

Comparative statistical analyses of LCModel‐derived metabolite concentrations (in institutional units, I.U.) and metabolite ratios were performed separately for TE = 35 ms and TE = 135 ms using paired measurements from corrected and uncorrected spectra. Analyses were restricted to tNAA, tCr, tCho, Glx, mI, and Lac. For metabolite concentrations, only estimates with CRLB ≤ 30% in both corrected and uncorrected spectra were included. For metabolite ratios relative to tNAA, tCr, and tCho, both the numerator and denominator metabolites were required to meet the same CRLB criterion in both conditions. Corrected and uncorrected values were summarized as median (interquartile range, IQR) and compared using Pratt signed‐rank tests. Effect size was reported as $r=\mid Z\mid /\sqrt{n}$. Multiple comparisons were controlled separately within each TE and outcome family using the Benjamini‐Hochberg false discovery rate procedure (N‐FDR = 6 for concentrations and N‐FDR = 5 for each ratio family), with statistical significance defined as $p_{\mathit{FDR}}<0.05$. An analysis of the between‐subject coefficients of variation (CVs) for each TE and selected metabolite concentrations and ratios before and after correction was also performed (Supporting Information in Section 1.4).

3. Results

Representative spectra demonstrated clear qualitative improvement after spectral registration, including greater apparent peak amplitude, improved definition of relatively low‐SNR and spectrally crowded resonances such as Glx and mI, and reduced linewidth relative to spectra processed without spectral registration (Figure 2). Across the 14 datasets, spectral registration and subsequent outlier rejection identified a mean of 10 ± 7 of 96 averages (11%) as outliers at TE = 35 ms and 7 ± 6 of 96 averages (8%) at TE = 135 ms, indicating that a meaningful fraction of individual transients was sufficiently corrupted to degrade conventional averaging. Representative average‐to‐average frequency and phase offsets are shown for both TEs in Figure S1. B₀ frequency drift across all datasets is shown in Figure S2.

FIGURE 2.

Representative fetal brain PRESS spectra (gestational age: 29 weeks) processed with and without spectral registration at TE = 35 ms (A) and TE = 135 ms (B). For each TE, the left panel shows the corrected individual spectra with the corresponding average overlaid, the middle panel shows the uncorrected individual spectra with the corresponding average overlaid, and the right panel directly compares the corrected and uncorrected average spectra. Spectral registration improves alignment across averages and yields a sharper averaged spectrum with improved peak definition.

3.1. Statistical Analysis

3.1.1. SNR and Linewidth

Paired dot plots (Figure 3) demonstrate consistent improvement in spectral quality after spectral registration across subjects (n = 14), with most subjects showing increased volume‐normalized SNR and reduced FWHM at both echo times. At TE = 35 ms, volume‐normalized SNR increased after correction (median 0.702 mL⁻¹ [IQR 0.600–0.816] uncorrected vs. 0.843 mL⁻¹ [0.720–0.944] corrected; signed‐rank p = 0.000122, Holm‐adjusted p = 0.000488), while FWHM decreased (5.42 Hz [4.93–7.52] vs. 3.94 Hz [3.45–4.81]; p = 0.0004883, Holm p = 0.001953). At TE = 135 ms, volume‐normalized SNR likewise increased (0.464 mL⁻¹ [0.388–0.608] vs. 0.555 mL⁻¹ [0.488–0.640]; p = 0.007812, Holm‐adjusted p = 0.007812) and FWHM decreased (5.92 Hz [4.93–7.52] vs. 4.68 Hz [3.45–5.92]; p = 0.0004883, Holm p = 0.001953). Paired t‐tests showed the same direction of effects and were also significant for all comparisons (TE = 35 volume‐normalized SNR p = 0.0497; TE = 35 FWHM p = 2.33 × 10⁻⁴; TE = 135 volume‐normalized SNR p = 0.00988; TE = 135 FWHM p = 3.41 × 10⁻⁴), consistent with the primary nonparametric analysis. A corresponding analysis considering the volume‐normalized SNR and FWHM valued estimated from the tCho peak is included in the Supporting Information in Section 1.3 and in Figure S3.

FIGURE 3.

Spectral registration improves LCModel spectral quality metrics. Paired dot plots compare uncorrected and corrected single‐voxel spectra across subjects at TE = 35 ms and TE = 135 ms, with median and interquartile range (IQR) overlaid (black). (A) Volume‐normalized signal‐to‐noise ratio (SNR) increases after correction at both echo times. (B) Full width at half maximum (FWHM) decreases after correction at both echo times. Brackets denote Holm–Bonferroni–adjusted paired Wilcoxon signed‐rank test significance (*p < 0.05; **p < 0.01; ***p < 0.001).

3.1.2. Cramér‐Rao Lower Bounds

CRLB analysis was restricted to tNAA, tCr, tCho, Glx, mI, and Lac, and threshold‐based results are summarized in Figure 4. Individual‐subject CRLBs are also reported in Table S1. At TE = 35 ms, pooled paired analysis across all 84 metabolite‐subject estimates showed a significant reduction in CRLB following spectral registration (Wilcoxon signed‐rank, p = 0.023). Consistent with this, the number of estimates meeting CRLB thresholds increased after correction, from 66/84 (78.6%) to 68/84 (81.0%) for CRLB ≤ 15% (Δ = +2, p = 0.500), from 71/84 (84.5%) to 74/84 (88.1%) for CRLB ≤ 30% (Δ = +3, p = 0.250), and from 72/84 (85.7%) to 80/84 (95.2%) for CRLB ≤ 50% (Δ = +8, p = 0.0078; Figure 4A). Although metabolite‐wise paired comparisons were not significant after false discovery rate correction, the largest descriptive improvement at TE = 35 ms was observed for Lac, with smaller gains for tNAA, tCho, and Glx.

FIGURE 4.

Threshold‐based summary of CRLB performance for tNAA, tCr, tCho, Glx, mI, and Lac before and after spectral registration. (A) TE = 35 ms. (B) TE = 135 ms. Bars show the percentage of metabolite‐subject estimates with CRLB ≤ 15%, ≤ 30%, and ≤ 50% for uncorrected and corrected spectra. Counts are shown above bars. Brackets and asterisks indicate significant paired differences by McNemar's exact test (*p < 0.05, **p < 0.01, ***p < 0.001).

At TE = 135 ms, pooled paired analysis showed no significant overall difference in CRLB between corrected and uncorrected spectra (p = 0.105). Consistent with this, only small changes were observed in the number of estimates meeting the CRLB thresholds after correction, from 52/84 (61.9%) to 53/84 (63.1%) for CRLB ≤ 15% (Δ = +1, p = 1.00), from 66/84 (78.6%) to 67/84 (79.8%) for CRLB ≤ 30% (Δ = +1, p = 1.00), and from 72/84 (85.7%) to 74/84 (88.1%) for CRLB ≤ 50% (Δ = +2, p = 0.688; Figure 4B). Metabolite‐specific changes at TE = 135 ms were small and heterogeneous, with the largest numerical improvement observed for mI. Together, these results indicate that spectral registration produced a modest but measurable improvement in quantitation reliability at TE = 35 ms, whereas corresponding improvements at TE = 135 ms were limited.

3.1.3. Metabolite Concentrations and Ratios

Comparisons between corrected and uncorrected spectra using the primary CRLB threshold of ≤ 30% are summarized in Table 1. For metabolite concentrations, no comparison remained significant after false‐discovery‐rate correction at either TE. At TE = 35 ms, tCho showed the largest nominal increase after correction, whereas at TE = 135 ms, Glx showed the largest nominal increase; however, neither remained significant after FDR adjustment. For metabolite ratios, only Glx/tCho at TE = 35 ms remained significant after FDR correction, with lower values after spectral registration than in the uncorrected data: 2.524 (2.224–2.673) versus 2.735 (2.398–2.929), corresponding to a median relative difference of −7.84% ($Z=-2.69$, $p=0.007$, $p_{\mathit{FDR}}=0.036$, $r=0.75$). No metabolite ratio at TE = 135 ms remained significant after FDR correction, although several comparisons showed nominal trends. No paired Lac observations at TE = 135 ms satisfied the CRLB inclusion criterion in both corrected and uncorrected spectra, precluding statistical testing for that metabolite at that TE. Overall, under the primary CRLB ≤ 30% criterion, spectral registration did not produce broad significant changes in metabolite concentrations or ratios, with the clearest robust effect being a reduction in Glx/tCho at TE = 35 ms. Results of the analysis of between‐subject CVs before and after corrections are summarized in Table S2.

TABLE 1.

Summary statistics comparing LCModel‐derived metabolite concentrations and metabolite ratios between spectra processed with and without spectral registration.

Outcome type	Measure	TE35 n	TE35 with SR	TE35 without SR	TE35 diff (%)	TE35 pFDR	TE35 r	TE135 n	TE135 with SR	TE135 without SR	TE135 diff (%)	TE135 pFDR	TE135 r
Metabolite concentrations (I.U.)	tNAA	13	0.874 (0.656–1.360)	0.994 (0.607–1.390)	1.49 (−12.07–4.99)	0.807	0.068	13	0.251 (0.226–0.377)	0.257 (0.238–0.375)	0.53 (−5.04–2.99)	0.861	0.048
	tCr	14	1.230 (0.954–1.530)	1.295 (1.001–2.402)	3.02 (−1.07–5.80)	0.588	0.319	14	0.381 (0.345–0.611)	0.387 (0.338–0.601)	1.41 (−2.62–4.12)	0.825	0.21
	tCho	13	0.721 (0.625–1.090)	0.693 (0.616–1.080)	1.46 (0.72–5.21)	0.115	0.649	14	0.317 (0.254–0.461)	0.335 (0.262–0.512)	1.58 (−2.37–8.48)	0.825	0.118
	Glx	13	1.990 (1.460–2.820)	2.320 (1.530–2.730)	−3.27 (−12.20–3.30)	0.588	0.291	11	0.737 (0.412–0.970)	0.672 (0.354–0.799)	10.52 (1.57–18.57)	0.104	0.697
	mI	14	2.830 (2.183–4.430)	3.440 (2.252–4.835)	0.17 (−2.27–2.58)	0.712	0.143	12	0.928 (0.547–1.287)	0.960 (0.648–1.770)	5.22 (−5.18–11.70)	0.825	0.136
	Lac	4	0.658 (0.518–0.845)	0.768 (0.534–0.976)	−1.80 (−7.84–1.30)	0.698	0.365	0	NA	NA	NA
Metabolite ratios to tCr	tNAA:tCr	13	0.709 (0.621–0.841)	0.749 (0.667–0.832)	−4.45 (−9.24–0.36)	0.098	0.572	13	0.678 (0.552–0.848)	0.796 (0.602–0.824)	−0.50 (−9.30–2.91)	0.866	0.126
	tCho:tCr	13	0.575 (0.562–0.668)	0.587 (0.56–0.661)	0.14 (−2.83–1.86)	0.7	0.107	14	0.795 (0.733–0.839)	0.791 (0.683–0.836)	−0.73 (−3.75–7.27)	0.925	0.025
	Glx:tCr	13	1.396 (1.251–1.798)	1.811 (1.382–1.963)	−10.19 (−13.50–0.72)	0.098	0.572	11	1.569 (1.107–1.612)	1.452 (0.896–1.598)	8.06 (6.91–17.74)	0.131	0.644
	mI:tCr	14	2.523 (1.884–2.82)	2.557 (1.871–2.873)	−1.18 (−2.70–1.70)	0.7	0.109	12	2.274 (1.941–2.703)	1.994 (1.781–3.062)	4.07 (−2.74–10.65)	0.776	0.249
	Lac:tCr	4	0.434 (0.387–0.531)	0.44 (0.387–0.605)	−4.46 (−13.02–1.73)	0.589	0.464	0	NA	NA	NA
Metabolite ratios to tNAA	tCr:tNAA	13	1.415 (1.19–1.613)	1.338 (1.199–1.505)	4.68 (−0.31–10.11)	0.16	0.533	13	1.476 (1.179–1.813)	1.257 (1.214–1.659)	0.74 (−2.90–10.98)	0.507	0.184
	tCho:tNAA	13	0.865 (0.772–0.957)	0.785 (0.742–0.936)	3.24 (−0.14–10.59)	0.16	0.514	13	1.155 (0.858–1.434)	1.009 (0.85–1.371)	4.24 (1.84–8.72)	0.073	0.533
	Glx:tNAA	13	2.151 (1.891–2.471)	2.231 (2.037–2.532)	−3.47 (−7.04–1.41)	0.166	0.455	11	2.017 (1.528–2.521)	1.703 (1.215–2.427)	14.65 (5.52–23.94)	0.073	0.67
	mI:tNAA	13	3.232 (2.877–3.43)	3.154 (2.559–3.657)	5.67 (−0.96–8.58)	0.166	0.417	11	3.87 (2.3–4.819)	3.706 (2.02–4.17)	8.19 (5.36–16.48)	0.073	0.59
	Lac:tNAA	4	0.754 (0.658–0.856)	0.713 (0.6–0.92)	5.57 (−4.52–8.98)	0.715	0.183	0	NA	NA	NA
Metabolite ratios to tCho	tNAA:tCho	13	1.156 (1.045–1.295)	1.274 (1.068–1.347)	−3.14 (−9.58–0.14)	0.137	0.533	13	0.866 (0.697–1.166)	0.991 (0.729–1.176)	−4.06 (−8.02–1.81)	0.149	0.494
	tCr:tCho	13	1.745 (1.495–1.778)	1.706 (1.513–1.784)	−0.17 (−2.23–2.88)	0.65	0.126	14	1.256 (1.192–1.365)	1.265 (1.196–1.466)	0.38 (−6.94–3.84)	0.826	0.059
	Glx:tCho	13	2.524 (2.224–2.673)	2.736 (2.398–2.929)	−7.84 (−10.71–1.82)	0.036	0.746	11	2.052 (1.516–2.318)	1.782 (1.256–2.218)	8.81 (2.71–20.85)	0.149	0.563
	mI:tCho	13	3.797 (3.465–4.064)	3.701 (3.41–4.253)	−1.42 (−3.20–1.74)	0.478	0.242	12	2.955 (2.612–3.03)	2.818 (2.575–3.245)	1.99 (−2.13–8.55)	0.363	0.317
	Lac:tCho	4	0.706 (0.633–0.819)	0.716 (0.611–0.951)	−4.20 (−11.35–0.08)	0.456	0.548	0	NA	NA	NA

Note: Results for TE = 35 ms and TE = 135 ms are presented side‐by‐side for comparison. Values are median (interquartile range). Analyses were restricted to tNAA, tCr, tCho, Glx, mI, and Lac, including only estimates with CRLB ≤ 30 in both corrected and uncorrected spectra; for ratio analyses, both numerator and denominator metabolites were required to satisfy the same CRLB criterion, and self‐ratios were excluded. Corrected and uncorrected values were compared using Pratt signed‐rank tests. Difference (%) denotes the paired relative difference with the uncorrected value as baseline. pFDR denotes the Benjamini–Hochberg false‐discovery‐rate‐adjusted p‐value computed separately within each TE and outcome family, and effect size is reported as r = |Z|/√n. NA indicates that no paired observations satisfied the inclusion criterion.

4. Discussion

This study shows that spectral registration improves the technical quality of in vivo fetal single‐voxel MRS and can partially compensate for phase and frequency inconsistencies across averages. The clearest effects were observed in the spectral quality metrics: volume‐normalized SNR increased and FWHM decreased at both TE = 35 ms and TE = 135 ms after correction, and the representative spectra showed sharper peak definition after averaging. These findings indicate that conventional averaging in fetal MRS is measurably degraded by inter‐average misalignment and that retrospective correction can recover spectral coherence.

The quantitative benefit of spectral registration was more apparent for reliability of fitting than for broad changes in metabolite estimates. At TE = 35 ms, pooled CRLB values were significantly lower after correction, and the proportion of estimates meeting the CRLB threshold increased. In contrast, CRLB improvements at TE = 135 ms were smaller and not significant. This pattern suggests that short‐TE fetal spectra, which contain greater spectral overlap and baseline complexity, may benefit more from improved alignment and linewidth reduction than long‐TE spectra.

Despite these gains in spectral quality, spectral registration did not produce widespread significant changes in metabolite concentrations (I.U.) or ratios after multiple‐comparison correction. Only Glx/tCho at TE = 35 ms remained significant after FDR correction. This likely reflects the fact that spectral registration primarily improved precision and fit stability rather than causing systematic shifts in central metabolite estimates. It may also reflect the modest sample size, the use of ratio‐based outcomes, and the conservative restriction of metabolite analyses to observations meeting CRLB ≤ 30% in both corrected and uncorrected spectra. In addition, the comparison in this study was not between raw uncorrected data and corrected data, but between spectra processed with eddy‐current correction alone and spectra processed with eddy‐current correction plus spectral registration and outlier rejection.

At the same time, spectral registration has inherent limitations. It cannot fully correct large bulk fetal motion that displaces the voxel from the intended anatomy, because retrospective phase and frequency alignment cannot restore spatial correspondence once signal originates from a different location. Spectral registration is therefore best viewed as a correction for inter‐average phase and frequency perturbations within an otherwise similar voxel rather than a complete solution to fetal motion. For this reason, it will likely be most effective when combined with motion‐aware acquisition strategies such as navigators, interleaved reference signals, prospective voxel steering, and dynamic B₀ and B₁ shimming adjustments [38, 39, 40, 41, 42].

Overall, these results support spectral registration as a useful and practical component of fetal MRS processing. Even in this modest cohort, it improved spectral coherence, narrowed linewidth, and enhanced quantitation reliability, with the strongest benefit observed at TE = 35 ms. These findings support incorporating spectral registration into fetal MRS pipelines and motivate further work in larger cohorts and in combination with prospective motion‐management approaches.

5. Conclusion

Spectral registration improved the quality of in vivo fetal single‐voxel MRS by increasing SNR, reducing linewidth, and improving spectral coherence relative to processing without spectral registration. It also modestly improved quantification reliability, particularly at TE = 35 ms, as reflected by lower pooled CRLB values and a greater proportion of estimates meeting CRLB thresholds after correction.

These findings support spectral registration as a practical addition to fetal MRS processing pipelines, especially in motion‐prone acquisitions where conventional averaging is vulnerable to degradation. Although it cannot correct large voxel‐displacing motion, spectral registration can improve the robustness and interpretability of fetal spectra and may be especially valuable when combined with prospective motion‐management strategies.

Funding

This work was supported in part by the UC Irvine Chancellor's Fellow program. This study was supported by the National Institutes of Health under award numbers R01HD109395 and R01EB031849.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Spectral‐registration frequency and phase offsets across individual averages. Representative transient‐wise frequency and zero‐order phase offsets estimated by spectral registration are shown across all acquired averages for spectra acquired at TE = 35 ms and TE = 135 ms. The upper row (A) shows the estimated offsets for a representative dataset at TE = 35 ms, and the lower row (B) shows the estimated offsets at TE = 135 ms. In each row, the left panel shows the frequency offsets ($\Delta f$) in Hz, and the right panel shows the phase offsets ($\Delta \phi$) in degrees. Accepted averages are shown as connected filled black circles, while averages identified as outliers are indicated by red crosses. These outlier averages were excluded from the final corrected spectral average.

Figure S2: B0 field drift estimated from spectral‐registration frequency offsets. Transient‐wise B0 drift is shown across all acquired averages for each dataset at TE = 35 ms and TE = 135 ms. The upper row (A) shows the TE = 35 ms data, and the lower row (B) shows the TE = 135 ms data. In the left panels, B0 drift curves are shown for all subject datasets, with accepted averages shown as connected filled circles and averages identified as outliers indicated by red crosses. In the right panels, linear drift trends are shown for each dataset after excluding outlier averages from the fitting. The B0 drift was estimated from the frequency offsets obtained during spectral registration and expressed in Hz. These plots summarize the magnitude and temporal pattern of field‐frequency instability across the acquired averages and echo times.

Figure S3: Spectral registration improves Cho‐based spectral quality metrics estimated directly from the spectra. Paired dot plots compare uncorrected and corrected single‐voxel spectra across subjects (n = 14) at TE = 35 ms and TE = 135 ms, with median and interquartile range (IQR) overlaid in black. (A) Volume‐normalized signal‐to‐noise ratio (SNR) estimated from the local baseline‐corrected total choline (tCho) resonance near 3.19 ppm increases after correction at both echo times. (B) Full width at half maximum (FWHM), estimated using a local Cr–Cho Lorentzian model fit to the 2.90–3.35 ppm region, decreases after correction at both echo times. Brackets denote Holm–Bonferroni–adjusted paired Wilcoxon signed‐rank test significance (*p < 0.05; **p < 0.01; ***p < 0.001).

Table S1: Individual‐subject LCModel Cramér–Rao lower bound (CRLB) values (%) for the reported metabolites tNAA, tCr, tCho, Glx, mI, and Lac at TE = 35 ms and TE = 135 ms across all 14 subjects. For each metabolite, values are shown separately for spectra processed with spectral registration (corrected, C) and without spectral registration (uncorrected, NC). The table is provided to allow metabolite‐specific and subject‐level assessment of baseline spectral quality and the effect of spectral registration on fitting uncertainty.

Table S2: Summary statistics comparing between‐subject coefficients of variation (CVs) for LCModel‐derived metabolite concentrations and metabolite ratios between spectra processed with and without spectral registration (SR). Results for TE = 35 ms and TE = 135 ms are presented side‐by‐side for comparison. Analyses were restricted to tNAA, tCr, tCho, Glx, mI, and Lac. For concentrations, only estimates with CRLB ≤ 30 in both corrected and uncorrected spectra were included; for ratio analyses, both numerator and denominator metabolites were required to satisfy the same CRLB criterion, and self‐ratios were excluded. Values for with SR and without SR are between‐subject CVs (%). Difference (%) denotes corrected minus uncorrected. pperm denotes the paired permutation p‐value, and pFDR denotes the Benjamini–Hochberg false‐discovery‐rate‐adjusted p‐value computed separately within each TE and outcome family. NA indicates that no paired observations satisfied the inclusion criterion.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.