Dynamic Contrast Enhanced—Magnetic Resonance Fingerprinting (DCE‐MRF) Improves Detection of Induced Vascular Perfusion Reduction in a Mouse Breast Cancer Model

ABSTRACT

Purpose

Dynamic contrast‐enhanced MRI (DCE‐MRI) can assess tumor perfusion using pharmacokinetic models. However, poor DCE‐MRI reproducibility from reliance on conventional T₁‐weighted MRI has limited clinical translation. We evaluated whether Dynamic Contrast Enhanced‐Magnetic Resonance Fingerprinting (DCE‐MRF), which directly generates quantitative T₁ relaxation time constant maps, provides improved precision and statistical power for detecting treatment‐induced vascular changes compared to conventional DCE‐MRI.

Methods

Twenty female mice bearing orthotopic 4 T1 breast tumors were randomly assigned to DCE‐MRF (n = 12) or conventional DCE‐MRI (n = 8) cohorts. Both methods acquired matched spatial and temporal resolution (23‐s) T₁ measurements at baseline, 3 h, and 48 h post‐treatment with combretastatin A4 phosphate (120 mg/kg), a known vascular disrupting agent. Perfusion assessments were obtained with a pharmacokinetic Linear Reference Region Model (RK ^trans and k _ep,T) and model‐independent initial area under the curve assessments and compared using Wilcoxon signed‐rank tests and Hedges' g effect sizes.

Results

DCE‐MRF demonstrated 1.3‐to‐5.1‐fold larger effect sizes for RK ^trans across all regions and timepoint comparisons, and 1.3‐to‐1.9‐fold larger effect sizes for tumor rim k _ep,T compared to conventional DCE‐MRI. RK ^trans and k _ep,T at 3 h post‐treatment in all regions, detected significant vascular recovery for whole tumor RK ^trans and both whole tuomor and tumor rim for k _ep,T at 48 h, whereas DCE‐MRI only detected significant changes at 3 h.

Conclusions

DCE‐MRF's improved measurement precision and increased effect size translates directly to enhanced statistical power for detecting treatment‐induced vascular changes, positioning it as a more reproducible tumor perfusion assessment for animal models and eventually cancer patients.

1. Introduction

For the past several decades, many prior imaging studies in both animal models and patients have described the potential advantages of Dynamic Contrast Enhanced (DCE) – MRI assessments of tumor perfusion to improve care for cancer patients [1, 2, 3, 4, 5]. DCE‐MRI leverages increased contrast accumulation into the tumor, leading to shorter T₁ and/or T₂ magnetic relaxation time constants and improved tumor visualization [6, 7], providing functional and physiologic insights beyond anatomic imaging.

While researchers have demonstrated the potential of DCE‐MRI to quantify tumor vascular perfusion rates (i.e., K ^trans) [8, 9, 10, 11], clinical translation remains hampered by the established high variability and poor reproducibility of these conventional DCE‐MRI perfusion methods [10, 12, 13, 14, 15, 16, 17]. As a result, most clinical applications are confined to simpler, semi‐quantitative or empirical assessments, such as time‐to‐peak enhancement, providing reduced physiological insight [18, 19]. The primary source of this high variability from DCE‐MRI perfusion measurements is the fact that the majority of DCE‐MRI methods are based on conventional T₁‐weighted images [14, 15, 16, 17, 20]. These T₁‐weighted images require multiple assumptions to perform pharmacokinetic modeling, and are highly susceptible to many MRI error sources including motion artifacts, coil position, temperature, signal‐to‐noise levels, and B₁ heterogeneities [14]. Therefore, a fundamentally new imaging approach is needed to make quantitative vascular assessments a robust and reliable clinical tool.

To address this long‐standing clinical limitation, we have developed a new Dynamic Contrast Enhanced—Magnetic Resonance Fingerprinting (DCE‐MRF) method designed for rapid, accurate, and precise quantification of tumor vascular perfusion [21]. In this initial development, DCE‐MRF generated maps of the T₁ relaxation time constant (referred to as “T₁” for brevity) with a 23‐s temporal resolution (8‐s MRF acquisition and a 15‐s delay) using multiple inversion preparations, highly‐undersampled spiral trajectories (R = 48), low tip angles to limit the impact of B₁ field heterogeneities [22, 23, 24, 25, 26], and a Low Rank (LR) reconstruction method to limit artifacts in the raw MRF images [21, 24, 27]. Importantly, this initial implementation of DCE‐MRF demonstrated a 2–3× reduction in the temporal T₁ variation in comparison to conventional DCE‐MRI methods when testing in in vitro phantoms [21]. Therefore, DCE‐MRF directly eliminates the core source of variability in traditional DCE‐MRI by basing pharmacokinetic analysis entirely on the directly‐acquired MRF‐based T₁ maps, thereby providing a more reproducible approach to assessing tumor vascular perfusion.

In the current study, we directly compared the performance of the new DCE‐MRF methodology with a conventional DCE‐MRI method to longitudinally monitor therapeutic response. Specifically, we evaluated these perfusion methods in a preclinical murine breast cancer model treated with a vascular disrupting agent, combretastatin A4 phosphate (CA4P), which is known to induce measurable reductions in tumor vascular perfusion [28]. Herein, we demonstrate that the inherently improved precision of DCE‐MRF translates into increased effect sizes and statistical power in detecting drug‐induced changes in tumor perfusion compared to conventional DCE‐MRI, ultimately positioning DCE‐MRF as a more reliable quantitative imaging biomarker methodology.

2. Methods

2.1. In Vivo Studies

All MRI studies were conducted using a 9.4 T Bruker Biospec preclinical MRI scanner with Paravision 360 v3.3 software and maximum gradient strength of 660 mT/m (Bruker Biospin, Billerica, MA). Data acquisition was performed with a 40‐mm dual‐channel transmit/receive volume coil.

All animal imaging studies were performed in accordance with approved Institutional Animal Care and Use Committee (IACUC) protocols at our institution. All experiments were performed on 20 female BALBC/cJ mice (Jackson Laboratories, Bar Harbor, ME, USA). To generate the mouse 4 T1 breast cancer model, 10 000 4 T1 cells were injected into the right 4th mammary gland. Tumors were given two weeks to develop prior to imaging with tumor diameters ranging between 4 and 13 mm. All mice were 13 weeks old upon MRI imaging.

Mice were maintained under anesthesia using 1%–1.5% isoflurane with 100% oxygen throughout the catheterization and imaging procedures. A 26‐gauge catheter (Coviden, Mansfield, MA) was placed in the tail vein for intravenous administration of 0.2 mg/kg Gadovist, an FDA‐approved gadolinium‐based MRI contrast agent, during DCE‐MRF and DCE‐MRI scanning. Animals were randomly assigned to receive either DCE‐MRF or conventional DCE‐MRI acquisitions. In both the DCE‐MRF (n = 12) and DCE‐MRI (n = 8) imaging cohorts, each mouse was imaged 2 days before treatment, as well as at 3 h and 48 h after treatment with a known vascular disrupting agent, combretastatin A4 phosphate (CA4P) at 120 mg/kg injected intraperitoneally, which is known to significantly reduce tumor perfusion in preclinical DCE‐MRI studies [28, 29].

2.2. In Vivo DCE‐MRF Studies

Animals were positioned supine for localizer scans. Coil sensitivity maps were then obtained using a modified RARE acquisition (TR = 3500 ms, TE = 32 ms, 2 averages, 1 repetition, echo spacing = 7.9 ms, RARE factor = 16, matrix size = 256 × 256, FOV = 50 × 50 mm², slice thickness = 1 mm, scan duration = 1.9 min) to allow for accurate combination of the two receive channels. The DCE‐MRF data were acquired using the previously described DCE‐MRF protocol with 23‐s temporal resolution over 60 repetitions [21]. Baseline T₁ measurements were obtained from the initial 5 repetitions over 1 min 55 s before contrast injection. Gadovist at 0.2 mmol/kg was manually administered at the start of the 6th repetition over a 10‐s period to minimize strain on the tail vein. The remaining 55 post‐contrast repetitions over 21 min 5 s captured the contrast agent pharmacokinetics (60 total repetitions).

All MRI experiments were performed on a Bruker Biospec 9.4 T MRI scanner using a previously‐described T₁‐only MRF acquisition to generate accurate T₁ maps with a 23‐s temporal resolution (8‐s MRF acquisition and a 15‐s delay) [21]. This acquisition used three sequential lobes with adiabatic inversion pulses (inversion time = 20 ms or 500 ms) followed by 240 sequential FISP kernels with a highly‐undersampled spiral trajectory (1 spiral interleaf/image, acceleration factor (R) = 48) [21]. The 720 total MRF images were acquired with TR = 10 ms, TE = 2 ms, variable flip angles (4°–14°) to mitigate B₁ field inhomogeneities [22], and $4\pi$ dephasing to limit banding artifacts [21, 23, 24]. The 8‐s MRF acquisition was combined with a 15‐s delay to limit the duty cycle of the magnetic field gradients, providing T₁ maps with 23‐s temporal resolution. Imaging parameters included FOV = 50 × 50 mm², matrix size = 256 × 256, slice thickness = 1 mm, and in‐plane resolution = $195\times 195\mathrm{\mu m}$ [21, 24].

The MRF dictionary was simulated for a single repetition in Python (Python Software Foundation, v3.10.12) using the Bloch equations and the mrftools [30] (v0.2.8) software utilizing the MRF acquisition parameters [21]. The T₁ relaxation time constants for the MRF dictionary ranged from 200 to 4000 ms (200–2020 ms with a 5 ms step size, 2020–4000 ms with a step size of 10 ms). T₂ relaxation time constants were included in the dictionary simulation to account for T₂‐weighting in the acquired FISP MRF acquisition [23, 31]. The T₂ relaxation time constants ranged from 2 to 600 ms (2–202 ms with a 2 ms step size, 202–600 ms with a step size of 10 ms) [21]. This generated 78 680 valid dictionary entries with 400 isochromats to account for the $4\pi$ dephasing gradient.

Reconstruction and dictionary matching was implemented with PyTorch (v2.3.1) on an Azure instance with NVIDIA A10 GPU, employing an iterative low‐rank (LR) subspace method with measured spiral trajectories and ESPIRiT [32]‐calculated coil sensitivity maps [21, 24, 27]. A temporal subspace rank of 10 was used to ensure over 99% of the energy represented by the singular values and the cost function was minimized using a conjugate gradient descent over 25 iterations [27, 33]. MRF‐based T₁ maps were generated via dictionary matching on a voxel‐by‐voxel basis typical of prior MRF methods [22, 24, 27]. The iterative LR MRF image reconstruction and subsequent dictionary matching methods required a total of 8 min and 22 s to reconstruct a single DCE‐MRF repetition.

The DCE‐MRF datasets were reconstructed and matched using the LR method described above [21] (total reconstruction time: 8 h 22 min for 60 repetitions) prior to pharmacokinetics modeling in MATLAB. Because DCE‐MRF directly measures the T₁ values at each timepoint, gadolinium concentration maps were calculated voxel‐wise using the established linear relationship between concentration and magnetic relaxivity r₁ (Equation 1):

$$\frac{1}{T_1}=\frac{1}{T_{1,0}}+r_1\times [\mathit{Gd}]$$ (1)

where 1/T₁ (R₁) represents the post‐contrast value at each imaging timepoint; 1/T_1,0 (R_1,0) is the average baseline pre‐injection value; r₁ is magnetic relaxivity of Gadovist (3.62 mM⁻¹ s⁻¹ at 9.4 T) [34]; and [Gd] indicates the Gadovist derived gadolinium concentration [35, 36].

2.3. In Vivo DCE‐MRI Studies

Similar to the DCE‐MRF studies above, animals were first positioned supine for localizer scans to verify appropriate tumor positioning for the axial single‐slice DCE‐MRI imaging studies. A baseline T₁ map was then acquired using a FLASH variable flip angle method (Flip Angle = 10°, 30°, 50°, 70°, 90°, TR = 90 ms, TE = 4 ms, matrix size = 256 × 256, FOV = 50 × 50 mm², and slice thickness = 1 mm). A dictionary was simulated using the FLASH equation with T₁ relaxation time constants 0–6000 ms with a step size of 1 ms. The signal intensity at each flip angle was then matched to the normalized dictionary simulated from the FLASH equation (Equation 2) [37] to generate a T_1,0 map.

$$S=\sin (\propto )·\left[\frac{1-e^{-\frac{\mathit{TR}}{T_1}}}{1-\cos (\alpha )e^{-\frac{\mathit{TR}}{T_1}}}\right]$$ (2)

S is the FLASH T₁ weighted signal intensity, TR is the repetition time, $\propto$ is the flip angle, and T₁ is the T₁ relaxation time constant.

DCE‐MRI data were then collected using a FLASH sequence with matched spatial and temporal resolution parameters to the DCE‐MRF acquisition (matrix size = 256 × 256, FOV = 50 × 50 mm², and slice thickness = 1 mm, temporal resolution = 23 s/repetition). The FLASH acquisition had the following parameters: TR = 90 ms, TE = 2 ms, and flip angle = 35°. Similar to the DCE‐MRF studies, Gadovist was manually injected ˜2 min into the DCE‐MRI acquisition during repetition 6.

The DCE‐MRI data were exported to MATLAB for offline reconstruction and analysis. To enable comparison with DCE‐MRF derived T₁ dynamics, dynamic T₁ values were estimated form the FLASH signal intensities using the steady‐state FLASH equation (Equation 3): [37]

$$\frac{1}{T_1}=-\frac{1}{\mathit{TR}}·\ln \left[1-\frac{S(t)}{S(0)}\left(1-e^{\left(-\frac{\mathit{TR}}{T_{1,0}}\right)}\right)\right]$$ (3)

where TR is the repetition time, S(t) represents the T₁‐weighted signal intensity at time t, S(0) is the baseline signal intensity, and T_1,0 is the pre‐injection T₁ value from the FLASH variable flip angle T₁ map.

Utilizing this estimated T₁ value, gadolinium concentration maps were calculated on a voxel‐wise basis using the established linear relationship between concentration and magnetic relaxivity r ₁ (Equation 1) [35, 36].

2.4. Pharmacokinetic Modeling and Model‐Independent Perfusion Analysis

All pharmacokinetics modeling was performed in MATLAB for both the DCE‐MRF and DCE‐MRI studies at all timepoints (baseline, 3 h post‐treatment, 48 h post‐treatment). First, regions of interest (ROIs) were manually selected for the whole tumor and the paraspinal muscle. The calculated voxel‐wise dynamic gadolinium concentration maps were used as input for the established Linear Reference Region Model (LRRM) [37, 38, 39, 40, 41] to generate maps of RK ^trans and k _ep for each mouse and imaging timepoint. The LRRM model is shown in Equation (4): [37, 38]

$$C_{\mathrm{T}}(t)={\mathit{RK}}^{\text{trans}}·C_{\mathrm{M}}(t)+\frac{K^{\text{trans},\mathrm{T}}}{v_{\mathrm{e},\mathrm{M}}}·{\int }_0^t{C}_{\mathrm{M}}\left(t^{\prime }\right)d{t}^{\prime }-k_{\mathrm{ep},\mathrm{T}}·{\int }_0^t{C}_{\mathrm{T}}\left(t^{\prime }\right)d{t}^{\prime },$$ (4)

where C _T(t) represents the tumor gadolinium concentration; C _M(t) represents the muscle gadolinium concentration; RK ^trans (a.u.) represents the ratio of K ^trans,T and K ^trans,M, the volume transfer constants describing tracer movement between the plasma and extravascular extracellular space in tumor and muscle (reference region), respectively; v _e,M denotes the extravascular extracellular volume fraction in the muscle; and k _ep,T ($\frac{K^{\text{trans},\mathrm{T}}}{v_{\mathrm{e},\mathrm{T}}}$) and k _ep,M ($\frac{K^{\text{trans},\mathrm{M}}}{v_{\mathrm{e},\mathrm{M}}}$) represent the rate constants (min⁻¹) of contrast agent efflux from tumor and muscle tissues back into the plasma.

Voxel‐wise parametric maps were generated using the LRRM with convergence tolerance of 1 × 10⁻¹⁰, maximum of 1000 iterations, and MATLAB's lsqcurvefit method [38]. Voxels with relative residual error $\le$ 0.2 were classified as “tumor rim” and those $>$ 0.2 were classified as “tumor core” (Figure S1), with better fits typically occurring in well‐perfused regions where contrast agent kinetics are more robust. While somewhat arbitrary for this initial in vivo study, this threshold provided an unbiased identification of rim regions across animals and timepoints, as evidenced by the reproducible pharmacokinetics parameter distributions observed.

Next, the LRRM analysis of the tumor rim ROI generated voxel‐wise maps of RK ^trans and k _ep,T. Pre‐injection R_1,0 (1/T_1,0) values were calculated by averaging R₁ (1/T₁) measurements (DCE‐MRF) or R₁ estimates (DCE‐MRI) from repetitions 2–4, while late‐phase R₁ values were determined by averaging repetitions 51–60, which was 17.6–21 min after injection. Additionally, box and whisker plots were used to compare the measured or estimated mean R₁ and R_1,0 values as well as the RK ^trans and k _ep,T mean values from the LRRM at all three timepoints using the previously mentioned whole tumor, tumor rim, and muscle ROIs. A paired Wilcoxon signed‐rank test was used to evaluate statistical significance of R₁, R_1,0, RK ^trans, and k _ep,T across all timepoints and all ROIs. Hedges' g effect size was calculated for baseline versus 3 h post‐treatment, baseline versus 48 h post‐treatment, and 3 h post‐treatment versus 48 h post‐treatment timepoints.

To verify that observed statistical differences between DCE‐MRF (n = 12) and DCE‐MRI (n = 8) cohorts were not artifacts of sample size disparities, a bootstrap subsampling validation was performed on the DCE‐MRF dataset. For each bootstrap iteration, 8 DCE‐MRF datasets were randomly selected to match the DCE‐MRI group composition: 5 datasets with all three timepoints (baseline, 3 post‐treatment, and 48 h post‐treatment) and 3 additional datasets with only baseline and 3 h post‐treatment measurements. This process was repeated 10 000 times. For each iteration, Wilcoxon signed‐rank tests were performed comparing: (1) baseline versus 3 h post‐treatment (n = 8), (2) baseline versus 48 h post‐treatment (n = 5), and (3) 3 h versus 48 h post‐treatment (n = 5) timepoints. The proportion of iterations yielding statistically significant results (p < 0.05) and the distribution of p‐values across iterations were calculated for each comparison to assess the robustness of findings independent of sample size advantages.

As a secondary comparison, the R₁ curves from DCE‐MRF and DCE‐MRI were used to calculate the initial area under the curve (IAUC) that represents a model‐independent assessment of tumor perfusion [42, 43]. For this IAUC analysis, the area under the R₁ curve was integrated over the first four scans following contrast administration (i.e., repetitions 6–9) starting with the scan acquired during the contrast injection (6th repetition) providing an assessment of the first 92 s of contrast administration. Similar to the pharmacokinetics analysis above, paired Wilcoxon signed‐rank tests were used to evaluate statistical significance of IAUC values across all timepoints, and Hedges' g effect size was calculated for baseline versus 3 h post‐treatment, baseline versus 48 h post‐treatment, and 3 h post‐treatment versus 48 h post‐treatment timepoints.

3. Results

We acquired pre‐contrast (n = 5 repetitions) and post‐contrast (n = 55 repetitions) DCE‐MRF and DCE‐MRI data from 20 total mice with orthotopic 4 T1 breast cancer tumors at baseline, 3 h, and 48 h post‐CA4P treatment (120 mg/kg). Figure 1 shows representative DCE‐MRF generated R₁ maps obtained from a single mouse before contrast administration and 2, 10, and 21 min after Gadovist contrast injection across all three imaging timepoints (baseline, 3 h post‐treatment, and 48 h post‐treatment). Similarly, Figure 2 shows representative estimated R₁ maps from DCE‐MRI at the same scan repetitions across all three imaging timepoints. Figures 1 and 2 both demonstrate characteristic accumulation of the Gadovist contrast agent in the tumor rim relative to the tumor core over time. DCE‐MRF R₁ maps (Figure 1) reveal clear changes in contrast enhancement following CA4P treatment, with reduced contrast uptake, especially in the tumor rim, at both post‐treatment timepoints. In contrast, the DCE‐MRI R₁ maps (Figure 2) show reduced visualization of the CA4P‐induced perfusion changes across timepoints, with limited visual evidence of treatment induced changes in contrast enhancement.

FIGURE 1.

Representative R₁ relaxation rate maps of the tumor ROI generated with DCE‐MRF from a single mouse before injection as well as at 2, 10, and 21 min post‐injection injection of Gadovist. These timepoints are replicated for all three imaging timepoints (baseline, 3 h post‐treatment, and 48 h post‐treatment). Note the increased accumulation of the contrast agent in the tumor rim relative to the tumor core over time. A representative paraspinal muscle ROI is denoted with the orange circle in the upper left image.

FIGURE 2.

Representative R₁ relaxation rate maps of the tumor ROI generated with DCE‐MRI from a single mouse before injection as well as at 2, 10, and 21 min post‐injection injection of Gadovist. These timepoints are replicated for all three imaging timepoints (baseline, 3 h post‐treatment, and 48 h post‐treatment). Note the increased accumulation of the contrast agent in the tumor rim relative to the tumor core over time.

Composite mean R₁ relaxation time curves were also generated for ROIs of the tumor rim, whole tumor, and paraspinal muscle regions of interest (ROIs) at each timepoint (Figure 3). DCE‐MRF demonstrated substantially greater R₁ increases following contrast administration across all ROIs and timepoints (Figure 3A,C,E) compared to conventional DCE‐MRI despite equal contrast agent injections (Figure 3B,D,F). Notably, DCE‐MRF demonstrated enhanced sensitivity to treatment‐induced changes in tumor vasculature. In the tumor rim ROI, DCE‐MRF clearly differentiated the baseline R₁ curve from the curves for both post‐treatment timepoints (Figure 3A), whereas DCE‐MRI showed limited separation between baseline, 3 h post‐treatment, and 48 h post‐treatment R₁ curves (Figure 3B). DCE‐MRF also exhibited reduced variation across the three imaging timepoints for the paraspinal muscle reference tissue (Figure 3E vs. Figure 3F), confirming reduced measurement noise in a relatively homogeneous tissue. As shown in Figures 3, S2, and Table S5, DCE‐MRF's improved R₁ precision results in reproducible dynamic R₁ assessments in the paraspinal muscles including a 3.1–6.6 fold observed increase in the change in the muscle R₁ before and after contrast injection. Note also that DCE‐MRF detected a significant shift in muscle contrast uptake at 3‐h post‐treatment that may be due to the acute of effects of CA4P.

FIGURE 3.

Following Gadovist administration, composite R₁ relaxation time courses were generated for (A, B) tumor rim, (C, D) whole tumor, and (E, F) paraspinal muscle regions of interest (ROIs) using (A, C, E) DCE‐MRF or (B, D, F) DCE‐MRI at all three imaging timepoints (baseline, 3 h post‐treatment, and 48 h post‐treatment). Note the enhanced sensitivity to treatment‐induced changes in tumor vasculature using DCE‐MRF relative to DCE‐MRI (A vs. B, C vs. D), and the enhanced precision of DCE‐MRF in the paraspinal muscle reference tissue (E vs. F). Error bars represent the standard deviation across all the respective DCE‐MRF and DCE‐MRI datasets.

Representative Linear Reference Region Model (LRRM)‐derived RK ^trans (Figure 4) and k _ep,T (Figure 5) are shown for both imaging methods across all timepoints. Figures 4A and 5A display RK ^trans and k _ep,T maps, respectively, from a representative animal imaged with DCE‐MRF at baseline, 3 h post‐treatment, and 48 h post‐treatment with 120 mg/kg CA4P. Correspondingly, Figures 4B and 5B show the same pharmacokinetic parameters from a different animal imaged with conventional DCE‐MRI at the same timepoints. At baseline, the RK ^trans and k _ep,T maps from both DCE‐MRF and DCE‐MRI demonstrate substantially elevated values in the tumor rim compared to the tumor core. This reflects the well‐established pattern of enhanced vascular perfusion in the tumor periphery relative to the often poorly perfused or necrotic central tumor regions [38, 44]. Following CA4P treatment, marked reductions in both RK ^trans and k _ep,T values are evident in the tumor rim at 3 h and 48 h post‐treatment, consistent with the known vascular disrupting mechanism of this agent [28, 29].

FIGURE 4.

Representative RK ^trans maps calculated from the Linear Reference Region Model (LRRM) are shown for both (A) DCE‐MRF and (B) DCE‐MRI methods across all three imaging timepoints (baseline, 3 h post‐treatment, and 48 h post‐treatment). At baseline, both DCE‐MRF and DCE‐MRI RK ^trans maps demonstrate substantially elevated values in the tumor rim compared to the tumor core. Following CA4P treatment, marked reductions in RK ^trans values are evident in the tumor rim at 3 h and 48 h post‐treatment.

FIGURE 5.

Representative k _ep,T maps calculated from the Linear Reference Region Model (LRRM) are shown for both (A) DCE‐MRF and (B) DCE‐MRI methods across all three imaging timepoints (baseline, 3 h post‐treatment, and 48 h post‐treatment). At baseline, both DCE‐MRF and DCE‐MRI k _ep,T maps demonstrate substantially elevated values in the tumor rim compared to the tumor core. Following CA4P treatment, marked reductions in k _ep,T values are evident in the tumor rim at 3 h and 48 h post‐treatment.

Quantitative analysis of the pharmacokinetic parameters for the tumor rim and whole tumor across all animals are presented in Figures 6 and 7. Box‐and‐whisker plots (Figure 6) display the distribution of RK ^trans and k _ep,T values for DCE‐MRF (Figure 6A,C,E,G) and conventional DCE‐MRI (Figure 6B,D,F,H) at each imaging timepoint. Additional descriptive statistics for the tumor ROI analysis are shown in Tables S1–S4. The baseline DCE‐MRF RK ^trans measurements showed median values 1.131 (Interquartile Range (IQR) = 0.373) for the tumor rim and 0.789 (IQR = 0.268) for the whole tumor. At 3 h post‐treatment, both the tumor rim and the whole tumor showed significant decreases (tumor rim: median = 0.288, IQR = 0.345; whole tumor: median = 0.121, IQR = 0.150; p < 0.0005 vs. baseline). By 48 h post‐treatment, the whole tumor demonstrated significant recovery of vascular function (median = 0.601, IQR = 0.374) compared to 3 h post‐treatment (p = 0.023), approaching baseline levels (p = 0.078 vs. baseline). DCE‐MRF k _ep,T values for both the tumor rim and whole tumor showed similar patterns, with baseline median values of 0.131 (IQR = 0.100) for the tumor rim and 0.122 (IQR = 0.064) for the whole tumor. At 3 h post‐treatment, both the whole rim and the whole tumor significantly decreased (tumor rim: median = 0.058, IQR = 0.008; whole tumor: median = 0.058, IQR = 0.005; p < 0.0005 vs. baseline). By 48 h post‐treatment, both the tumor rim and whole tumor demonstrated significant recovery of vascular function (tumor rim: median = 0.093, IQR = 0.027; whole tumor: median = 0.089, IQR = 0.031; p = 0.039 vs. 3 h post‐treatment; p = 0.023 vs. baseline). In comparison, DCE‐MRI RK ^trans measurements showed significant reductions at 3 h post‐treatment only for the whole tumor (baseline: median = 1.036, IQR = 0.338; 3 h: median = 0.278, IQR = 0.178; p < 0.008), with no significant changes observed in the tumor rim. At 48 h post‐treatment, neither the tumor rim (median = 1.163, IQR = 0.783) nor the whole tumor (median = 0.615, IQR = 1.087) showed significant changes compared to baseline or 3 h. Baseline DCE‐MRI k _ep,T values (tumor rim: median = 0.130, IQR = 0.048; whole tumor: median = 0.129; IQR = 0.028) decreased significantly at 3 h post‐treatment for both regions (tumor rim: median = 0.070, IQR = 0.024; whole tumor: median = 0.066, IQR = 0.013; p < 0.008 vs. baseline). At 48 h post‐treatment, neither region showed significant changes compared to baseline or 3 h post‐treatment (tumor rim: median = 0.098, IQR = 0.016; whole tumor: median = 0.096, IQR = 0.009). Overall, these results demonstrated that DCE‐MRF detected more significant differences in RK ^trans and k _ep,T across the tumor rim and whole tumor for the 3‐h and 48‐h post‐treatment timepoints in comparison to DCE‐MRI.

FIGURE 6.

Box‐and‐whisker plots showing the distribution of tumor rim (A, B) RK ^trans and (E, F) k _ep,T values and whole tumor (C, D) RK _trans and (G, H) k _ep,T values at each imaging timepoint assess by (A, C, E, G) DCE‐MRF and (B, D, F, H) conventional DCE‐MRI. (A) DCE‐MRF assessments of the tumor rim detected significant decreases in RK ^trans at 3 h post‐treatment compared to baseline (p < 0.0005), with a trend towards recovery at 48 h post‐treatment (p = 0.055 vs. baseline, p = 0.195 vs. 3 h post‐treatment) that did not reach statistical significance. (B) DCE‐MRI assessments of the tumor rim did not achieve significance at any timepoint (p $\ge$ 0.13). (C) DCE‐MRF assessments of the whole tumor detect significant decreases in RK ^trans at 3 h post‐treatment (p < 0.0005) relative to baseline, with a trend towards recovery at 48 h post‐treatment (p = 0.078 vs. baseline). Additionally, RK ^trans was significantly lower at 3 h compared to 48 h post‐treatment (p = 0.023). (D) DCE‐MRI assessments of the whole tumor detected significant decreases in RK ^trans only for the baseline to 3 h post‐treatment comparison (p < 0.008). (E) DCE‐MRF assessments of the tumor rim detected significant decreases in k _ep,T at both the 3 h post‐treatment (p < 0.0005) and 48 h post‐treatment (p = 0.023) timepoints relative to baseline, as well as the 3 h versus 48 h post‐treatment comparison (p = 0.039). (F) DCE‐MRI assessments of the tumor rim detected significant decreases in k _ep,T only for the baseline to 3 h post‐treatment comparison (p < 0.008). (G) DCE‐MRF assessments of the whole tumor detected significant decreases in k _ep,T at both the 3 h post‐treatment (p < 0.0005) and 48 h post‐treatment (p = 0.023) timepoints relative to baseline, as well as the 3 h versus 48 h post‐treatment comparison (p = 0.039). (H) DCE‐MRI assessments of the whole tumor detected significant decreases in k _ep,T only for the baseline to hours post‐treatment comparison (p < 0.008).

FIGURE 7.

Longitudinal mean tumor rim (A, B) RK ^trans and (E, F) k _ep,T and whole tumor (C, D) RK ^trans and (G, H) k _ep,T results at each imaging timepoint assessed by (A, C, E, G) DCE‐MRF and (B, D, F, H) conventional DCE‐MRI. Note the consistent trajectories with DCE‐MRF showing clear decreases at 3 h post‐treatment with trends towards recovery of vascular function at 48 h post‐treatment, relative to the substantially greater variability and sometimes divergent trajectories with DCE‐MRI for both the RK ^trans and k _ep,T parameters. Lines in each plot represent the longitudinal findings from individual mice.

To ensure the observed DCE‐MRF statistical significance was not an artifact of its larger sample size (n = 12 vs. n = 8 in DCE‐MRI), a bootstrap subsampling analysis with 10 000 iterations was performed. In each iteration, 8 DCE‐MRF animals were randomly selected to match the DCE‐MRI group composition (5 animals with all three timepoints and 3 animals with only baseline and 3 h post‐treatment timepoints). Wilcoxon signed‐rank tests comparing (1) baseline versus 3 h post‐treatment, (2) baseline versus 48 h post‐treatment, and (3) 3 h versus 48 h post‐treatment were performed on each subsampled dataset. For RK ^trans, the baseline versus 3 h post‐treatment effects were reliably sustained, achieving statistical significance in 100% of bootstrap iterations (mean p = 0.00781). In contrast, the baseline versus 48 h post‐treatment comparison was not significant in any iteration (mean p = 0.197), and the 3 h versus 48 h post‐treatment comparison was also not significant (mean p = 0.431, median p = 0.438). Similarly, for k _ep,T, the baseline versus 3 h post‐treatment comparison retained the significant difference achieving statistical significance in 100% of bootstrap iterations (mean p = 0.00781). In contrast, the baseline versus 48 h post‐treatment comparison was not significant in any iteration (mean p = 0.123, median p = 0.125), and the 3 h versus 48 h post‐treatment comparison was also not significant (mean p = 0.157, median p = 0.188). This confirms that DCE‐MRF's superior precision enables reliable detection of early treatment effects at 3 h even with reduced sample sizes matching the DCE‐MRI cohort, while changes at 48 h post‐treatment are not significantly different from baseline or 3 h post‐treatment values in this subsampled analysis.

Longitudinal perfusion trajectories from individual animals (Figure 7) provide additional insight into measurement reproducibility and treatment response patterns. DCE‐MRF measurements of RK ^trans for the tumor rim and whole tumor (Figure 7A,C, n = 12) demonstrate remarkably consistent baseline values across all animals, with nearly parallel trajectories showing clear decreases at 3 h post‐treatment with trends towards recovery of vascular function at 48 h post‐treatment. This uniformity of response is similarly evident in the k _ep,T trajectories (Figure 7E,G), where all animals exhibit coordinated decreases at 3 h following CA4P administration and a trend towards recovery of vascular function at 48 h post‐treatment. The tight clustering of individual traces reflects the improved precision and reduced inter‐animal variability of DCE‐MRF. In contrast, DCE‐MRI trajectories for both RK^trans (Figure 7B,D, n = 8) and k _ep,T (Figure 7F,H) show substantially greater variability and divergent patterns across animals. Several DCE‐MRI traces exhibit paradoxical increases at post‐treatment timepoints, while other animals demonstrated decreased perfusion or minimal changes, resulting in non‐parallel, intersecting trajectories that obscure the overall treatment effect. The substantially greater spread of individual DCE‐MRI traces, particularly evident for k _ep,T measurements, contrasts markedly with the consistent, concordant patterns observed with DCE‐MRF.

To quantify the overall capability of each imaging method to detect treatment‐induced changes in tumor vascular parameters, we calculated Hedges' g effect sizes comparing baseline measurements to post‐treatment timepoints and between the two post‐treatment timepoints (Table 1). For tumor rim RK ^trans measurements, DCE‐MRF demonstrated substantially larger effect sizes compared to conventional DCE‐MRI at both the 3 h post‐treatment timepoint (Hedges' g = 2.58 vs. 0.83) and 48 h post‐treatment timepoint (Hedges' g = 1.18 vs. 0.23) compared to baseline, representing approximately 3.1‐fold and 5.1‐fold larger effect sizes, respectively. The 3 h versus 48 h post‐treatment comparison also showed a larger effect size for DCE‐MRF (Hedges' g = 1.27 vs. 0.54), representing a 2.4‐fold larger effect size than DCE‐MRI. For whole tumor RK ^trans measurements, DCE‐MRF also demonstrated larger effect sizes at 3 h post‐treatment relative to baseline (Hedges' g = 2.65 vs. 2.04), 48 h post‐treatment relative to baseline (Hedges' g = 0.85 vs. 0.34), and in the 3 h vs. 48 h post‐treatment comparison (Hedges' g = 1.80 vs. 1.42) representing a 1.3‐fold, 2.5‐fold, and 1.3‐fold larger effect sizes, respectively. Similarly, for tumor rim k _ep,T measurements, DCE‐MRF yielded effect sizes of 2.46 (3 h post‐treatment) and 1.42 (48 h post‐treatment) relative to baseline, whereas DCE‐MRI produced smaller effect sizes of 1.88 (3 h post‐treatment) and 0.86 (48 h post‐treatment) versus baseline, corresponding to a 1.3‐fold and 1.7‐fold larger effect sizes, respectively. The 3 h versus 48 h post‐treatment comparison showed effect sizes of 1.72 for DCE‐MRF and 0.89 for DCE‐MRI (1.9‐fold larger effect size). For whole tumor k _ep,T measurements, both methods showed large effect sizes at 3 h post‐treatment (DCE‐MRF: 2.51; DCE‐MRI: 2.68) and 48 h post‐treatment (DCE‐MRF: 1.29; DCE‐MRI: 1.36) relative to baseline, though DCE‐MRF demonstrated a larger effect size in the 3 h versus 48 h post‐treatment comparison (2.24 vs. 1.14, representing a 2.0‐fold larger effect size). These results demonstrate that the enhanced precision of DCE‐MRF translates directly to improved sensitivity for detecting treatment‐induced vascular changes, with 1.3‐to‐5.1‐fold greater effect sizes for RK ^trans across all regions and comparisons, and 1.3‐to‐1.9‐fold greater effect sizes for tumor rim k _ep,T compared to conventional DCE‐MRI.

TABLE 1.

Hedges' g effect sizes for mean RK ^trans and k _ep,T at baseline versus 3 h post‐treatment, baseline versus 48 h post‐treatment, and 3 h versus 48 h post‐treatment for both DCE‐MRF and DCE‐MRI methods.

	DCE‐MRF	DCE‐MRI	DCE‐MRF/DCE‐MRI
	Hedges' g effect size	Hedges' g effect size	Effect size ratio
Tumor rim: RKtrans (a.u.)
Baseline vs. 3 h	2.58	0.83	X3.11
Baseline vs. 48 h	1.18	0.23	X5.13
3 h vs. 48 h	1.27	0.54	X2.35
Whole tumor: RKtrans (a.u.)
Baseline vs. 3 h	2.65	2.04	X1.30
Baseline vs. 48 h	0.85	0.34	X2.50
3 h vs. 48 h	1.80	1.42	X1.27
Tumor rim: kep,T (min−1)
Baseline vs. 3 h	2.46	1.88	X1.31
Baseline vs. 48 h	1.42	0.86	X1.65
3 h vs. 48 h	1.72	0.89	X1.93
Whole tumor: kep,T (min−1)
Baseline vs. 3 h	2.51	2.68	X0.94
Baseline vs. 48 h	1.29	1.36	X0.95
3 h vs. 48 h	2.24	1.14	X1.96

Figure 8 shows box‐and‐whisker plots comparing the model‐independent IAUC assessments (first 92 s post contrast injection) from DCE‐MRF (Figure 8A,C) and DCE‐MRI (Figure 8B,D) for the tumor rim and the whole tumor. Consistent with the pharmacokinetic assessments, DCE‐MRF IAUC values showed significant reductions from baseline at 3 h (tumor rim: p < 0.0005; whole tumor: p < 0.0005) and 48 h (tumor rim: p < 0.008; whole tumor: p = 0.016) post‐treatment. In comparison, DCE‐MRI showed a significant IAUC reduction only for the whole tumor at 3 h post‐treatment (p = 0.023), with no other significant changes observed. Effect size analysis for model‐independent IAUC assessments demonstrated that DCE‐MRF provided greater sensitivity for detecting treatment‐induced changes compared to DCE‐MRI (Table 2). For tumor rim IAUC measurements, DCE‐MRF showed larger effect sizes at 3 h (Hedges' g = 2.85 vs. 0.81) and 48 h (Hedges' g = 2.71 vs. 0.62) post‐treatment, representing 3.5‐fold and 1.9‐fold larger effect sizes, respectively. The 3 h versus 48 h post‐treatment comparison showed minimal change for DCE‐MRF (Hedges' g = 0.05) compared to DCE‐MRI (Hedges' g = 0.65). For the whole tumor IAUC measurements, DCE‐MRF demonstrated larger effect sizes at 3 h post‐treatment (Hedges' g = 1.67 vs. 1.28), 48 h post‐treatment (Hedges' g = 1.04 vs. 1.00), and in the 3 h versus 48 h post‐treatment comparison (Hedges' g = 1.51 vs. 0.45), representing 1.3‐fold, 1.0‐fold, and 3.4‐fold increases in effect sizes, respectively.

FIGURE 8.

Box‐and‐whisker plots showing the distribution of tumor rim (A, B) and whole tumor (C, D) initial area under the curve (IAUC) for the first 92 s post contrast injection at each imaging timepoint assess by (A, C) DCE‐MRF and (B, D) conventional DCE‐MRI. (A, C) DCE‐MRF IAUC values showed significant reductions from baseline at 3 h (tumor rim: p < 0.0005; whole tumor: p < 0.0005) and 48 h (tumor rim: p < 0.008; whole tumor: p = 0.016) post‐treatment. (B, D) DCE‐MRI showed a significant IAUC reduction only for the whole tumor at 3 h post‐treatment (p = 0.023), with no other significant changes observed.

TABLE 2.

Hedges' g effect sizes for IAUC assessments at baseline versus 3 h post‐treatment, baseline versus 48 h post‐treatment, and 3 h versus 48 h post‐treatment for both DCE‐MRF and DCE‐MRI methods.

	DCE‐MRF	DCE‐MRI	DCE‐MRF/DCE‐MRI
IAUC (mM*min)	Hedges' g effect size	Hedges' g effect size	Effect size ratio
Tumor rim: IAUC 92 s post‐injection
Baseline vs. 3 h	2.85	0.81	X3.52
Baseline vs. 48 h	2.71	1.43	X1.90
3 h vs. 48 h	0.05	0.62	X0.08
Whole tumor: IAUC 92 s post‐injection
Baseline vs. 3 h	1.67	1.28	X1.30
Baseline vs. 48 h	1.04	1.00	X1.04
3 h vs. 48 h	1.51	0.45	X3.36

4. Discussion

In this study, we directly compared the new DCE‐MRF methodology [21] with conventional DCE‐MRI to evaluate their respective capabilities for longitudinal treatment monitoring in an in vivo preclinical breast cancer model. Using the known vascular disrupting agent CA4P [28], we demonstrated that DCE‐MRF's inherently improved precision translates to enhanced statistical power for detecting drug induced vascular perfusion changes. Specifically, DCE‐MRF achieved 1.3–5.1 times larger effect sizes for RK ^trans across all regions and comparisons and 1.3–1.9 times larger effect sizes for tumor rim k _ep,T compared to conventional DCE‐MRI. DCE‐MRF demonstrated superior statistical detection of treatment‐induced vascular changes, achieving significance for RK ^trans at the early 3 h post‐treatment timepoint in both tumor rim and whole tumor regions and detecting significant vascular recovery in the whole tumor (3 h vs. 48 h), while also showing k _ep,T significance across all timepoint comparisons in both regions, whereas DCE‐MRI only detected significant changes in the acute response at 3 h post‐treatment despite similar median parameter changes. Therefore, this study establishes DCE‐MRF as a more reliable and sensitive quantitative imaging biomarker for assessing therapeutic response, with potential to overcome the reproducibility limitations that have hindered clinical translation of conventional DCE‐MRI perfusion imaging.

The enhanced sensitivity of DCE‐MRF was evident in the R₁ curves of the tumor and the paraspinal muscle reference tissue (Figure 3). DCE‐MRF's improved sensitivity enabled clear separation of baseline and post‐treatment time courses in the tumor rim and the whole tumor (Figure 3A,C), while conventional DCE‐MRI demonstrated a more muted effect (Figure 3B,D). As noted above, DCE‐MRF also detected subtle changes in the uptake of the contrast agent in the muscle reference tissue at 3‐h post‐treatment (Figure 3E) which may provide an opportunity for further refinements of the LRRM model. DCE‐MRF also demonstrated superior statistical power compared to DCE‐MRI in assessing tumor perfusion. Across all the comparisons of RK ^trans for both the tumor rim and the whole tumor (Table 1), DCE‐MRF demonstrated a 1.3‐to‐5.1 fold increase in the effect size between each of the imaging timepoints in comparison to conventional DCE‐MRI. The measured effect sizes of k_ep were also generally increased for DCE‐MRF relative to DCE‐MRI, especially for the tumor rim (1.3‐to‐1.9 fold). The effect sizes for k_ep were more similar across the whole tumor (0.9‐to‐2.0 fold), which may be partially due to tumor heterogeneity. The bootstrapping subsampling analysis confirmed these advantages arise from the improved precision of DCE‐MRF rather than the larger sample size. When matching the DCE‐MRI sample size (n = 8), DCE‐MRF maintained significance for the baseline versus 3 h post‐treatment comparison for both parameters (RK ^trans and k _ep,T: mean p = 0.0078, median p = 0.0078) in 100% of 10 000 bootstrap iterations. The loss of statistical significance for comparisons involving the 48 h post‐treatment timepoint in the subsampled data likely reflects the reduced statistical power from fewer animals at this timepoint (n = 5 vs. n = 8 at 3 h post‐treatment) and the well‐known waning of CA4P's vascular disrupting effects over time [28]. Regardless, these findings confirm that acute vascular changes at 3 h post‐treatment are robustly detectable with DCE‐MRF even at small sample sizes.

The majority of prior DCE‐MRI studies use the Tofts model or its variants to measure K ^trans and k _ep in a tumor by comparing tumor uptake kinetics to an Arterial Input Function (AIF) from a nearby blood vessel [8, 13]. However, localizing a blood vessel in an MR image is challenging, and movement of the blood vessel during the DCE‐MRI scans can produce significant errors in the AIF estimate [45, 46]. The Linear Reference Region Model (LRRM) used in this study is an alternative to the Tofts model, uses a muscle region as a reference instead of the AIF in the pharmacokinetic model, and results in estimates RK ^trans (a ratio of K ^trans for tumor vs. muscle) and k _ep [40, 41]. Localizing muscle tissue in an MR image is much easier than finding a blood vessel, and muscle is much less susceptible to movement during DCE‐MRI scans. We used the linearized version (LRRM) to analyze both datasets, which provides faster and more precise fitting [37, 38]. However, the lower muscular uptake of contrast agent compared to tumor blood vessels can result in errors in perfusion estimates, inhibiting clinical translation of the Reference Region Model. Overall, this study suggests that the enhanced precision of DCE‐MRF enables a more reliable detection of muscle R₁, thereby expanding the potential role of the LRRM in both preclinical and clinical DCE‐MRF studies.

DCE‐MRF, like conventional DCE‐MRI, is also adaptable to many pharmacokinetic models as well as model‐independent assessments of tumor perfusion. In addition to the LRRM assessments of RK ^trans and k _ep discussed above, the model‐independent IAUC measurements also demonstrated DCE‐MRF's improved capability to assess perfusion changes, especially in the tumor rim (Figure 8, Table 2). As mentioned above, the majority of prior DCE‐MRI studies use multi‐compartment Tofts models that estimate the AIF and therefore require shorter DCE‐MRF acquisition times. One approach to shortening the DCE‐MRF scan time would be to reduce or eliminate the 15‐s delay between repetitions that was included in this initial DCE‐MRF implementation as an abundance of caution to protect the gradient coil insert. Eliminating this 15‐s delay entirely would provide the opportunity to achieve 8‐s temporal resolution. The DCE‐MRF scan time could be further reduced by optimization of the spiral gradient waveforms providing shorter overall data acquisition windows and shorter MRF repetition times. These potential DCE‐MRF acquisition improvements, as well as the implementation of the corresponding DCE‐MRF dictionaries, would likely provide the opportunity to obtain AIF estimates and suggests the opportunity to utilize a wide variety of pharmacokinetic models used in many preclinical and clinical imaging research studies.

Beyond preclinical applications, several factors support the feasibility of DCE‐MRF for human imaging studies. While clinical scanners typically operate at lower field strengths (1.5–3 T) than the 9.4 T system used here, the substantially larger voxel volumes in clinical imaging (1–10 mm³ vs. 0.04 mm³) provide markedly higher SNR per voxel. Combined with shorter T₁ values in human tissues at clinical field strengths, this facilitates reductions in the number of MRF images and T₁ preparations required, further improving the temporal resolution and reducing the reconstruction complexity. These findings address a critical barrier to clinical translation of quantitative perfusion imaging by DCE‐MRF. Poor reproducibility of conventional DCE‐MRI, stemming from reliance on T₁‐weighted images to estimate T₁ rather than direct quantitative T₁ measurements, has limited clinical adoption despite decades of research [14, 15, 16, 17]. By directly measuring T₁ at each timepoint, DCE‐MRF eliminates assumptions required for T₁‐weighted signal conversions of conventional DCE‐MRI. At the same time, MRF has already demonstrated reduced susceptibility to errors and artifacts due to motion, B₁ inhomogeneities, coil positioning, temperature variations, and signal‐level dependencies [22, 47, 48]. This fundamental difference in measurement strategy enables the improved reproducibility necessary for reliable treatment monitoring. The increased effect sizes and robust detection of treatment response at small sample sizes positions DCE‐MRF as a viable quantitative imaging biomarker capable of overcoming the reproducibility challenges that have limited routine clinical use of conventional DCE‐MRI perfusion assessments.

Study limitations include the modest cohort sizes (n = 12 for DCE‐MRF, n = 8 for DCE‐MRI), though these were sufficient to demonstrate statistically significant differences and reveal fundamental precision differences between the two methods. Single‐slice imaging provided limited spatial tumor coverage, but enabled direct methodological comparison at matched resolution, isolating the impact of quantitative versus qualitative measurement strategies on precision. Additionally, the iterative LR reconstruction method, while necessary for achieved adequate SNR at the high spatial resolution ($195\mathrm{\mu m}$) required for preclinical imaging at 9.4 T, presents a significant computational limitation with reconstruction times of > 8 h for complete 60‐repetition DCE‐MRF datasets. However, this limitation is expected to be substantially reduced or eliminated in clinical applications where larger voxel sizes provide markedly higher SNR, potentially enabling accurate parameter estimation using simpler, faster reconstruction approaches. Importantly, multiple recent clinical MRF studies have shown that MRF data reconstructions can be achieved in near real‐time using available cloud computing architectures suggesting the opportunity for rapid translation for future clinical DCE‐MRF studies [49, 50, 51].

In conclusion, DCE‐MRF's improved precision translates directly into enhanced sensitivity for detecting treatment‐induced changes in tumor vascular perfusion. The 1‐to‐5‐fold improvement in RK ^trans effect sizes, 1‐to‐2‐fold improvement in tumor rim k _ep,T effect sizes, achievement of statistical significance where conventional DCE‐MRI failed, and markedly superior measurement reproducibility establish DCE‐MRF as a more reliable quantitative imaging biomarker methodology. These findings suggest DCE‐MRF has potential to overcome longstanding reproducibility challenges that have limited clinical translation of quantitative DCE‐MRI, enabling more robust and reliable assessments of treatment response for improved cancer patient care.

Funding

This work was supported by National Institute of General Medical Sciences (T32 GM007250); National Center for Advancing Translational Sciences (LR1 TR002549); National Cancer Institute (P30 CA043703, R01 CA269354); National Institute of Neurological Disorders and Stroke (R01 NS124206); National Institute of Biomedical Imaging and Bioengineering (T32 EB007509); American Heart Association (23PRE1017924).

Conflicts of Interest

Our research team has a master research support agreement with Siemens Healthineers. However, this support did not bias the presented results in any way.

Supporting information

Figure S1: Representative RK ^trans map calculated from the Linear Reference Region Model (LRRM) and the whole tumor relative residual error map. Example relative residual error thresholds (0.10, 0.15, 0.20, 0.25) applied to the whole tumor to differentiate tumor rim from the core based on pharmacokinetic model fit quality. Note the decreased relative residual errors near the tumor rim as better model fits typically occur in well‐perfused regions where contrast agent kinetics are more robust. Additionally, note the threshold level of 0.2 closely mirrors the tumor rim versus tumor core pattern in the RK ^trans map.

Figure S2: Box and whisker plots showing the mean (A, B) tumor rim, (C, D) whole tumor, and (E, F) paraspinal muscle R₁ values before (R_1,0) contrast and after contrast (R₁) administration. The pre‐injection R_1,0 was determined by averaging repetitions 2–4. The mean post‐injection R₁ value was determined by averaging the last 10 repetitions (repetitions 51–60). (A, C, E) Box and whisker plots from the DCE‐MRF data (n = 12) showed higher pre‐contrast R_1,0 values for the tumor rim, whole tumor, and paraspinal muscle at all three imaging timepoints (baseline, 3 h, and 48 h) than the post‐contrast R₁ values (p ≤ 0.003). (B, D, F) Box and whisker plots from the DCE‐MRI data (n = 8) showed higher pre‐contrast R_1,0 values for the tumor rim, whole tumor, and paraspinal muscle at all three imaging timepoints (baseline, 3 h, and 48 h) than the post‐contrast R₁ values (baseline and 3 h: p ≤ 0.014; 48 h: p = 0.063). However, the DCE‐MRF method showed more pronounced R₁ change following contrast administration in comparison to DCE‐MRI.

Table S1: Statistics for box and whisker plots of the mean tumor rim RK ^trans (a.u.) generated from the LRRM analysis for both the DCE‐MRF and DCE‐MRI data.

Table S2: Statistics for box and whisker plots of the mean tumor rim k _ep,T (min⁻¹) generated from the LRRM analysis for both the DCE‐MRF and DCE‐MRI data.

Table S3: Statistics for box and whisker plots of the mean whole tumor RK ^trans (a.u.) generated from the LRRM analysis for both the DCE‐MRF and DCE‐MRI data.

Table S4: Statistics for box and whisker plots of the mean whole tumor k _ep,T (min⁻¹) generated from the LRRM analysis for both the DCE‐MRF and DCE‐MRI data.

Table S5: Hedges' g effect sizes for pre‐contrast versus post‐contrast R₁ values for the tumor rim, whole tumor, and paraspinal muscle at all three imaging timepoints for both DCE‐MRF and DCE‐MRI.

Data Availability Statement

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