Venous Oxygenation Measurements in the Upper Arm: Reproducibility and Sensitivity to a Physical Challenge

ABSTRACT

Purpose

The feasibility and reproducibility of adapting the T2‐Relaxation‐Under‐Spin‐Tagging (TRUST) MRI sequence for noninvasive measurement of venous oxygenation in the upper arm were investigated, aiming to enable future in vivo validation of T _2b ‐oxygenation calibration curves in sickle cell disease (SCD).

Methods

The TRUST sequence, originally developed for the brain, was optimized for the upper arm by comparing control, label and difference signal. Accuracy was assessed using a phantom experiment. Twelve healthy volunteers underwent repeated TRUST scans using two different protocols: Three‐averages and One‐average. Reproducibility and temporal resolution were investigated, and a handgrip exercise was used to assess sensitivity to physiological changes. Statistical analyses included coefficient of variation (CV), within‐subject CV (wsCV), and Bland–Altman analysis.

Results

Phantom experiments demonstrated accurate T ₂ estimation at low expected velocities with deviations from reference less than 16 ms (11.2%). In vivo, the adapted sequence provided faster and reproducible T _2b (86.0 ± 30.0 ms) and Y _v measurements (66.1% ± 11.7%). Reducing the number of averages improved temporal resolution without compromising measurement reproducibility (bias: 16.3 to 7.6 ms, wsCV: 39.8% to 35.3%). During exercise, significant decreases in T _2b and Y _v were observed in both the superficial and deep vein: mean T _2b decreased from 87.2 to 58.9 ms (p < 0.001) in the cephalic vein, and from 93.3 to 71.3 ms (p < 0.001) in the brachial vein.

Conclusion

The adapted TRUST protocol enables reproducible assessment of venous oxygenation in the upper arm, providing a technique for future validation and recalibration of T _2b –oxygenation curves in SCD.

1. Introduction

T ₂ relaxometry is a well‐established technique for noninvasively measuring venous blood oxygenation (Y _v ) in the brain [1, 2, 3]. Several methods exploit the change in magnetic susceptibility when oxyhemoglobin transitions to a deoxygenated state to derive blood T ₂ (T _2b ). One of these methods is the T2‐Relaxation‐Under‐Spin‐Tagging (TRUST) technique, which uses T ₂‐preparation pulses to induce T ₂‐weighting on blood, allowing an exponential fit of the signal curve [2]. This way, T _2b can be estimated and converted into blood oxygenation (Y _v ) using empirically derived calibration models. Such models have been derived from bovine blood [4], blood from healthy controls [5], and blood from patients with severe types of sickle cell disease (SCD) [6], namely the HbSS and HbSβ⁰ genotypes. Milder forms of the disease, HbSC and HbSβ⁺, having a less severe presentation of the disease with less damage to their red blood cells (RBCs), are lacking a specific calibration model [6].

To date, no consensus exists on which T _2b calibration should be used for patients with milder forms of SCD, leading to contradicting reports of oxygen‐derived metrics like oxygen extraction fraction (OEF), which reflects the fraction of oxygen extracted from the blood. Depending on the calibration model used, OEF has been reported as higher [7, 8], lower [9] or comparable [10] to that of healthy controls, putting the fidelity of the models into question. There is also uncertainty as to whether in vitro calibrations for SCD blood accurately reflect T ₂ calibrations in vivo, given the inherent fragility of sickle RBCs.

Validating these models is challenging, since blood samples cannot be directly obtained from the superior sagittal sinus (SSS), the intracranial vein where blood oxygenation is traditionally measured to derive the brain's OEF. Moreover, ex vivo experiments are complicated by blood aggregation and RBC damage during deoxygenation and temperature variations. T _2b measurements in the antecubital vein alongside contemporaneous CO‐oximetry of blood samples could offer an alternative in vivo validation. However, applying the TRUST sequence to the arm introduces technical challenges, including smaller vein size, lower Signal‐to‐Noise‐Ratio (SNR), smaller blood pool volumes to label, slower flow velocities and surrounding muscle and fat tissue around the vein of interest. Also, an accurate validation would require higher temporal resolution in TRUST measurements, especially during a physical challenge, which could provide additional data points across a broader range of blood oxygenation levels for comparison with blood samples.

To address these challenges, the present study aimed to adapt and optimize the TRUST sequence for use in the arm to noninvasively obtain venous T _2b in the superficial cephalic vein. This will enable future studies to compare TRUST measurements with CO‐oximetry and validate, while potentially recalibrating, the Y _v calibration curves for patients with SCD. Specifically, the objectives of this study were threefold: (1) to determine among the signal types inherent to the TRUST sequence (control, label or difference) which provides the most accurate and reliable T _2b estimation in the upper arm vasculature; (2) to test reproducibility; and (3) to evaluate sensitivity to physiological changes induced by a physical challenge.

2. Methods

The study performed on healthy volunteers was approved by the local medical ethics review committee, in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants.

To assess the feasibility of TRUST measurements in the arm, four experiments were performed. Experiment 1 optimized the sequence in vivo comparing both the SSS and the superficial cephalic vein; Experiment 2 assessed the accuracy of T ₂ estimation under controlled flow conditions in a phantom, comparing with a reference measurement; Experiment 3 evaluated the reproducibility of two acquisition schemes with different temporal resolutions; and Experiment 4 examined the sensitivity of the sequence to exercise‐induced physiological changes. These experiments included both phantom and in vivo measurements, as depicted in Figure 1. The duration of the protocol was around 30 min per scan session for in vivo and 5 min for the phantom measurements.

FIGURE 1.

Schematic of the scanning protocol. (A) In vivo measurements: Each volunteer was scanned twice, 24 h apart. Experiment 1 compared the different signal types; Experiment 3 evaluated reproducibility with the Three‐averages approach (NSA = 3) compared to the One‐average approach (NSA = 1). Experiment 4 examined sequence sensitivity with a handgrip physical challenge with the One‐average approach (NSA = 1). (B) Phantom measurements: Experiment 2 assessed T ₂ accuracy with TSE and One‐average TRUST (NSA = 1) scans without flow for reference, followed by scans with increasing flow $Q$.

All experiments were performed on a Philips 3 T MRI scanner (Ingenia Elition X, Philips Medical Systems, Best, Netherlands) using a 32‐channel head coil for the head and a 16‐channel small extremity coil for the arm. Measurements on the head were performed similarly as described by Lu and Ge [2]: field of view (FOV) = 220 × 220 mm², voxel size = 3.24 × 3.24 × 5 mm³, TR = 1978 ms, TI = 1022 ms, TE = 3.6 ms, flip angle (FA) = 90°, labeling thickness = 100 mm, labeling gap = 2.5 mm, Carr‐Purcell‐Meiboom‐Gill interpulse spacing (${\tau }_{\mathit{\text{CPMG}}}$) = 10 ms, effective echo times (eTEs) = 0, 40, 80, and 160 ms corresponding to 0, 4, 8, and 16 refocusing pulses in the T ₂‐preparation, transfer insensitive labeling technique (TILT) inversion pulse and number of signal averages = 1, whereas 3 averages were used in the original TRUST implementation by [2]. Positioning of the imaging slice of TRUST scans were matched and placed according to guidelines described by Liu et al. [1], approximately parallel to the anterior commissure–posterior commissure (AC–PC) plane and 20 mm above the sinus confluence.

Measurements on the right arm were performed with the hand in supine position. Foam padding and a weighted pillow were used to minimize movement. Other imaging parameters included eTEs = 0, 40, 80 and 160 ms, FOV = 100 × 100 mm², voxel size = 1.47 × 1.47 × 5 mm³, a single‐shot echo planar imaging (EPI) readout was used, TR = 3000 ms, TI = 1000 ms, labeling thickness = 100 mm, labeling gap = 2.5 mm. These parameters were empirically optimized in a pilot study (unpublished) to ensure robust labeling for the arm protocol for both the superficial cephalic vein and deep brachial vein, across the expected range of venous blood velocities. As part of our optimization, three and one signal average were used for measurements in the arm to ensure that the increased temporal resolution would not lead to poorer reproducibility. In contrast to the original TRUST implementation where TE was kept at its shortest value, here an in‐phase TE of 6.91 ms was used to minimize chemical shift artifacts from fat signal. If a fat artifact was present over the vein of interest, the FOV was rotated to avoid overlap. The labeling slab was positioned distal from the imaging slice.

Due to the increased tortuosity of veins in the arm compared to the SSS, an initial scout TRUST scan was performed approximately 10 cm distal to the axilla to localize suitable venous anatomy. Minor adjustments in slice position were made, typically within ±3 cm, to maximize the difference between the observed control and label signal, thereby maximizing the difference signal, while ensuring the slice intersected with a suitable vein segment. Despite the tortuous vasculature, imaging the arm in the transverse plane and positioning the slice perpendicular to the long axis of the arm ensured robust coverage of the veins. Additionally, the use of a larger labeling slab (100 mm) further minimized the risk of missing the labeled spins. Once an optimal position was identified, the selected slice location was kept consistent across all measurements for that participant.

For the phantom experiments, a Turbo Spin Echo (TSE) sequence was used as a reference to determine the T ₂ value. The following parameters were used: FOV = 100 × 100 mm², voxel size = 0.57 × 0.57 × 4 mm³, single‐shot EPI, FA = 90°, refocusing angle = 180°, TSE factor = 24, TE = 0, 26.7, 72.5, 118.3, and 164.1 ms, TR = 8000 ms, and repetitions = 1.

In all in vivo experiments, both control and label images were acquired, in line with traditional TRUST imaging, to allow the conventional difference signal to be obtained. However, only the control image was used for the analyses, unless stated otherwise.

2.1. Experiment 1: Sequence Optimization

Experiment 1 aimed to optimize the TRUST sequence for robust T _2b and Y _v estimation in the upper arm. In the original TRUST implementation, control and label images were subtracted and the average difference signal across four voxels was calculated. By doing this subtraction, partial volume effects (PVE) were minimized. In this experiment, all three signal types, that is, control, label and the computed difference image, were systematically compared for T _2b estimation in both brain (SSS) and superficial cephalic vein in the upper arm. The superficial cephalic vein data analyzed here are the same as those used in Experiment 3, but are presented separately for a systematic comparison of the different signal types for T _2b estimation.

2.2. Experiment 2: Accuracy

A phantom experiment was conducted to verify the accuracy of T ₂ estimation with the control‐only image under well‐defined conditions. Measurements were performed using a [1:5] solution of pineapple juice and water, resulting in a T ₂ value within the range expected for upper arm venous blood (≈160 ms). The solution was pumped from outside the bore using an infusion pump and circulated through a CareFusion PVC stopcock extension line set (internal diameter 2 mm) at selected constant flows (0.1, 0.2, 0.3, 0.4, and 0.5 mL/s) in random order. The tube itself was also submerged in pineapple juice. Reference T ₂ values were obtained without flow using the multi‐echo TSE described earlier. The mean velocity was calculated by dividing the flow rate of the infusion pump by the cross‐sectional area of the line. For the TRUST measurement, only a single average per eTE was acquired, since high SNR and absence of physiological motion allowed robust parameter estimation without additional averaging.

2.3. Experiment 3: Reproducibility and Repeatability

To assess the precision of the TRUST measurements in the arm in vivo, a group of healthy volunteers (N = 12, 7 females, age 28.5 ± 2.2 years) were scanned twice, at least 24 h apart. Each session included two series of measurements over 5 min, each with different temporal resolutions to assess reproducibility. The first series acquired TRUST data using three averages per eTE (NSA = 3), in line with the original TRUST implementation, producing a T _2b estimation every 72 s and resulting in 4 measurements. The second series used a single acquisition per eTE (NSA = 1), producing a T _2b estimation every 24 s and resulting in 12 measurements.

2.4. Experiment 4: T2 Response to Handgrip Exercise

Finally, to assess sensitivity of TRUST in the arm to physiological changes, participants underwent a physical challenge during repeated TRUST measurements. A single acquisition per eTE (NSA = 1) was used, resulting in a temporal resolution of 24 s. T _2b was continuously measured for 6 min in total: 2 min at baseline, 2 min during a handgrip exercise at a constant load of 50 N (Current Designs Inc. PA, USA) and 2 min post‐exercise. As a reference, the Maximal Voluntary Contraction (MVC) was separately determined after both MRI measurements, using a digital handgrip dynamometer (T.K.K. 5401, Grip‐D Takei, Tokyo, Japan) and two consecutive attempts.

2.5. Data Analysis

Data was processed using in‐house developed MATLAB 2022b (The Mathworks Inc., Natick, MA, USA) software. The processing pipeline was similar to the analysis by Lu and Ge [2]. First, where applicable, control and label images were subtracted for each eTE to generate the difference images. To minimize user bias, a region of interest (ROI) was drawn around the vessel of interest. From there, on the mean of all second eTEs the four highest connecting voxels were selected as a preliminary mask. Additionally, if motion was visible—particularly relevant during the challenge experiments—this step was replaced by manually drawing that mask surrounding the vessel of interest every four eTEs. Within this preliminary mask, the three connected voxels with the highest signal were automatically selected for each in vivo image. This number of voxels was chosen because we found that all subjects had at least three superficial cephalic vein voxels visible (at the resolution of 1.47 × 1.47 × 5 mm³). For the phantom scans, a single voxel with the highest signal was selected due to the smaller cross‐sectional area. The average diameter of the superficial cephalic vein in healthy controls is 2.7 ± 1.6 mm [11] while the SSS averages 7.3 ± 0.9 mm [12]. To avoid increased PVE in the superficial arm veins, this decrease in diameter was approximately matched by our imaging resolution (1.47 × 1.47 × 5 mm³ for the arm vs. 3.24 × 3.24 × 5 mm³ for the SSS).

Afterwards, the spatially averaged signals from the selected voxels were fitted to an exponential decay model as a function of eTE from which T _2b can be derived. Specifically, the blood signal in the control scan is given by:

$$S_{\mathit{\text{control}}}=S_0·e^{-\frac{\mathit{eTE}}{T2b}}$$ (1)

where $S_0$ is the equilibrium magnetization of blood. For the traditional TRUST implementation using the inversion pulse, an adjustment for blood T ₁ (T _1b ) is required to describe the signal of the label scan:

$$S_{\mathit{\text{label}}}=S_0·\left(1-2·e^{-\frac{\mathit{TI}-\mathit{eTE}}{T1b}}\right)·e^{-\frac{\mathit{eTE}}{T2b}}$$ (2)

in which TI is the inversion time, and T _1b is assumed to be 1624 ms [13]. Subsequently, the difference signal can be described as follows:

$$S_{\mathit{\text{difference}}}=S_{\mathit{\text{control}}}-S_{\mathit{\text{label}}}={S_0}^{\prime }·e^{-\mathit{eTE}·\left(\frac{1}{T1b}-\frac{1}{T2b}\right)}$$ (3)

where ${S_0}^{\prime }=2·e^{\frac{\mathit{TI}}{T_{1b}}}$.

Conversion from T _2b to Y _v was performed using the empirically defined model by Bush and colleagues [5]:

$$\frac{1}{T_{2_b}}=77.5·\mathrm{Hct}·{\left(1-Y_v\right)}^2+27.8·{\left(1-Y_v\right)}^2+6.95·\mathrm{Hct}+2.34$$ (4)

Hematocrit (Hct) was assumed to be 46% for men and 42% for women [14]. Quality control included computing the 95% confidence interval for T _2b (CI(T _2b )) from the fitting procedure using the nlparci MATLAB function, along with the coefficient of determination R ². Values with R ² < 0.8 were excluded. The lowest Hct level reported by the literature for healthy controls is ≈36% [15, 16]. Therefore, we considered the T _2b value corresponding to 100% Y _v at this Hct as a physiological upper limit and excluded values above 206.5 ms as not physiologically reasonable. The arm features two major venous return pathways: the superficial cephalic vein and the deep brachial vein [17]. Control‐only measurements have the advantage of being independent of blood velocities, making it possible to obtain oxygenation values for both veins in a single acquisition. For simplicity, only data from the superficial vein are presented in Experiment 1 and 3. In Experiment 4, data from both superficial and deep veins are shown to provide additional insight into blood oxygenation redistribution during the physical challenge.

For completeness and to provide a more intuitive interpretation, all analyses were also performed for Y _v and are reported in the Supporting Information.

2.6. Statistical Analysis

Statistical analyses were performed using RStudio (v4.4.1; RStudio Inc., Boston, MA, USA 2015). To quantify variability in T _2b and CI 95, the Coefficient of Variation (CV) was calculated for each signal type across sessions as:

$$\mathrm{CV}=\frac{\mathrm{SD}}{\mu }\times 100\%$$

where $\mathrm{SD}$ is the standard deviation and $\mu$ is the mean of the parameter. Normality was assessed using the Shapiro–Wilk test as well as visual inspection of quantile‐quantile plots. For repeated measures across volunteers and scan sessions, linear mixed‐effects models (LMMs) were applied to normally distributed data and generalized linear mixed‐effects models (GLMMs) were used otherwise. Volunteer was included as a random effect, with time point and scan session as fixed effects. Paired sample t‐tests or Wilcoxon signed‐rank tests were used both to compare control, label and difference T _2b values, and to assess changes between baseline and the end of the physical challenge. A significance level of 0.05 was applied. Where significant main effects were observed, post hoc pairwise comparisons were performed using Bonferroni correction.

To assess repeatability and reproducibility, Bland–Altman analysis was performed. For inter‐session reproducibility, the measurements of both sessions were first averaged per volunteer, after which the bias was calculated as the difference between the means so that each volunteer contributed only once to the analysis. For intra‐session repeatability, measurements from the first half of the session were paired with those from the second half (i.e., 1–6 with 7–12 for the One‐average approach, and 1 and 2 with 3 and 4 for the Three‐averages approach). These were averaged within each volunteer to yield one value per subject, and the difference between paired values was used to calculate the bias, which was assessed for statistical significance using a paired t‐test.

Variability was also assessed using the within‐subject coefficient of variation (wsCV) as follows:

$$\text{wsCV}=\frac{{\mathrm{SD}}_{\mathrm{\Delta }}}{\mu }\times 100\%$$

where ${\mathrm{SD}}_{\mathrm{\Delta }}$ is the standard deviation of the differences between the repeated measurements and $\mu$ is the mean of all repeated measurements.

To assess concordance, Lin's Concordance Correlation Coefficient (CCC) was used [18].

3. Results

3.1. Experiment 1: Sequence Optimization

Figure 2 shows representative TRUST control and label scans, and their resulting difference image in the SSS (Figure 2A) and in the cephalic vein (Figure 2B,C). For the SSS and the first cephalic vein acquisitions, the estimated T _2b and Y _v were in agreement across all signal types, whereas that was not the case for the second cephalic vein acquisition. This highlighted the potential unreliability of the difference signal in T _2b measurements of the superficial cephalic vein.

FIGURE 2.

Representative images of TRUST in the SSS and cephalic vein, as well as a comparison of TRUST‐derived T _2b estimates using control, label, and difference images with their respective normalized signal decay. (A) TRUST scan in the SSS showing similar T _2b estimates from control, label, and difference images; (B) TRUST scan of the cephalic vein with similar T _2b values from the control, label, and difference images; (C) TRUST scan of the cephalic vein where T _2b from the difference and label image deviate from the control image. In all images, the target vein is highlighted with a red circle.

Figure 3 shows boxplots across all volunteers of the One‐average approach for each signal type—control, label and difference—per scan session. Four metrics are shown: session‐averaged T _2b in (A) and CV of T _2b in (B), as well as the session‐averaged of CI(T _2b ) in (C) and CV of CI(T _2b ) in (D). Session‐averaged T _2b values did not differ significantly between signal types (Figure 3A). However, the CV across T _2b values from the difference signal was significantly higher compared to the control‐derived and label‐derived T _2b values (Figure 3B). Significant differences were observed for the control‐difference comparison in both sessions for CIs of the T _2b estimates (Figure 3C). The variation in CIs of the T _2b estimates was significant for the control‐difference pair in the first session but not in the second session (Figure 3D). Detailed statistics are reported in Table S1.

FIGURE 3.

Comparison of T _2b metrics across control, label, and difference signals for both scan sessions using the One‐average approach. Boxplots of (A) mean of T _2b value $\overline{T_{2b}}$; (B) CV of T _2b $\mathrm{CV}\left(T_{2b}\right)$; (C) mean of CI 95 T _2b $\overline{\mathrm{CI}\left(T_{2b}\right)}$; (D) CV of CI 95 T _2b $\mathrm{CV}\left(\mathrm{CI}\left(T_{2b}\right)\right)$. All metrics are estimated from the control, label, and difference signals separated by sessions 1 and 2 using the One‐average approach. Gray dots indicate individual estimates per volunteer. Groups were tested for normality, and the appropriate statistical test was applied. Asterisks denote statistically significant differences (n.s. = non‐significance, p < 0.05*, p < 0.01**, p < 0.001***).

Figure 4A shows the Bland–Altman plots comparing the T _2b measurements derived from control and label images. T _2b was unbiased with a mean difference of −7.3 ± 19.6 ms (p = 0.3). Figure 4B shows the scatter plot and concordance between these measurements, with CCC = 0.75 (95% CI [0.26, 0.93]). The Pearson correlation coefficient was r = 0.77, with a bias correction factor of Cb = 0.96, indicating that the limited concordance was driven by random scatter rather than systematic bias. Metrics from these analyses, as well as agreement between the different signal types are summarized in Table 1. No significant bias was observed between control‐derived and label‐derived T _2b values. When comparing both control‐derived and label‐derived to difference‐derived T _2b values, only 72% of difference‐derived estimations were retained, reflecting reduced reliability.

FIGURE 4.

Comparison between control and label‐derived T _2b estimates from the first scan sessions. (A) Bland–Altman plot comparing control‐ and label‐derived T _2b estimates. (B) Scatter plot showing the correlation between control and label‐derived T _2b estimates, with the identity line (solid black) and the linear regression fit (dotted red). Black dots indicate individual volunteer T _2b estimates. Lin's Concordance Correlation Coefficient (CCC) and its 95% confidence interval (95% CI) quantifying agreement between methods are also reported.

TABLE 1.

Summary of Bland–Altman and correlation metrics for all pairwise comparisons of T _2b estimates derived from control, label, and difference images.

	Bland–Altman plot			Correlation plot
Comparison	Bias (ms)	95% LoA (ms)	p	CCC	95% CI	Pearson's r	Cb	Points retained (%)
Control vs. label	−7.3	±19.6	0.3	0.75	[0.26, 0.93]	0.77	0.96	100
Control vs. diff	8.8	±29.8	0.4	0.65	[0.039, 0.91]	0.73	0.83	72
Label vs. diff	16.7	±49.1	0.3	0.095	[−0.57, 0.68]	0.12	0.79	72

Note: Reported values include bias, limits of agreement (LoA), and p value from the Bland–Altman analysis as well as Lin's Concordance Correlation Coefficient (CCC), its 95% confidence interval (95% CI), the Pearson correlation coefficient (Pearson's r) and a bias correction factor (Cb). Points retained refers to the amount of data points that were retained after applying the exclusion criteria.

Based on these findings, we opted to proceed by using only the control‐derived T _2b estimates in subsequent analyses.

3.2. Experiment 2: Accuracy

Results of the phantom experiment are depicted in Figure 5 and demonstrate the accuracy of the TRUST‐derived T _2b estimates under controlled conditions. Up to the expected blood velocity in the upper veins (7.2 ± 0.7 cm/s [19]), the deviation from the reference T ₂ value ranged from 1.4% to 11.2%. At higher velocities, the error in T _2b estimation increased to a maximum of 43.6%. A linear regression analysis demonstrated a significant negative relationship between velocity of the fluid and T ₂ (slope = −4.6 × 10³ ms² cm⁻¹, p < 0.001, R ² = 0.95), indicating that T ₂ decreases with increasing flow in the phantom.

FIGURE 5.

Schematic overview and results of the phantom experiment. (A) Phantom schematic of axial and (B) sagittal view with flow direction (black arrow). The line (lighter yellow) was submerged in pineapple juice (darker yellow) with the respective imaging plane (gray box); (C) TRUST axial view scan of the phantom; (D) T ₂ value estimates plotted against flow velocities. Horizontal gray dashed‐dot line indicates reference T ₂ values at 0 cm/s (measured with TSE) and vertical red dashed line indicates the expected blood velocity of the superficial veins at 7.2 ± 0.7 cm/s measured by Ooue et al. [19].

3.3. Experiment 3: Reproducibility and Repeatability

Following the exclusion of four scans—one for absence of a visible vein, one for perceived numbness and two for insufficient signal—20 scans of the cephalic vein were analyzed across both series of measurements. Figure 6 presents all measured T _2b and Y _v values, including session means. For the Three‐averages approach, the mean T _2b was 83.5 ± 32.8 ms (Figure 6A) and the mean Y _v was 64.9% ± 12.5% (Figure 6D) and LMM analysis revealed no significant effects for time points (T _2b : p = 0.2; Y _v : p = 0.5), sessions (T _2b : p = 0.1; Y _v : p = 0.07) or their interaction (T _2b : p = 1.0; Y _v : p = 0.8). For the One‐average approach, the mean T _2b was 86.0 ± 30.0 ms (Figure 6B) and the mean Y _v was 66.1% ± 11.7% (Figure 6E). Here, LMM analysis showed no significant effects for time points (T _2b : p = 0.2; Y _v : p = 0.5) or their interaction with session (T _2b : p = 0.5; Y _v : p = 0.6) but did show a significant effect across sessions with higher values in session 2 (T _2b : p = 0.02; Y _v : p = 0.03).

FIGURE 6.

Variability of T _2b and Y _v estimates during 5 min of measurements. Left column: Three‐averages approach resulting in (A) T _2b and (D) Y _v values every 72 s. Central column: One‐average approach resulting in (B) T _2b and (E) Y _v values every 24 s. Right column: One‐average approach resulting in (C) T _2b and (F) Y _v values but after block averaging, matching the temporal resolution of the original TRUST protocol (72 s)—Block average (i.e., the mean of every three consecutive measurements: 1–3, 4–6, 7–9, etc.). Gray lines connect individual data points from the same subject; thick lines and shaded areas indicate the session mean ± SD. Red squares mark excluded points.

To match the temporal resolution of the original TRUST protocol, block averaging was applied to the One‐average data, that is, the mean of every three consecutive measurements was calculated (e.g., measurements 1–3, 4–6, 7–9, etc.). After block averaging, the mean values were 86.4 ± 29.8 ms for T _2b (Figure 6C) and 66.2% ± 11.7% for Y _v (Figure 6F). No significant effects were observed for time points (T _2b : p = 0.7; Y _v : p = 0.9), sessions (T _2b : p = 0.2; Y _v : p = 0.2) or their interaction (T _2b : p = 0.6; Y _v : p = 0.6).

Figure 7 summarizes the bias for both intra‐ and inter‐session repeatability and reproducibility. For inter‐session reproducibility (Figure 7A), neither the Three‐average (16.3 ± 34.0 ms, p = 0.2) nor the One‐average approach (Figure 7B) exhibited a bias (7.6 ± 30.8 ms, p = 0.5); wsCV was also comparable at 39.8% and 35.3%, respectively. For intra‐session repeatability, the Three‐averages approach (Figure 7C) resulted in a bias of −2.6 ± 10.4 ms (p = 0.3) and a wsCV of 13.3%, whereas the One‐average approach (Figure 7D) had a bias of 2.3 ± 6.4 ms (p = 0.3) and a wsCV of 7.6%. Across all Bland–Altman plots, linear regression of the differences on the corresponding mean confirmed that no significant proportional bias was present.

FIGURE 7.

Bland–Altman plots for the T _2b estimates. On the left: (A) inter‐session and (C) intra‐session reproducibility of the Three‐averages approach. On the right: (B) inter‐session and (D) intra‐session reproducibility of the One‐average approach. Black dots indicate individual volunteer T _2b estimates.

Comparing inter‐ and intra‐session results, both bias and wsCV were lower for intra‐session repeatability than for inter‐session reproducibility, regardless of the averaging approach. Additionally, the One‐average approach generally resulted in lower bias and wsCV compared to the Three‐averages approach, particularly for intra‐session repeatability, although these differences did not reach statistical significance.

3.4. Experiment 4: T2 Response to Handgrip Exercise

The optimized TRUST sequence detected significant physiological changes in Y _v during the physical challenge. Mean maximum grip strength was 40.0 ± 13.3 kg, with the challenge of 50 N corresponding to an MVC of 13.9% ± 4.2%. Figure 8 shows Y _v values during the physical challenge. For the superficial cephalic vein, five scans were excluded—one for excessive motion, one for absence of visible vein, one for water–fat shift artifact over the vein of interest, and two for insufficient signal—resulting in 19 scans included in the analysis. For the deep brachial vein, three scans were excluded—two for water–fat shift artifacts and one due to excessive data loss after quality control, leaving 21 scans.

FIGURE 8.

Venous oxygenation Y _v over time during the physical challenge. The start and end of the physical challenge are marked out by the blue dashed‐dot lines. Top row: Superficial cephalic vein with (A) increased temporal resolution (24 s); (B) temporal resolution similar to the original TRUST (72 s); and (C) normalized data (values divided by the mean of the first five baseline points). Bottom row: Deep brachial vein with (D) increased temporal resolution (24 s); (E) 72 s resolution; and (F) normalized data. Gray lines connect individual data points from the same subject; thick lines and shaded areas indicate the session mean ± SD. Red squares mark excluded points, which may include superimposed points. On average, all protocols show a drop in oxygenation during the physical challenge, followed by a recovery stage.

For the superficial cephalic vein, Figure 8A–C illustrates the Y _v time course at increased temporal resolution (24 s), conventional temporal resolution (72 s), and values normalized to each individual's baseline (average of the first five time points), respectively. Figure 8D–F shows the corresponding data for the deep brachial vein. Individual curves are provided in Figure S4.

Across all included sessions at increased temporal resolution, mean T _2b in the superficial vein decreased from 87.2 ± 34.7 to 58.9 ± 24.9 ms during exercise (p < 0.001), with a corresponding Y _v decrease from 65.7% ± 13.7% to 53.5% ± 13.8% (p < 0.001). For the deep vein, T _2b decreased from 93.3 ± 30.5 to 71.3 ± 31.3 ms during exercise (p < 0.001), with a corresponding Y _v decrease from 68.5% ± 11.3% to 59.2% ± 14.4% (p < 0.001). However, a subset of participants exhibited an increase in Y _v in either the superficial or deep vein.

4. Discussion

In this study, the TRUST sequence was adapted for noninvasive measurements of venous oxygenation in veins in the upper arm and its feasibility was investigated. Specifically, we omitted the labeling part of the original TRUST sequence to overcome instability in the difference signal observed in superficial upper arm veins. Phantom experiments demonstrated that the modified TRUST sequence still provides accurate T _2b measurements at low flow velocities. In vivo, a higher temporal resolution compared to the TRUST in the SSS was demonstrated, enabling dynamic, time‐resolved measurements of T _2b . We demonstrated good reproducibility both intra‐ and inter‐session, as well as sensitivity to oxygenation changes during a physical challenge.

We found that T _2b estimates obtained from difference images showed the highest variability, whereas both control and label images provided more reliable results. Estimating T _2b from the control image over the label is less complex, enabling an extra two‐fold temporal resolution increase and imaging veins with different velocities. This is in contrast with earlier findings in the sagittal sinus of the brain, where such instabilities in the difference signal estimations were not observed [2]. This discrepancy can be explained by several factors. The velocity of labeled blood in the arm is approximately three times lower than that of blood in the SSS (7.2 ± 0.7 cm/s [19] vs. 20.0 ± 3.0 cm/s [20]) and changes direction regularly due to venous valves [21]. Additionally, veins in the upper arm are more tortuous [22]. Finally, imaging conditions are further challenged by both scanning off‐isocenter and field inhomogeneities due to susceptibility effects of surrounding tissues such as muscle, fat, and air interfaces. The extent to which these factors contribute to the observed instability in the difference signal was not directly studied here and warrants further investigation.

Phantom data showed that there is a good agreement between the control‐signal‐based T ₂ estimates and the reference standard at lower velocities. The relationship between higher velocities and T ₂ inaccuracies confirmed earlier findings attributed to B ₀‐inhomogeneities [23]. An effect of PVE is less likely as the tube wall will have a very short T ₂, making its contribution negligible to the signal. Furthermore, we confine our analysis to the highest signal voxel, which suffers the least from PVE. It should be noted that the B ₀‐inhomogeneities in the phantom setup are likely more extreme than those encountered in vivo, making the phantom a worst‐case scenario—phantom tube's material is PVC, which creates additional non‐homogeneous magnetic field. Still, we obtained accurate T ₂ measurements in the expected velocity range of superficial veins, supporting the validity of using the control image for T _2b estimation in vivo. Based on these findings, a flow‐induced bias in T _2b measurements in vivo is thus not to be expected.

Our reproducibility experiments demonstrated that the number of signal averages could safely be reduced from three to one per eTE, without compromising robustness with respect to bias and reproducibility compared to the currently accepted Three‐averages TRUST protocol. Bland–Altman analysis and wsCV calculations further support the reliability of the One‐average approach. Repeatability within sessions was high, while inter‐session reproducibility was worse, reflecting both greater physiological variability across days and the technical challenges mentioned above. Notably, the One‐average approach showed comparable repeatability to the Three‐averages protocol, while enabling a three‐fold increase in temporal resolution. This increase is advantageous for tracking dynamic changes during physiological challenges and for synchronizing with blood sampling or external measurements. It is important to notice that, given the high sensitivity of the T _2b –Y _v calibration curve in the physiological expected range, observed changes in T _2b result in minor variations in Y _v .

Our intra‐session mean Y _v and standard deviation of 66.1% ± 11.7% observed in our study are higher than the earlier reported 61.7% ± 4.6% in the brain [24]. These higher absolute values likely reflect differences in vascular regulation: cerebral perfusion is tightly coupled to energetic demand, whereas flow regulation in the arm is looser, allowing anaerobic activity.

The sensitivity study presented here is, to our knowledge, the first to measure T _2b values (and oxygenation Y _v values) in both superficial and deep veins before, during, and after exercise. This potentially provides new insights to understand venous return pathways and enables more accurate comparisons with blood samples.

During the challenge, a group‐level drop in oxygenation was observed in both veins, consistent with expected physiological responses and prior observations [25]. Compared to recent work by Skow and colleagues [26], who reported a larger oxygenation drop using a different MRI approach and exercise protocol, our results likely reflect differences in exercise intensity and methodology.

At the individual level, responses varied: while both veins showed on average the expected decrease in oxygenation, in some participants oxygenation increased or decreased in either vein. This individual variability may reflect differences in physiological adaptation as well as the distinct physiological roles of superficial and deep veins in venous return and response to exercise. Our results are in line with previous studies showing blood flow changes during thermoregulatory [27] and exercise challenges [28]. While T _2b changes during exercise are assumed to reflect alterations in venous oxygenation, it is important to note that vessel dilation or restriction could also influence the measurement by altering PVE. However, the reduction in T _2b would require an unphysiological reduction in vessel diameter to be solely explained by vasoconstriction. While exercise load was not individualized, these results demonstrate the potential of TRUST for monitoring dynamic changes in venous oxygenation. Future studies including direct flow measurements and venous compliance assessments are warranted to further clarify these mechanisms.

4.1. Limitations

In Experiment 3, the order of the series of measurements, Three‐averages and One‐average, was not randomized, which may have introduced an effect of time, temperature, or other session‐related factors. These uncontrolled variables could explain the significant differences observed across sessions. Additionally, the control‐only approach may reintroduce magnetization transfer (MT) effects, absent in standard TRUST. Phase contrast or venogram scans could have improved PVE, vein visualization, and flow assessment. Although PVEs were mitigated by selecting the highest‐signal voxels within the vein, residual influence from surrounding tissue cannot be completely excluded.

In Experiment 4, exercise load was not individualized, and ambient temperature was not measured, which could explain differences between participants. One scan had to be excluded due to excessive motion. This is related to the nature of the handgrip challenge as recruitment of the biceps muscle can displace the vein of interest, whereas using primarily muscles of the forearm would minimize such movement. While co‐registration could further mitigate this effect, no reliable algorithm was found for upper arm imaging.

5. Conclusion

This study demonstrates the feasibility and reproducibility of an adapted TRUST MRI sequence for noninvasive, continuous monitoring of venous oxygenation in both superficial and deep veins of the upper arm. By optimizing the protocol to a higher temporal resolution and employing a control‐only approach, we achieved sensitive detection of physiological changes in venous oxygenation during exercise while maintaining reproducibility comparable to current TRUST protocols. The ability to assess both veins simultaneously and potentially disentangle respective physiological functions, especially during exercise, provides a valuable tool for future research. Additionally, using a control‐only approach also permits an extra two‐fold increase of temporal resolution. We consider these advantages a reasonable trade‐off, as they may come at the expense of some SNR or increased PVE but enable direct comparison of oxygenation dynamics between the two venous compartments under identical physiological conditions. Beyond hemoglobinopathies, this technique may provide insight into disorders characterized by impaired oxygen delivery or extraction, such as chronic venous insufficiency or varicose veins. Evaluating drug interventions in older anemic patients or decision making in blood removal with polycythemia vera are other possible research applications. In addition, dynamic assessment of venous desaturation and post‐exercise reoxygenation kinetics may offer a noninvasive means to probe muscle oxidative capacity, potentially useful in studies related to muscle functioning. Most importantly, this work establishes a methodological foundation in future studies for in vivo validation and potential recalibration of T _2b ‐oxygenation conversion curves in clinical populations, particularly individuals with SCD. By enabling direct comparison of TRUST‐derived T _2b values with contemporaneous CO‐oximetry blood samples from the arm, this approach provides a unique opportunity to improve the accuracy of noninvasive oxygenation assessment in SCD.

Funding

The authors have nothing to report.

Conflicts of Interest

Matthan W. A. Caan is shareholder of Nico‐lab International Ltd. The other authors declare no conflicts of interest.

Supporting information

Table S1: P values corresponding to the statistical comparisons presented in Figure 3 for T _2b . Groups were tested for normality and the appropriate statistical test was applied. Bold denotes statistically significant differences (p < 0.05*, p < 0.01**, p < 0.001***).

Table S2: P values corresponding to the statistical comparisons presented in Figure S2 for Y _v . Groups were tested for normality and the appropriate statistical test was applied. Bold denotes statistically significant differences (p < 0.05*, p < 0.01**, p < 0.001***).

Table S3: Summary of Bland–Altman and correlation metrics for all pairwise comparisons of Y _v estimates derived from control, label, and difference images. Reported values include bias, limits of agreement (LoA), and p value from the Bland–Altman analysis as well as Lin's Concordance Correlation Coefficient (CCC), its 95% Confidence Interval (95% CI), the Pearson correlation coefficient r and a bias correction factor Cb. Points retained refers to the amount of data points that were retained after the exclusion criteria.

Figure S1: Comparison of Y _v metrics across control, label, and difference signals for both scan sessions using the One‐average approach. Boxplots of (A) mean of Y _v value $\overline{Y_v}$; (B) CV of Y _v $\mathrm{CV}\left(Y_v\right)$; (C) mean of CI 95 Y _v $\overline{\mathrm{CI}\left(Y_v\right)}$; (D) CV of CI 95 Y _v $\mathrm{CV}\left(\mathrm{CI}\left(Y_v\right)\right)$. All metrics are estimated from the control, label and difference signals separated by sessions 1 and 2 using the One‐average approach. Gray dots indicate individual volunteer parameter estimates. Groups were tested for normality, and the appropriate statistical test was applied. Asterisks denotes statistically significant differences (n.s.—non‐significance, p < 0.05*, p < 0.01**, p < 0.001***).

Figure S2: Error and agreement analysis between Control and Label‐derived Y _v estimates across both scan sessions. (A) Bland–Altman plot comparing Control‐ and Label‐derived Y _v estimates. Each point represents a measure. (B) Scatter plot showing the correlation between Control and Label‐derived Y _v estimates, with the identity line (solid black) and the linear regression fit (dotted red). Black dots indicate individual volunteer T _2b estimates.

Figure S3: Bland–Altman plots for the Y _v estimates. On the left: (A) inter‐session and (C) intra‐session reproducibility of the Three‐averages approach. On the right: (B) inter‐session and (D) intra‐session reproducibility of the One‐average approach. Black dots indicate individual volunteer Y _v estimates.

Figure S4: Individual venous oxygenation Y _v over time for Experiment 3: Sensitivity separated per volunteer for both veins. The start and end of the physical challenge are marked out by the blue dashed‐dot line. Left side: (A) T _2b and (B) Y _v for superficial cephalic vein. Right side: (C) T _2b and (D) Y _v for deep brachial vein. Red line indicates first scan session and blue line indicate second scan session. Shaded regions represent the CI 95 at each time point. Red squares mark excluded point.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.