Spatially and temporally resolved assessment of quantitative skeletal muscle perfusion with non-contrast MRI in patients with PAD and healthy controls

Introduction

Peripheral arterial disease (PAD) is a circulatory disorder caused by plaque accumulation in the peripheral arterial wall that reduces blood flow to distal limbs, leading to symptoms such as claudication, ischemic pain, and in severe cases, tissue loss. PAD affects approximately 8 to 12 million people in the US and 200 million individuals worldwide, with prevalence increasing with age, affecting up to 20% of individuals over the age of 60 years (1). The pathophysiology of PAD involves intricate interactions between atherosclerotic stenosis, microvascular dysfunction, and impaired tissue perfusion that collectively contribute to exercise intolerance, skeletal muscle atrophy, and increased cardiovascular mortality (2-5). Historically, PAD has been underdiagnosed using diagnostic modalities that have relatively low sensitivity and precision (5). This has been shown to contribute to delays in diagnosis and timely therapeutic interventions (6). Notably, PAD has traditionally been viewed as a large-vessel (macrovascular) atherosclerotic disease, but recent studies emphasize the microvascular component: endothelial dysfunction, capillary rarefaction, and sympathetic overactivity in muscle that can adversely impact PAD outcomes (5,7). Microvascular disease (MVD) is common among individuals with PAD and plays a synergistic role in predicting mortality and amputation, with risk of major amputation increasing up to a 22.7-fold with concomitant PAD, compared to 13.9-fold with isolated PAD (8,9).

Despite patients with PAD experiencing reduced macrovascular and microvascular blood flow to the lower extremities, microvascular dysfunction is not routinely assessed by current clinical testing modalities. The traditional gold standard for PAD diagnosis remains digital subtraction angiography (DSA), offering high-resolution visualization of the arterial anatomy. However, due to the invasive nature and cost of the procedure, DSA is generally reserved for patients being considered for revascularization (2). In routine clinical practice, the ankle-brachial index (ABI) serves as the primary diagnostic and risk stratification tool for microvascular assessment for its reasonable accuracy, relatively low cost, and noninvasive nature (10). Although widely used, ABI has important limitations, particularly in the presence of non-compressible arteries, which lead to inaccurate results in patients with important comorbidities such as diabetes mellitus (DM) and chronic kidney disease (CKD) (11). Alternative diagnostic modalities such as the duplex ultrasound (DUS), computed tomography angiography (CTA), positron emission tomography (PET), and magnetic resonance angiography (MRA) have also been implemented (8,12). While these advanced imaging techniques offer superior delineation of macrovascular anatomy, none of them can quantitatively assess tissue-level perfusion. Moreover, given the chronic and progressive nature of PAD, serial imaging is frequently required to evaluate therapeutic response and ensure the long-term patency of revascularization procedures (13). Therefore, a contrast-free imaging approach emphasizing quantitative and highly sensitive tissue-level perfusion, rather than solely focusing on macrovascular anatomy, has the potential to provide important clinically-relevant insights about PAD severity, risk of progression, and post-revascularization outcomes.

Arterial spin labeling (ASL) is a noninvasive magnetic resonance imaging (MRI) technique that provides quantitative skeletal muscle blood flow (SMBF) measurements without the use of exogenous contrast, potentially permitting serial measurement of tissue perfusion (14-17). Wu et al. first implemented a continuous ASL in patients with PAD to measure calf muscle perfusion post-ischemic reactive hyperemia induced by cuff occlusion, demonstrating correlations between time to peak hyperemic flow with disease severity (18). Pollak et al. applied a pulsed ASL (PASL) method in patients with PAD to measure peak exercise perfusion immediately after plantar flexion exercise, demonstrating reproducible peak perfusion measurements that can identify patients with PAD from normal controls (15). All of these studies avoided perfusion measurements during the exercise to minimize motion artifacts. To differentiate PAD from healthy controls (HCs), many studies employ cuff occlusion perfusion protocols, as exercise-based assessments are often limited by symptoms and particularly challenging in severe disease (16,18,19). This limitation can introduce substantial variability in calf perfusion measurements due to differences in achievable workload.

In this study, a flow-sensitive alternating inversion recovery (FAIR) type of ASL and deep learning enabled model was applied to measure dynamic SMBF distribution in the calf muscle of HCs and patients with PAD using exercise-induced hyperemic protocol. The purpose of this study is two-fold: (I) to examine whether dynamic SMBF time course patterns across entire study protocol (rest, isometric exercise, and recovery) can be used to reliably differentiate PAD from age matched HCs and to assess spatial heterogeneity of SMBF, with a moderate exercise strength; (II) to determine whether a deep-learning enabled SMBF quantification can improve temporal resolution. Collectively, these advances support the use of ASL MRI as a reliable tool for differentiating and assessing microvascular dysfunction in PAD, with potential implications for improving clinical evaluation and disease monitoring. We present this article in accordance with the STARD-AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0675/rc).

Methods

Study population

Ten age- and sex-matched healthy volunteers [age, 60±4 years; body mass index (BMI), 26.3±6.3 kg/m²] and ten patients with PAD (age, 66±9 years; BMI, 29.1±6 kg/m²) were recruited from the local community, Washington University School of Medicine in St. Louis, and Barnes-Jewish Hospital to assess calf muscle perfusion. Of the 20 participants recruited, one healthy volunteer was excluded due to improperly performing the exercise, and one patient with PAD was excluded due to severe motion artifacts. All healthy volunteers were nonsmokers, free of cardiovascular, metabolic, musculoskeletal diseases, and did not have a history of PAD or peripheral neuropathy. Patients with PAD had confirmed clinical diagnosis of PAD (ankle-brachial index, i.e., ABI <0.9, or significant peripheral arterial stenosis detected by computer tomography or magnetic resonance angiography), were nonsmokers, and had no history of diabetes. Although DM is a common comorbidity in patients with PAD, patients with DM were intentionally excluded from this study to minimize the confounding effects of diabetes-related microvascular dysfunction on SMBF. Furthermore, the impact of DM on calf SMBF has been previously investigated by our group, where significantly impaired perfusion was demonstrated in diabetic patients (20). Individuals were excluded if they were hemodynamically unstable, pregnant, claustrophobic, or have chronic limb threatening ischemia (CLTI), severe lower extremity infection, or contraindication to MRI scanning. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Washington University Human Research Protection Office (No. 201603075) and informed consent was obtained from all individual participants prior to the study.

Study protocol

Exercise protocol consisted of a 3-minute isometric plantar flexion with a foot on the pedal of an MR-compatible custom-built ergometer device (14). During exercise, participants were instructed to exert force to fully depress the pedal of the device connected to a pressurized pneumatic cylinder. To ensure consistent resistance across participants, the exercise intensity was adjusted to 25% of the participant-specific maximal voluntary isometric contraction (MVIC). MVIC was estimated prior to the imaging protocol based on the peak cylinder pressure the participant could fully compress. The exercise device was pressurized using compressed room air according to the patient’s MVIC, and the pressure was verified with a calibrated gauge prior to each imaging session.

Participants were positioned supine on the MRI table with the right foot (or, for patients with PAD, the most symptomatic foot) securely strapped onto the pedal of the custom-built ergometer. Velcro straps were fastened over the subject’s leg with the knee flexed to approximately 30º. All imaging was performed using a 3T Prisma Siemens whole-body MR system (Siemens Healthineers, Malvern, PA, USA) and a flexible surface coil for signal receiver wrapped around the calf.

We employed an ASL technique originally validated in the myocardium and more recently adapted for use in skeletal muscle, as described in detail in previous reports (14,21,22). Briefly, the acquisition sequence is an inversion recovery prepared single-shot gradient-echo sequence with slice-selective inversion (SS) and volume-selective inversion (VS) acquisitions (23). There were 8 single-shot acquisitions after each inversion pulse, resulting in a total of 16 acquisitions. There was a 4-sec delay after the first 8 acquisitions prior to the second inversion pulse to ensure completely recovered magnetization. The ASL sequence parameters for quantification of SMBF included: gradient-echo acquisition, repetition time/echo time (TR/TE) =500 ms/1.25 ms; initial TI =220 ms after each inversion pulse, flip angle =5º; field of view =165×220; acquisition matrix =128×96; bandwidth =601 Hz pixel^-1 single slice; slice thickness =8 mm; number of averages = 1; total acquisition time =15 sec. In this dynamic ASL study, the ASL sequence was executed consecutively in four sets (scan time was 1 minute). There was a 3-sec delay between each two sets of ASL sequence to ensure complete magnetization recovery.

An illustration of the study protocol is shown in Figure 1. SMBF measurements were obtained under three physiological conditions: 3-minute rest, during a 3-minute sustained isometric plantar-flexion exercise, and throughout a 4-minute post-exercise hyperemia and recovery phase. Three sets of ASL acquisitions were obtained at rest and exercise, for a total of six. Upon cessation of exercise, four sets of ASL acquisitions were performed sequentially to assess active hyperemia and the temporal recovery of SMBF towards baseline levels (τ_hyperemia).

Figure 1 Flow diagram of study protocol. Yellow arrow indicates direction of applied patient force. SMBF, skeletal muscle blood flow.

Image analysis

Deep learning method

We developed a physics-informed deep learning model for skeletal muscle ASL, termed DeepSMASL, to estimate calf SMBF. The DeepSMASL architecture is directly adopted from the published DeepMASL framework (24). The superiority of this deep learning reconstruction over conventional physics-based pixel-wise T1 fitting approach has been previously demonstrated in that study, where it reduced hyperemic blood flow measurement error from 33–49% to less than 10% against invasive microsphere gold-standard measurements. Repeating this comparison in the present study is therefore not necessary. The purpose is to generate more uniform SMBF maps from relatively noisy ASL data sets. The original physics model was validated in a cardiac canine model by using microsphere blood flow measurements as the gold standard (24). Briefly speaking, DeepSMASL utilized synthetic data sets to train and test the model. The use of synthetic training data is deliberate and established strategy for this class of physics-based deep learning models, motivated by the absence of ground-truth perfusion maps in vivo. This approach was previously validated in the published DeepMASL study (24), where a network trained entirely on synthetic data was subsequently shown to generalize to real in vivo conditions, achieving less than 10% error against invasive microsphere gold-standard measurements in a canine model of coronary artery disease. Synthetic datasets were generated using the established ASL kinetic model (14):

SMBF=λT1,vsT1,blood(1T1,ss−1T1,vs)[1]

Where T_1,vs and T_1,ss are the longitudinal relaxation times following volume-selective and slice-selective inversion pulses, respectively, T_1,blood is the blood T1, and λ is the blood-tissue water partition coefficient (mL/g).

For use in skeletal muscle, SMBF values were simulated in the range of 5–90 mL/min/100 g, and T1 values in the range of 1,100–1,300 ms. Signal intensities were then simulated using the modified Look-Locker equation accounting for T1 saturation between single-shot acquisitions (25):

SNsin(α)=An2−1+M0×(1−e−τNT1)+(An2×cos(α)×e−τNT1+SN−1sin(α)×Bn2×cos(α)×e−τNT1)×Bn2−1[2]

Where M₀ is the equilibrium magnetization, n is the RF pulse index, τ_N is the effective inversion time for the N-th acquisition, and A and B account for the relaxation effects between RF pulses. Gaussian noise corresponding to SNR levels between 20–90 was added to mimic in vivo acquisition conditions. Calf muscle masks from real MR datasets were applied to ensure realistic spatial representation.

The DeepSMASL architecture combines a U-Net backbone for spatial feature extraction and multi-stage dense layers for voxel-wise regression expressed in Figure 2. The encoder extracts hierarchical spatial features through repeated convolution and pooling operations, while the decoder reconstructs high-resolution feature maps using up-sampling and skip connections. Dense layers are applied at multiple stages to integrate global contextual information and model complex voxel-wise relationships between ASL series and SMBF values. Residual connections are incorporated to enhance gradient flow and improve network stability. Batch normalization and dropout are used to prevent overfitting and improve generalization. Input to the network is the ASL image series, and output is the quantitative SMBF map with identical spatial dimensions.

Figure 2 Illustration of a modified U-net structure for the DeepSMASL model to calculate skeletal muscle blood flow. MR, magnetic resonance.

Training was performed using 12,000 synthetic datasets with mean absolute error (MAE) as the loss function. The U-Net encoder-decoder combined with multi-stage dense layers is directly adopted from the published DeepMASL framework, where the architectural design including the integration of dense layers was developed, ablation-tested, and validated against invasive microsphere gold-standard measurements (24). Readers are referred to the prior publication for full architectural details and design justification. During inference, DeepSMASL directly produces voxel-level SMBF maps without requiring explicit parameter fitting or manual input, enabling rapid and robust perfusion quantification in calf muscles.

In addition, to accelerate temporal resolution of the ASL method, the number of single-shot acquisitions after each inversion pulse was reduced to 4 in training data sets, termed as under-sampling (US) to compare with fully sampled data sets (FS). In this way, the temporal resolution would be theoretically improved from 15 sec to 10 sec. DeepSMASL was then trained with reduced number of data acquisitions to calculate SMBF with US. The similarity of SMBF values obtained by US and FS scheme was assessed. One observation was noted that the first SMBF in each of 4 ASL data sets appeared higher in some subjects, likely due to the saturation of magnetization from previous ASL data acquisition. To ensure consistency of data modeling in all subjects, the first SMBF value of each ASL data set was removed for further data analysis (Figure 1), i.e., only 3 SMBF maps were created for each ASL data set.

Data analysis

Using anatomical landmarks identified on axial 2D T1-weighted TurboFLASH gradient-echo ASL images, ImageJ software (National Institutes of Health, Bethesda, MD, USA) was used to place regions of interest (ROIs) around five muscle groups (Figure 3A): medial head of gastrocnemius (MG), soleus, lateral gastrocnemius (LG), tibial anterior, and the lateral compartment. ROIs were manually adjusted at each acquisition to avoid large vessels and account for inter-condition motion. The ROIs were then copied onto the FS and the US SMBF maps created with the DeepSMASL technique (Figure 3B). The SMBF was quantified as mean ± SD in units of mL/min/100 g (Figure 3C). All SMBF values less than 0.01 mL/min/100 g or greater than 200 mL/min/100 g were omitted from quantification, attributed to either noise SMBF or contamination from large vessels. Across all analyzed ROIs, approximately 10% of pixels were omitted by this thresholding procedure, with 95% of excluded pixels having values below 0.01 mL/min/100 g and 5% exceeding 200 mL/min/100 g. Pixels with values larger than 200 mL/min/100 g were primarily associated with contamination from large vessels.

Figure 3 Representative anatomical images and SMBF maps demonstrating muscle segmentation, reconstruction performance, and perfusion measurements across physiological conditions in a healthy control and a patient with PAD. (A) Anatomical axial T1-weighted image used for muscle segmentation with major calf muscle groups labeled. (B) Corresponding SMBF maps generated from the same ASL acquisition using FS and US reconstruction methods. (C) Representative SMBF maps from a healthy control (top row) and a patient with PAD (bottom row) under three physiological conditions: rest, exercise, and active hyperemia. Color scale indicates SMBF (mL/min/100 g). ASL, arterial spin labeling; FS, fully-sampled; LG, lateral gastrocnemius; MG, medial gastrocnemius; PAD, peripheral arterial disease; SMBF, skeletal muscle blood flow; US, under-sampled.

The peak exercise flow (PEF) and percent change in peak exercise flow relative to baseline flow (∆PEF%) were measured. During the active hyperemia phase, immediately after the cessation of exercise, SMBF time-course for each muscle ROI was modeled to decay from a maximal hyperemic SMBF towards baseline SMBF using a mono-exponential decay model:

SMBF=SMBF0∗e−(tτ)+SMBFconstant[3]

where SMBF0 is the initial peak perfusion at the start of the hyperemic phase, SMBF_constant is the asymptotic baseline perfusion acquired using the last few points of recovery, t is the time since start of hyperemia, and τ is the recovery time constant. Time at recovery onset (t=0) was defined as either (I) the time point corresponding to the maximal observed SMBF in the hyperemia condition or (II) in cases where the maximal observed point appeared to either be an outlier or nearly equivalent to the previous time point, the immediately preceding point was used.

Statistical analysis

Primary outcome measures were PEF, ∆PEF% and τ_hyperemia. Analyses were focused on assessing differences and trends in entire SMBF time courses between two participant groups across two primary muscle groups (medial gastrocnemius and soleus). Continuous variables are presented as mean ± standard deviation (SD) for normally distributed data, or as median (25^th, 75^th percentile) for non-normally distributed data. Categorical variables are presented as n (%). Data distribution normality of continuous variables was first evaluated utilizing the Shapiro-Wilk test, and Levene’s test was used to determine whether the variances were equal. If neither of these required assumptions were violated, a two-sided paired t-test was used to assess within-group parameters between the FS method and the US method. A two-sided independent t-test was used for group comparisons between the two groups. When assumptions were violated, Welch, Wilcoxon, Mann-Whitney or Chi-squared test were used as appropriate.

Additionally, correlation analysis was used to assess the associations between primary outcome (PEF, ∆PEF% and τ_hyperemia) and clinical variables (BMI, ABI, and TBI) using the combined data of the two groups. If data residuals were non-normally distributed, Spearman’s coefficient of rank correlation (rho) was determined; otherwise, a correlation coefficient (r) was determined for normally distributed data.

SMBF kinetics were modeled using a linear mixed-effects model with restricted cubic splines to capture the nonlinear temporal response during the exercise protocol, estimated using restricted maximum likelihood with Kenward-Roger degrees of freedom. Fixed effects included group (HC vs. PAD), muscle region (five calf muscles), time, time × muscle interaction, and three-way interaction between time × group × muscle, with subject specified as a random intercept to account for correlation within-subject observations.

Most statistical analyses were performed using MedCalc Statistical Software version 18.2.1 (MedCalc Software bvba, Ostend, Belgium; http://www.medcalc.org; 2018) and GraphPad Prism version 10.5.0 for Windows (GraphPad Software, Boston, MA, USA). Linear mixed-effects regression model was used to evaluate heterogeneity of time-dependent changes in SMBF among five muscle groups and between two cohorts using SAS software (version 9.4; SAS Institute Inc. Cary, NC, USA). A P value <0.05 was considered statistically significant.

Results

Demographics of participants

Baseline characteristics of the study participants are summarized in Table 1. Age, sex, and BMI did not differ significantly between the two cohorts. Patients with PAD exhibited a significantly lower ABI compared to the healthy population (0.60±0.21 vs. 1.25±0.1, P<0.0001). This observation was similar in TBI testing with the PAD population exhibiting significantly diminished ratios compared to HCs (0.42±.2 vs. 0.81±0.12, P=0.0003), indicating impaired distal perfusion.

Quantitative calf muscle perfusion-group comparison correlation/differences (HC vs. PAD)

In the HCs, the medial gastrocnemius (53.2±13 mL/min/100 g) and the lateral gastrocnemius (44.3±21.9 mL/min/100 g) were the muscles with highest peak perfusion during the exercise protocol. In contrast, patients with PAD experience highest peak perfusion primarily in the lateral gastrocnemius (55.8±23 mL/min/100 g), followed in the medial gastrocnemius (46.6±18.7 mL/min/100 g) and the lateral compartment (45.2±19.1 mL/min/100 g). Table 2 summarizes the quantified MRI-derived perfusion metrics averaged across each patient cohort, while Table 3 presents perfusion parameters including: PEF, ∆PEF%, and τ_hyperemia across all muscle regions by patient group. PEF and ∆PEF% did not differ significantly between patients with PAD and HCs across any calf muscle compartment; except that the MG exhibited a trend toward lower ∆PEF% in PAD compared with controls [PAD: 54.0%±41.4% vs. HC: 83.4% (66.2%, 128.4%), P=0.05]. The time-dependent variation of percent change in SMBF relative to each individual’s baseline flow (∆SMBF%) across muscle compartment can be visualized in Figure 4, with HCs demonstrating increased changes in SMBF within the MG, LG, and lateral compartment, while the soleus and tibial remained relatively unaffected by group status.

Figure 4 Time-dependent line plot showing the mean percent change in SMBF from averaged baseline SMBF across five calf muscle regions for 9 HCs (in black) and 9 patients with PAD (in red) using fully sampled data. HC, healthy control; LG, lateral gastrocnemius; MG, medial gastrocnemius; PAD, peripheral arterial disease; SMBF, skeletal muscle blood flow; ∆SMBF%, percent change in SMBF relative to individual’s baseline flow.

In contrast, τ_hyperemia values were significantly prolonged across four of the five muscle compartments in PAD (Figure 5), indicating delayed post-exercise perfusion recovery compared with HC [MG: 140.4±99.9 vs. 44±35.6 s, P=0.021; soleus: 153.5±95.4 vs. 31 (23.4, 73.4) s, P=0.019; LG: 60.5 (47.9, 141.9) vs. 34.5±20.7 s, P=0.019; lateral: 74.4±47.5 vs. 25.8±19.7 s, P=0.017]. The tibial compartment demonstrated a similar but non-significant trend [53.1±48 vs. 11.8 (4.5, 42.1) s, P=0.094].

Figure 5 Comparison of exercise-induced perfusion responses and active hyperemia recovery time constants between HCs and patients with PAD across individual calf muscle compartments. (A) Percent change in peak exercise flow from patient specific baseline flow (∆PEF%) for individual calf muscle groups in HC and patients with PAD. Bars represent mean ± SD; P values indicate between-group comparisons for each muscle. (B) Active hyperemia flow recovery time constant (τ_hyperemia) for each muscle group in HC and PAD. Box-and-whisker plots show median, interquartile range, with “+” symbol representing mean value. P values denote between-group differences. HC, healthy control; LG, lateral gastrocnemius; MG, medial gastrocnemius; PAD, peripheral arterial disease; SD, standard deviation; ∆PEF%, percent change in peak exercise flow from patient specific baseline flow.

Correlation between patient demographics and quantitative MRI values

Correlation coefficients for ABI, TBI, and BMI across all muscle groups are summarized in Table 4. In the pooled cohort, ABI moderately and positively correlated with PEF and ∆PEF% in the MG (r=0.541, P=0.02; r=0.49, P=0.039, respectively), and demonstrated a significantly moderate and negative correlation with τ_hyperemia in the MG (r=−0.52, P=0.027) and the soleus (r=−0.503, P=0.033). A similar trend was observed for TBI; however, correlations were generally weaker with reduced statistical significance, except in the case of TBI-τ_hyperemia for the soleus where a significantly negative correlation (r=−0.651, P=0.006) was demonstrated. In contrast to ABI and TBI, BMI exhibited reversed associations with MRI-derived perfusion measurements: correlating negatively with PEF and ∆PEF% and positively with τ_hyperemia across all muscle groups.

Interaction of time, muscle groups, and patient groups

Linear mixed-effects modeling revealed significant main effects of group, muscle, and time (RCS), plus significant time × muscle (all P<0.05; Table 5), indicating distinct SMBF time-course patterns across muscles and cohorts. Notably, RCS curves represent population-level model predictions on the linear predictor scale and therefore reflect relative differences in shape and timing of the perfusion response rather than absolute SMBF values. Linear mixed-effects analysis showed significant higher resting SMBF in patients with PAD compared with controls (difference +8.3 mL/min/100 g, P=0.011). However, the primary finding was the highly significant three-way interaction among time × group × muscle (F=6.09, P<0.0001), demonstrating that muscle-specific temporal patterns were altered in PAD compared with HC. Consistent with the Type III tests, Figure 6 presents both locally estimated scatterplot smoothing (LOESS) trends and model-estimated RCS curves. The heterogeneous muscle-specific SMBF responses are clearly seen in HC, with the MG exhibiting the highest SMBF throughout the protocol, followed by lateral, LG, and soleus muscles, while the tibial region remained relatively unaffected. In contrast, patients with PAD exhibited a more homogeneous perfusion response across muscle regions, characterized by reduced separation between muscles and attenuated peak responses. Model-estimated RCS curves revealed a characteristic rise-to-peak and subsequent decline in SMBF for all muscle regions in both HC and PAD cohorts, except the tibial anterior muscle. Compared with HC, patients with PAD displayed attenuated temporal SMBF responses in the MG, lateral, and soleus muscles, suggesting limited hemodynamic reserve during exercise-induced demands.

Figure 6 Temporal response of SMBF across five calf muscle regions in HCs and patients with PAD. The left panel shows nonparametric LOESS-smoothed trends of the observed data, while the right panel displays model-estimated curves derived from a linear mixed-effects model using RCS. Red-shaded box represents start and end of exercise protocol. HC, healthy control; LG, lateral gastrocnemius; LOESS, locally estimated scatterplot smoothing; MG, medial gastrocnemius; PAD, peripheral arterial disease; RCS, restricted cubic splines; SMBF, skeletal muscle blood flow.

Method comparison (FS vs. US)

Beyond the validation of the DeepSMASL framework established in the published DeepMASL study (24), the agreement between FS and US DeepSMASL outputs in the present study provides further internal validation of the model’s consistency and reliability on real in vivo skeletal muscle data. In Figure 7, correlation analysis demonstrated strong linear associations between FS and US created SMBF measurements under all conditions: rest (r=0.769, P<0.0001), exercise (r=0.884, P<0.0001), and active hyperemia (r=0.85, P<0.0001). Regression slopes approximated unity, indicating excellent quantitative concordance between the two quantification methods across the three physiological states. The calculation of τ_hyperemia was also highly consistent between FS and US datasets, showing strong linear correlation (r=0.912, P<0.0001).

Figure 7 Agreement and correlation analyses demonstrate the consistency between FS and US SMBF measurements across different physiological conditions and for the τ_hyperemia. (A) The Bland-Altman plots (top row) and corresponding correlation plots (bottom row) comparing the FS and US SMBF measurements across three physiological states (rest, exercise, and active hyperemia). (B) Agreement (left) and correlation (right) for the τ_hyperemia between the FS and US measurements. FS, fully-sampled; MRI, magnetic resonance imaging; SD, standard deviation; SMBF, skeletal muscle blood flow; US, under-sampled; τ_hyperemia, active hyperemia flow recovery time constant.

Bland-Altman analysis revealed a small systematic bias between the FS and US techniques across all conditions, with mean differences of −1.2, −2.0, and −1.5 mL/min/100 g for rest, exercise and active hyperemia, respectively. These findings indicate close agreement between the two methods at lower flow ranges, whereas at higher flows, the US approach showed a tendency to slightly overestimate SMBF relative to FS. The τ_hyperemia calculation demonstrated minimal bias (mean difference =3.7 s) between techniques, supporting its higher latent precision for assessing perfusion dynamics with higher temporal resolution.

Discussion

Current work results demonstrate that the DeepSMASL technique can noninvasively capture spatially heterogeneous perfusion kinetics across different muscle compartments of the calf without the use of exogenous contrast agents. Moreover, the US model showed excellent agreement with the reference FS model, supporting the feasibility of DeepSMASL as a framework to improve temporal resolution with preserved spatial resolution, which would improve the quantification precision of ASL measurements, such as PEF, ∆PEF% and τ_hyperemia.

In a study conducted by Sultan Mahmud and Adil Bashir (2022), the recovery time constant of popliteal arterial blood flow in healthy subjects was measured at 38.03±6.91 s, which is comparable to the SMBF recovery observed in the present study (26). This study introduced a highly sensitive SMBF parameter to differentiate PAD from HCs: the recovery time constant of SMBF (τ_hyperemia). The τ_hyperemia was significantly prolonged (P<0.05) in patients with PAD across all major calf muscle compartments except the anterior tibialis region, while it also negatively correlated with ABI. This pattern is consistent with delayed vasodilatory onset, reduced capillary density, and slower perfusion wash-out described in other prior studies (18,27-29), and suggests that τ_hyperemia may serve as a sensitive quantity for differentiating PAD from normal vascular function.

HCs exhibited nearly a two-fold higher ∆PEF% relative to patients with PAD, suggesting decreased SMBF reserve in PAD. This finding aligns with prior work by Pollak et al. (2013), who reported a similar near two-fold difference in whole-calf peak exercise perfusion between normal and PAD subjects (HC: 80±23, PAD: 48±16 mL/min/100 g) (15). While our results did not demonstrate significant muscle-specific differences in ∆PEF% between cohorts, consistent with findings by Lopez et al., the RCS time-trend curves revealed significant differences in muscle heterogeneity and SMBF responses to exercise across cohorts (19). With sufficient spatial resolution, our findings also suggest that the reduced perfusion is muscle dependent rather than a homogeneous reduction in patients with PAD. Unlike prior ASL studies where resting SMBF was too low for reliable quantification (15,18,19), our dynamic ASL enhanced with DeepSMASL revealed a unique observation where resting SMBF was significantly higher in PAD than controls (difference +8.3 mL/min/100 g, P=0.011), a pattern also observed in some PET and contrast-enhanced ultrasound studies, where resting SMBF is higher in PAD but non-significant (30,31).

The isometric plantar-flexion exercise employed in this study predominantly recruits the gastrocnemius and soleus muscles, as indicated by 6 of 9 HCs demonstrating highest peak perfusion in the MG, consistent with findings from the linear mixed model and prior studies (17,32). In contrast, patients with PAD exhibited a more homogeneous distribution of peak perfusion, characterized by less distinguishable peaks, and with PEF occurring across a broader range of muscles (4 of 9 in the LG, 2 of 9 in the MG, and 2 of 9 in the soleus). In addition, the MG appeared more susceptible to disease status in the exercise protocol, exhibiting roughly a two-fold reduction in ∆PEF% and significantly prolonged τ_hyperemia against healthy controls. The soleus muscle functions predominantly as a postural muscle composed of approximately 70% to 80% slow-twitch type I fibers, characterized by high capillary density, oxidative capacity, and abundant mitochondrial content, rendering it more resilient to hypoxic stress during exercise in patients with PAD (33). Conversely, the gastrocnemius is evenly comprised of both type I fibers (50%) and fast-twitch type II fibers, relying heavily on glycolytic metabolism during high-intensity contractions. These differences in myofiber composition and metabolic profiles might account for the prolonged τ_hyperemia observed in the soleus across both cohorts compared to other muscle compartments, and the relatively preserved ∆PEF% in this region in PAD.

The primary methodological comparison in this study is between the FS and US DeepSMASL schemes, which directly addresses the novel contribution of this work—demonstrating that temporal resolution can be improved while maintaining quantitative fidelity. Comparison against the conventional non-deep-learning framework, and to our knowledge no, validated alternative deep learning method currently exists for skeletal muscle ASL perfusion quantification at this field strength and exercise protocol (24).

An important consideration in interpreting the DeepSMASL results relates to the parameters used for synthetic training data generation and the potential bias introduced by under-sampling. The model was trained on simulated SMBF values ranging from 5–90 mL/min/100 g and T1 values between 1,100–1,300 ms when using a 3T MRI system, which may impose implicit distributional priors on reconstructed perfusion maps. Such constraints can introduce range-compression effects, whereby very low physiological flows are biased upward toward the lower bound of the training manifold and higher flows are attenuated toward the upper bound. This mechanism may partially explain the relatively elevated resting SMBF observed in our cohort compared with prior reports (17,30,31), as well as the modestly lower PEF relative to studies employing maximal or symptom-limited exercise paradigms (15,32). In addition, the US acquisition scheme reduces the number of inversion-recovery sampling points, increasing sensitivity to noise, motion, and residual magnetization effects. At higher flow states, where signal evolution is more nonlinear and temporally dynamic, reduced sampling density may contribute to the mild high-flow bias observed in Bland–Altman analysis.

Several limitations should be acknowledged for the current technique employed in this study. First, the relatively small sample size limits the statistical power and generalizability of the findings, particularly for subgroup analyses across individual muscle groups. While the age- and sex-matched design between healthy and PAD cohorts strengthens internal validity, future studies with larger and more diverse populations are warranted to confirm the reproducibility of these findings. Secondly, although ASL offers a noninvasive approach for spatially and temporally quantifying SMBF, its inherently low signal-to-noise ratio and susceptibility to motion artifacts in skeletal muscle remain significant technical challenges for accurately measuring resting SMBF (14). Therefore, exercise-induced SMBF should be considered the primary biomarker for assessing microcirculation. Future studies, similar to Fahlström et al., using hybrid PET/MRI or other independent SMBF measurements should be implemented to validate our SMBF measurements (34,35). Deformable muscle motion artifacts introduced by voluntary isometric plantar flexion further contribute to signal variability. While such artifacts can be mitigated through cuff occlusion protocol, this approach may fail to capture physiological recruitment and perfusion response to voluntary exercise, or to reflect how compensatory mechanisms may manifest under PAD disease conditions (16,18,19,36-38).

Future technical development in DeepSMASL will therefore focus on expanding the synthetic training conditions to encompass a broader physiological range of perfusion and relaxation parameters, incorporating subject-specific or empirically measured T1 values, and explicitly modeling incomplete magnetization recovery and motion-related perturbations. Reweighting extreme low- and high-flow conditions during training may further mitigate distributional bias. Methodologically, optimized or adaptive under-sampling strategies could improve temporal resolution while preserving quantitative fidelity, particularly during rapid hyperemic transitions. External validation against independent perfusion standards, such as contrast-enhanced MRI or PET/MRI, will be critical to establish quantitative accuracy across varying disease severities.

Conclusions

In conclusion, this study emphasizes the value of temporally resolved imaging by using a newly developed DeepSMASL method as a key indicator for distinguishing patients with PAD from HC, as evidenced by the highly significant τ_hyperemia and attenuated responses observed in the model-estimated RCS curves. Leveraging DeepSMASL’s ability to enhance temporal resolution, quantification precision would be improved. These SMBF time course patterns captured using dynamic DeepSMASL provide a noninvasive, contrast-free, muscle-specific characterization of microvascular perfusion deficits, offering a promising tool for assessing PAD severity and monitoring therapeutic interventions.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STARD-AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0675/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0675/dss

Funding: This work was supported by the National Institute of Health (No. R21 AR065672 to J.Z., Nos. K08HL132060 and R01HL153262 to M.Z., No. R01DK105322 to J.Z. and M.K.H); and Washington University Institute of Clinical and Translational Sciences of National Institute of Health (No. UL1 TR002345 to J.Z.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0675/coif). J.Z. serves as an unpaid editorial board member of Quantitative Imaging in Medicine and Surgery. M.Z. is a consultant for Acera Surgical, Amgen, Medtronic, and GlucoTrack. In the last 36 months, M.Z. received consulting fees from Amgen, Medtronic and GlucoTrack. He received reimbursement from Medtronic and Boston Scientific to attend a meeting. He is a co-founder and holds equity in Caeli Vascular, Inc., AirSeal Cardiovascular, Inc., and Vascorra, LLC. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Washington University Human Research Protection Office (No. 201603075) and informed consent was obtained from all individual participants prior to the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Allison MA, Armstrong DG, Goodney PP, Hamburg NM, Kirksey L, Lancaster KJ, Mena-Hurtado CI, Misra S, Treat-Jacobson DJ, White Solaru KT. Health Disparities in Peripheral Artery Disease: A Scientific Statement From the American Heart Association. Circulation 2023;148:286-96. [Crossref] [PubMed]
2. Institute of Medicine (US) Committee on Social Security Cardiovascular Disability Criteria. Cardiovascular disability: updating the social security listings. Washington (DC): National Academies Press (US); 2010.
3. Khagi B, Belousova T, Short CM, Taylor AA, Bismuth J, Shah DJ, Brunner G. Contrast-enhanced magnetic resonance imaging based calf muscle perfusion and machine learning in peripheral artery disease. Sci Rep 2025;15:4996. [Crossref] [PubMed]
4. Addison O, Prior SJ, Kundi R, Serra MC, Katzel LI, Gardner AW, Ryan AS. Sarcopenia in Peripheral Arterial Disease: Prevalence and Effect on Functional Status. Arch Phys Med Rehabil 2018;99:623-8. [Crossref] [PubMed]
5. Bethel M, Annex BH. Peripheral arterial disease: A small and large vessel problem. Am Heart J Plus 2023;28:100291. [Crossref] [PubMed]
6. Mohler ER 3rd. Peripheral arterial disease: identification and implications. Arch Intern Med 2003;163:2306-14. [Crossref] [PubMed]
7. Koutakis P, Weiss DJ, Miserlis D, Shostrom VK, Papoutsi E, Ha DM, Carpenter LA, McComb RD, Casale GP, Pipinos II. Oxidative damage in the gastrocnemius of patients with peripheral artery disease is myofiber type selective. Redox Biol 2014;2:921-8. [Crossref] [PubMed]
8. Callegari S, Feher A, Smolderen KG, Mena-Hurtado C, Sinusas AJ. Multi-modality imaging for assessment of the microcirculation in peripheral artery disease: Bench to clinical practice. Am Heart J Plus 2024;42:100400. [Crossref] [PubMed]
9. Behroozian A, Beckman JA. Microvascular Disease Increases Amputation in Patients With Peripheral Artery Disease. Arterioscler Thromb Vasc Biol 2020;40:534-40. [Crossref] [PubMed]
10. Norgren L, Hiatt WR, Dormandy JA, Nehler MR, Harris KA, Fowkes FG; . Inter-Society Consensus for the Management of Peripheral Arterial Disease (TASC II). J Vasc Surg 2007;45 Suppl S:S5-67.
11. AbuRahma AF, Adams E, AbuRahma J, Mata LA, Dean LS, Caron C, Sloan J. Critical analysis and limitations of resting ankle-brachial index in the diagnosis of symptomatic peripheral arterial disease patients and the role of diabetes mellitus and chronic kidney disease. J Vasc Surg 2020;71:937-45. [Crossref] [PubMed]
12. Hirsch AT, Haskal ZJ, Hertzer NR, Bakal CW, Creager MA, Halperin JLAmerican Association for Vascular Surgery, et al. Society for Vascular Surgery; Society for Cardiovascular Angiography and Interventions; Society for Vascular Medicine and Biology; Society of Interventional Radiology; ACC/AHA Task Force on Practice Guidelines Writing Committee to Develop Guidelines for the Management of Patients With Peripheral Arterial Disease; American Association of Cardiovascular and Pulmonary Rehabilitation; National Heart, Lung, and Blood Institute; Society for Vascular Nursing; TransAtlantic Inter-Society Consensus; Vascular Disease Foundation. ACC/AHA 2005 Practice Guidelines for the management of patients with peripheral arterial disease (lower extremity, renal, mesenteric, and abdominal aortic): a collaborative report from the American Association for Vascular Surgery/Society for Vascular Surgery, Society for Cardiovascular Angiography and Interventions, Society for Vascular Medicine and Biology, Society of Interventional Radiology, and the ACC/AHA Task Force on Practice Guidelines (Writing Committee to Develop Guidelines for the Management of Patients With Peripheral Arterial Disease): endorsed by the American Association of Cardiovascular and Pulmonary Rehabilitation; National Heart, Lung, and Blood Institute; Society for Vascular Nursing; TransAtlantic Inter-Society Consensus; and Vascular Disease Foundation. Circulation 2006;113:e463-654. [Crossref] [PubMed]
13. Stacy MR, Sinusas AJ. Novel Applications of Radionuclide Imaging in Peripheral Vascular Disease. Cardiol Clin 2016;34:167-77. [Crossref] [PubMed]
14. Zheng J, An H, Coggan AR, Zhang X, Bashir A, Muccigrosso D, Peterson LR, Gropler RJ. Noncontrast skeletal muscle oximetry. Magn Reson Med 2014;71:318-25. [Crossref] [PubMed]
15. Pollak AW, Meyer CH, Epstein FH, Jiji RS, Hunter JR, Dimaria JM, Christopher JM, Kramer CM. Arterial spin labeling MR imaging reproducibly measures peak-exercise calf muscle perfusion: a study in patients with peripheral arterial disease and healthy volunteers. JACC Cardiovasc Imaging 2012;5:1224-30. [Crossref] [PubMed]
16. Chen HJ, Roy TL, Wright GA. Perfusion measures for symptom severity and differential outcome of revascularization in limb ischemia: Preliminary results with arterial spin labeling reactive hyperemia. J Magn Reson Imaging 2018;47:1578-88. [Crossref] [PubMed]
17. Schewzow K, Fiedler GB, Meyerspeer M, Goluch S, Laistler E, Wolzt M, Moser E, Schmid AI. Dynamic ASL and T2-weighted MRI in exercising calf muscle at 7 T: a feasibility study. Magn Reson Med 2015;73:1190-5. [Crossref] [PubMed]
18. Wu WC, Mohler E 3rd, Ratcliffe SJ, Wehrli FW, Detre JA, Floyd TF. Skeletal muscle microvascular flow in progressive peripheral artery disease: assessment with continuous arterial spin-labeling perfusion magnetic resonance imaging. J Am Coll Cardiol 2009;53:2372-7. [Crossref] [PubMed]
19. Lopez D, Pollak AW, Meyer CH, Epstein FH, Zhao L, Pesch AJ, Jiji R, Kay JR, DiMaria JM, Christopher JM, Kramer CM. Arterial spin labeling perfusion cardiovascular magnetic resonance of the calf in peripheral arterial disease: cuff occlusion hyperemia vs exercise. J Cardiovasc Magn Reson 2015;17:23. [Crossref] [PubMed]
20. Zheng J, Li R, Zayed MA, Yan Y, An H, Hastings MK. Pilot study of contrast-free MRI reveals significantly impaired calf skeletal muscle perfusion in diabetes with incompressible peripheral arteries. Vasc Med 2021;26:367-73. [Crossref] [PubMed]
21. Zhang H, Shea SM, Park V, Li D, Woodard PK, Gropler RJ, Zheng J. Accurate myocardial T1 measurements: toward quantification of myocardial blood flow with arterial spin labeling. Magn Reson Med 2005;53:1135-42. [Crossref] [PubMed]
22. Belle V, Kahler E, Waller C, Rommel E, Voll S, Hiller KH, Bauer WR, Haase A. In vivo quantitative mapping of cardiac perfusion in rats using a noninvasive MR spin-labeling method. J Magn Reson Imaging 1998;8:1240-5. [Crossref] [PubMed]
23. Kim SG. Quantification of relative cerebral blood flow change by flow-sensitive alternating inversion recovery (FAIR) technique: application to functional mapping. Magn Reson Med 1995;34:293-301. [Crossref] [PubMed]
24. Li R, Berberet C, Huang Q, Woodard PK, Zheng J. Improved accuracy for myocardial blood flow mapping with deep learning-enabled CMR arterial spin labeling (DeepMASL): Validation by microsphere in vivo. J Cardiovasc Magn Reson 2025;27:101989. [Crossref] [PubMed]
25. Look DC, Locker DR. Time saving in measurement of NMR and EPR relaxation times. Rev Sci Instrum 1970;41:250-1.
26. Mahmud SZ, Bashir A. Repeatability assessment for simultaneous measurement of arterial blood flow, venous oxygen saturation, and muscle perfusion following dynamic exercise. NMR Biomed 2023;36:e4872. [Crossref] [PubMed]
27. da Silva ND Jr, Andrade-Lima A, Chehuen MR, Leicht AS, Brum PC, Oliveira EM, Wolosker N, Pelozin BRA, Fernandes T, Forjaz CLM. Walking Training Increases microRNA-126 Expression and Muscle Capillarization in Patients with Peripheral Artery Disease. Genes (Basel) 2022;14:101. [Crossref] [PubMed]
28. Robbins JL, Jones WS, Duscha BD, Allen JD, Kraus WE, Regensteiner JG, Hiatt WR, Annex BH. Relationship between leg muscle capillary density and peak hyperemic blood flow with endurance capacity in peripheral artery disease. J Appl Physiol 1985;2011:81-6. [Crossref] [PubMed]
29. Englund EK, Langham MC, Ratcliffe SJ, Fanning MJ, Wehrli FW, Mohler ER 3rd, Floyd TF. Multiparametric assessment of vascular function in peripheral artery disease: dynamic measurement of skeletal muscle perfusion, blood-oxygen-level dependent signal, and venous oxygen saturation. Circ Cardiovasc Imaging 2015;8:e002673. [Crossref] [PubMed]
30. Divakaran S, Harms HJ, Robertson M, Merugumala SK, Park MA, Kijewski MF, Martell LB, Morgan V, Barrett L, Perillo A, Yang D, Jarolim P, Feinberg MW, Gerhard-Herman MD, Belkin M, Lin AP, Creager MA, Bonaca MP, Di Carli MF. Post-walking exercise skeletal muscle perfusion and energetics in patients with symptomatic lower extremity peripheral artery disease. J Nucl Cardiol 2025;46:102143. [Crossref] [PubMed]
31. Meneses AL, Nam MCY, Bailey TG, Magee R, Golledge J, Hellsten Y, Keske MA, Greaves K, Askew CD. Leg blood flow and skeletal muscle microvascular perfusion responses to submaximal exercise in peripheral arterial disease. Am J Physiol Heart Circ Physiol 2018;315:H1425-33. [Crossref] [PubMed]
32. Zhang JL, Layec G, Hanrahan C, Conlin CC, Hart C, Hu N, Khor L, Mueller M, Lee VS. Exercise-induced calf muscle hyperemia: quantitative mapping with low-dose dynamic contrast enhanced magnetic resonance imaging. Am J Physiol Heart Circ Physiol 2019;316:H201-11. [Crossref] [PubMed]
33. Askew CD, Green S, Walker PJ, Kerr GK, Green AA, Williams AD, Febbraio MA. Skeletal muscle phenotype is associated with exercise tolerance in patients with peripheral arterial disease. J Vasc Surg 2005;41:802-7. [Crossref] [PubMed]
34. Fahlström M, Appel L, Kumlien E, Danfors T, Engström M, Wikström J, Antoni G, Larsson EM, Lubberink M. Evaluation of Arterial Spin Labeling MRI-Comparison with 15O-Water PET on an Integrated PET/MR Scanner. Diagnostics (Basel) 2021;11:821. [Crossref] [PubMed]
35. Haddock B, Holm S, Poulsen JM, Enevoldsen LH, Larsson HB, Kjær A, Suetta C. Assessment of muscle function using hybrid PET/MRI: comparison of (18)F-FDG PET and T2-weighted MRI for quantifying muscle activation in human subjects. Eur J Nucl Med Mol Imaging 2017;44:704-11. [Crossref] [PubMed]
36. Englund EK, Rodgers ZB, Langham MC, Mohler ER 3rd, Floyd TF, Wehrli FW. Measurement of skeletal muscle perfusion dynamics with pseudo-continuous arterial spin labeling (pCASL): Assessment of relative labeling efficiency at rest and during hyperemia, and comparison to pulsed arterial spin labeling (PASL). J Magn Reson Imaging 2016;44:929-39. [Crossref] [PubMed]
37. Brunner G, Bismuth J, Nambi V, Ballantyne CM, Taylor AA, Lumsden AB, Morrisett JD, Shah DJ. Calf muscle perfusion as measured with magnetic resonance imaging to assess peripheral arterial disease. Med Biol Eng Comput 2016;54:1667-81. [Crossref] [PubMed]
38. Chen HJ, Wright GA. A physiological model for interpretation of arterial spin labeling reactive hyperemia of calf muscles. PLoS One 2017;12:e0183259. [Crossref] [PubMed]