De-contrasted Image Registration Improves Quantification of Extracellular Volume and Fractional Myocardial Blood Volume

Abstract

Background

Accurate quantification of extracellular volume (ECV) and fractional myocardial blood volume (fMBV) in cardiac magnetic resonance (CMR) relies on precise alignment between pre- and post-contrast images. Variable image contrast often undermines conventional motion correction, causing misalignment due to respiration or cardiac motion. Herein, we present a registration approach that accounts for varying image contrast levels and cardiac motion to achieve more precise and high-quality quantitative cardiac mapping.

Methods

Patients with suspected myocardial diseases underwent cardiac MRI with Gadavist (0.1 mmol/kg, N=11) and ferumoxytol (4.0 mg/kg cumulative, N=9) enhancement for ECV and fMBV measurements, respectively. T1 maps were generated using the MOLLI sequence. To remove contrast variations across different inversion times and contrast doses, pre- and post-contrast MOLLI images were grouped and processed using correlation-weighted representations based on the myocardium and blood pool signals. Group-wise registration is performed based on the maximization of mutual information. The image registration accuracy and mapping precision of the proposed method were assessed relative to those of conventional methods.

Results

Compared with the conventional group-wise registration approach, the proposed de-contrasted approach showed superior alignment between images of different contrasts, as evidenced by the higher Dice scores (mean 0.77 vs. 0.69, p<0.001). It also eliminated artifacts commonly observed owing to image misalignment (all 11 cases showed improvement). Improved myocardial mapping precision was observed for both ECV (median coefficient of variation, 0.14 vs. 0.27; p<0.001) and fMBV (median coefficient of variation, 0.59 vs. 0.71; p<0.001). It also reduced individual myocardial segmental variations in the ECV (5.8 to 3.58, p<0.001) and fMBV maps (9.86 to 7.93, p<0.001).

Conclusion

Overall, de-contrasted image registration improves the precision of contrast-enhanced myocardial parametric mapping by reducing the misalignment between multi-contrast images. This framework may be extended to other post-processing tasks in cardiac MRI that involve variable image contrasts.

GRAPHICAL ABSTRACT

Myocardial extracellular volume (ECV) and fractional myocardial blood volume (fMBV) maps from pre- and post-contrast T1 imaging are prone to mis-registration. We propose a de-contrast preprocessing step to reduce contrast variance, enhancing group-wise registration. Applied to 11 ECV and 9 fMBV datasets, this approach improved accuracy and reduced map variance. Iterative de-contrasting effectively improves group-wise registration and parametric mapping quality in contrast-enhanced myocardium.

BACKGROUND

Contrast-enhanced cardiac magnetic resonance (CMR) imaging techniques allow the quantification of important physiological parameters, such as extracellular volume (ECV)[1] and fractional myocardial blood volume (fMBV)[2]. The ECV and fMBV reflect different myocardial compartments and are used as imaging biomarkers for the detection and monitoring of various pathological processes such as fibrosis, edema, and ischemia. ECV represents the proportion of extracellular tissue volume, including both extravascular and intravascular spaces. ECV is computed using pre- and post-contrast myocardial T1 values normalized for the hematocrit, and post-contrast images are obtained after the administration of an extracellular, non-protein bound gadolinium-based contrast agent. The computation of the ECV assumes that the change in the relaxation rate (ΔR1) between the pre- and post-contrast is proportional to gadolinium concentration. In contrast, fMBV represents the proportion of myocardial tissue volume occupied by the intravascular space and can be estimated using a 2-compartment water exchange model that considers the distribution of an intravascular contrast agent, such as ferumoxytol. For fMBV estimation, multiple myocardial and blood T1 values at different ferumoxytol contrast doses were used as inputs for the model. As an off-label contrast agent for CMR, ferumoxytol has a long blood pool half-life, excellent safety profile in patients with renal impairment, and enhanced sensitivity for detecting small vessels[2–4]. These benefits make it particularly valuable for patients with chronic kidney disease and for a detailed assessment of microcirculation.

Both current implementations of ECV and fMBV rely on the Modified Look-Locker Inversion (MOLLI) recovery pulse sequence[5] to compute pixel-wise myocardial and blood T1 parametric maps. The MOLLI sequence acquires multiple T1-weighted images at different inversion times, and the data are fitted to a T1 recovery curve to generate pixelwise T1 maps. Accurate quantification of the ECV and fMBV at the pixel level using T1 maps is highly dependent on the precise alignment between multiple acquisitions, which is difficult to achieve due to differences in image contrast and imperfect physiologic motion correction. Current approaches suffer from image misregistration due to variations in patient respiration and cardiac motion, which reduces the accuracy of quantitative parameter estimation[6]. Established motion correction techniques primarily focus on correcting motion within a single acquisition[7, 8]. These intra-sequence corrections do not adequately address motion differences between sequential acquisitions and lead to temporal misalignment of images acquired at different time points during an exam. The lack of effective motion correction methods for sequential image acquisition can lead to spatial misregistration of images, particularly between pre- and post-contrast acquisitions, which is required for the computation of ECV and fMBV.

Enhanced motion correction techniques that can provide robust alignment across multiple acquisitions and ensure reliable myocardial tissue characterization are needed to improve the accuracy of ECV and fMBV pixel-wise computation. One of the challenges for accurate image registration in contrast-enhanced acquisitions is the dramatic change in tissue contrast from pre- and post-contrast images, which often leads to failure in image registration[9]. Our correlation-weighted contrast separation is conceptually related to robust sparse decomposition approaches that disentangle slowly varying backgrounds from sparse contrast-changing components in DCE-MRI[10]. In this study, we propose an image de-contrasting and group-wise registration approach for myocardial parametric mapping. We hypothesized that a combined image processing step, including image contrast decomposition and groupwise registration, would reduce image misalignment and result in more precise ECV and fMBV estimation.

METHODS

This study was approved by the local Institutional Review Board, and all experiments adhered to the HIPAA regulations. A summary of the workflow is presented in Fig. 1.

Figure 1.

Image processing workflow. Generation of extracellular volume (ECV) or fractional myocardial blood volume (fMBV) maps. This process involves repeated cycles of de-contrast and registration from the original MOLLI image set.

Image acquisition

Difficult registration cases from two larger groups of patients with known or suspected myocardial disease were included in this study to test the proposed method. Underwent CMR using two different contrast agents. A group of 11 patients received gadobutrol (Gadavist; Bayer AG, Leverkusen, Germany), while another group of nine patients received ferumoxytol (Feraheme; Covis Pharmaceutical, Waltham, MA, USA). All scans were acquired using a clinical 3T magnet (Skyra; Siemens Healthineers, Ehrlangen, Germany). Gadobutrol was administered at 0.1 mmol/kg body weight for ECV, and ferumoxytol was administered at cumulative doses of 0.125, 2.0, and 4.0 mg/kg body weight for fMBV mapping[2]. T1 maps for fMBV measurements were acquired using the 5(3)3(3)3 Modified Look-Locker Inversion Recovery (MOLLI) sequence with a balanced steady-state free precession (bSSFP) readout. T1 maps for ECV measurements were acquired using 5(3)3 MOLLI for native imaging and 5(3)3(3)3 MOLLI for post-contrast imaging. The ECV map was derived from two T1 maps, whereas the fMBV map was calculated from four T1 maps [2, 11]. The detailed sequence parameters are listed in Table 1.

Table 1.

Acquisition parameters for the MOLLI T1 mapping sequence and contrast concentrations for ECV and fMBV mapping. bSSFP, balanced steady-state free precession; FOV, field of view; TE, echo time; TI, inversion time; TR, repetition time.

Acquisition Parameters	Parameter value (ECV maps)	Parameters value (fMBV maps)
FoV	380 × 285 mm2	360 × 274 mm2
Matrix size	256 × 192	192 × 156
In-plane resolution	1.48 mm × 1.48 mm	1.88 mm × 1.76 mm
TR	374.68 ms	376.82 ms
TE	1.09 ms	1.01 ms
Slice thickness	8 mm	8 mm
Pixel Bandwidth	1085 Hz/Px	1085 Hz/Px
Flip angle	20°	35°
TIs pre	139 – 218 ms (5(3)3)	100 – 260 ms (5(3)3(3)3)
TIs post	100 – 260 ms (5(3)3(3)3)	100 – 260 ms (5(3)3(3)3)
Gadavist concentration	0.1 mmol/kg	/
Ferumoxytol concentration	/	0.125, 2.0, 4.0 mg/kg
Readout type	bSSFP	bSSFP

Image De-contrasting

We first concatenated all frames required for registration into a single temporal stack: native (pre-contrast) MOLLI, post-contrast MOLLI, or every ferumoxytol-dose MOLLI series. Let the Total number of frames be $T_{all}$.

For each frame in this concatenated stack, we drew two ROIs: a mid-myocardial line ROI and a polygonal ROI in the LV blood pool (Figure S1). For each acquisition separately (native 5(3)3; post-contrast and for fMBV each ferumoxytol dose 5(3)3(3)3), we averaged within these ROIs frame-by-frame to produce two global temporal reference curves across the entire stack: one that predominantly reflects myocardial T1 recovery and one that reflects blood-pool kinetics (including dose-dependent changes).

Let $s\in {\mathbb{R}}^{T_{all}\times 1}$ denote the voxel-wise temporal signal, $A\in {\mathbb{R}}^{T_{all}\times M}$ The “dictionary” is built from the whole image set, and $\alpha \in {\mathbb{R}}^{M\times 1}$ The coefficient map. We estimate $\alpha$ by solving

$$\widehat{\alpha }=arg\underset{\alpha }{\text{min}}\frac{1}{2}\parallel s-A\alpha {\parallel }_2^2+\lambda \sum _{j=1}^2\parallel {\gamma }_{(\bullet ,j)}☉\alpha {\parallel }_1$$

Where $\lambda =0.1$ is the regulator, $\odot$ denotes element-wise multiplication, and $\left[{\gamma }_{(\bullet ,1)},\left.{\gamma }_{(\bullet ,2)}\right]\in {\mathbb{R}}^{M\times 2}\right.$ is a class-weighting matrix derived from the correlations between the two ROI curves. Specifically for each pixel $m=1,...,M$ and class $j\in \left\{\mathit{\text{myocardium}},\left.\mathit{\text{blood}}\mathit{\text{pool}}\right\}\right.$,

$${\gamma }_{m,j}=\left\{\begin{array}{cc}exp\left(-{\rho }_{m,j}/0.9\right),& {\rho }_{m,j}\ge 0\\ 1,& {\rho }_{m,j}<0\end{array}\right.$$

with ${\rho }_{m,j}$ The Pearson correlation between the temporal curve of pixel m and the corresponding ROI reference curve $j$. We used proximal gradient descent to solve this objective.

$$\tilde{s}=s-A\widehat{\alpha }$$

Conceptually, our residual $\tilde{s}$ Plays a role analogous to removing sparse contrast components, but the weights are ROI-driven from myocardium/blood temporal curves over the concatenated stack (pre, post, doses). These de-contrasted frames were then used in the registration stage, and the final transformations were applied to the original (non-separated) image.

Image Registration

After the single joint de-contrast step above, we register all de-contrasted frames to a common reference. Let $\left\{{\tilde{I}}_k\}\right.$ denote the de-contrasted images from every time point $k\in \{1,...,T_{all}\}$. We estimate transforms $\left\{T_k\right\}$ by maximizing the sum of mutual information (MI) to a single reference frame R,

$$arg\underset{T_k}{\text{max}}\sum _{k}MI(R,{\tilde{I}}_k\bullet T_k)$$

We chose the last TI frame of the post-contrast MOLLI as R, which we found to empirically reduce bias and avoid unstable reference updates. We use MI with 50 bins and all pixels because MI tolerates residual intensity offsets, and de-contrast suppresses systematic contrast differences, yielding a more stable objective for cross-acquisition alignment. We used 2D in-plane similarity transforms (rotation, translation, isotropic scale, no shear) with physiologic scale bounds. Optimization used a one-plus-one evolutionary strategy (growth factor 1.05; epsilon of 1.5×10⁻⁶; initial radius 1.79×10⁻³; max 1000 iterations). Final transforms are then applied to the original (non-separated) frames to build the registered stacks for T1/ECV/fMBV computation. All registrations were performed using MATLAB R2023a.

Statistical analysis

Data are summarized as mean and standard deviation (SD) or median and interquartile range (IQR). Normality was tested using the Shapiro-Wilk test. The parametric maps of the proposed method for the ECV and fMBV values were evaluated against those generated using traditional techniques. Two baseline approaches were included for comparison. The first baseline was groupwise registration without a de-contrast step, and the second baseline was de-contrast and registration performed separately for each MOLLI sequence. Statistical analysis was conducted using paired t-tests for normally distributed data and Wilcoxon tests for skewed distributions to assess the differences between the proposed approach and the two baseline approaches. Accuracy and reliability were assessed using the Dice score and coefficient of variance (CV). All statistical analyses were performed using MATLAB software (MATLAB R2023a). A subjective image score (0–2) was applied to all 11 ECV subject images, both with and without de-contrast images. In this scoring system, 0 indicates a clear image without noticeable artifacts, whereas 2 indicates that artifacts significantly impact image interpretation.

RESULTS

The introduction of the de-contrast step significantly improved the group-wise registration performance for multi-contrast images, particularly in the context of fMBV and ECV mapping. This de-contrast step minimizes the intensity variance between the pre- and post-contrast images, leading to more accurate co-registration across different contrast levels.

Improvement in registration

The proposed groupwise registration showed significant improvements with the implementation of a contrast-separation preprocessing step. Images registered with contrast separation (Figures 2A and 2B) had sharper border definitions (red arrows) and improved overall alignment between the moving and base images. The accuracy of image registration was assessed using the Dice similarity coefficient, which measures the overlap between the segmented myocardial regions in the registered images. The Dice scores (Figure 2C) for images that underwent contrast separation and groupwise registration were significantly higher than those without (mean 0.77 vs. 0.69, p <0.001), indicating superior alignment of the multi-contrast images.

Figure 2.

Groupwise registration with and without contrast separation pre-processing. For the extracellular volume (ECV) maps (A) and fractional myocardial blood volume (fMBV) maps (B), registration with contrast separation showed better alignment with the base image (magenta is the moving image; green is the static image). (C) shows a significant improvement in the Dice scores using the proposed registration method (p<0.001).

Improvement in quantitative parametric mapping

Reduction of artifacts. A similar improvement was observed in ECV mapping, where contrast separation minimized the misregistration artifacts commonly observed in conventional techniques (in all 11 instances, an enhancement was observed). The average subjective score with de-contrast was 0.27, while without de-contrast, it was 1.55). As shown in Figure 3, the conventional registration method produced dark-band artifacts at the myocardial border, particularly in post-contrast images (Figure 3B). These artifacts result from poor alignment between pre- and post-contrast images and significantly affect the accuracy of ECV quantification. The proposed approach mitigated these errors, resulting in sharper borders and fewer artifacts. The fMBV parametric maps generated using the proposed method (Figure 4B) displayed more homogeneous fMBV values with fewer artifacts, particularly in regions near the tissue boundaries.

Figure 3.

Coefficient of variation (CV) for extracellular volume (ECV) maps. (A) CV boxplots for all ECV values in 66 with and without de-contrast preprocessing and registration. Both the mean and variability of the CV values were smaller than those obtained without the preprocessing. (B) ECV maps with and without preprocessing and registration. The red arrow indicates a dark band caused by misregistration.

Figure 4.

Coefficient of variation for fractional myocardial blood volume (fMBV) map. A) Boxplots for all segments (144 in total) of the coefficient of variation (CV) of the fMBV with and without de-contrast pre-process registration. Both the mean and spread CV values were lower than those of their counterparts. B) fMBV maps of a ventral apical slice demonstrating reduced image artifacts with contrast separation and groupwise registration are shown.

Reduction in the ECV and fMBV coefficient of variance. The improvements introduced by the contrast separation step were further substantiated by the results of the quantitative analysis of variance. The CV for the ECV and fMBV maps with and without the proposed processing steps are shown in Figures 3A and 4A, respectively. With the proposed processing steps, the CV for ECV (median CV 0.14 vs. 0.27, p<0.001) and fMBV (median CV 0.59 vs. 0.71, p<0.001) were significantly lower, suggesting more consistent values across the myocardial segments.

Reduction in the ECV and fMBV standard deviations. The impact of contrast separation on reducing variability was further substantiated by analyzing the standard deviation of the segmental ECV and fMBV values. Figures 5 and 6 show the segmental standard deviations of the ECV and fMBV across the ventricular basal, mid, and apical slices using the American Heart Association’s 17-segment model. Compared to images registered without contrast separation, the standard deviation for both ECV and fMBV with contrast separation showed lower values across the myocardial segments (median decrease from 5.8 to 3.58 for ECV maps, p <0.001; median decrease from 9.86 to 7.93 for fMBV maps, p<0.001). These findings suggest that the contrast separation step not only improves the registration accuracy but also enhances the reproducibility of the parametric maps, which enhances the reliability of the measurements across all segments of the myocardium.

Figure 5.

Representative segmental ECV standard deviation at the mid-left ventricular myocardium. (A) provides the standard deviation of ECV values when pre- and post-contrast T1 MOLLI images are used to generate ECV maps with contrast separation and groupwise joint registration (the proposed approach). (B) Standard deviation of ECV values when pre- and post-contrast T1w MOLLI images are used to generate ECV maps without contrast separation. Note that the standard deviation for the segments in Part B is larger (p<0.001).

Figure 6.

Representative segmental fractional myocardial blood volume (fMBV) standard deviation at the base, mid, and apical left ventricular myocardium. (A) provides the standard deviation of fMBV values when pre- and post-contrast T1 MOLLI images are used to generate fMBV maps with contrast separation and groupwise joint registration (the proposed approach). (B) Standard deviation of fMBV values when pre- and post-contrast T1w MOLLI images are used to generate fMBV maps without contrast separation. Note that the standard deviation for the segments in Part B is larger (p<0.001).

Compared with the baseline de-contrast approach without groupwise registration, the proposed groupwise registration method yielded a higher level of alignment of the myocardium (myocardial Dice scores were 0.63 and 0.77 for the baseline and proposed methods, respectively, p <0.001), and lower myocardial segmental variation (CV was 0.23 and 0.14 77 for baseline and the proposed method, respectively, p <0.001) in the ECV data.

DISCUSSION

Sequential multicontrast images used for the estimation of myocardial parametric mapping, including ECV and fMBV, often suffer from misalignments during image registration. The results of the current study suggest improved accuracy in ECV and fMBV estimation when the image processing step includes contrast separation and groupwise registration. De-contrast reduced contrast-induced ambiguities in MI, mitigating failure modes that produced overly homogeneous fMBV and ECV/fMBV regions without de-contrast. Both the registration Dice score and precision of the ECV and fMBV maps improved with the inclusion of contrast separation. These results highlight the importance of performing image deconvolution registration in multicontrast studies.

Our approach differs from that proposed by Kellman et al.[12] It addresses both intra- and inter-sequence motion simultaneously while accounting for contrast differences. For ECV mapping, Kellman et al. developed an approach that separately provided motion correction within a single acquisition using the MOLLI sequence and between pre- and post-contrast images. The results were synthetic images free of motion from an initial T1 map and known inversion times. Each original MOLLI image was registered to its synthetic counterpart to correct intrasequence motion. For inter-sequence correction, the image with the longest inversion time was chosen as a reference, allowing robust registration because of the similar magnetization states. Xue et al[13] later proposed the use of phase-sensitive inversion recovery (PSIR) reconstruction to remove contrast inversion in magnitude images. After the background phase was removed, each frame was registered within a variational, non-rigid framework. However, preserving the T1 contrast in the MOLLI sequence may still influence the similarity-based registration metrics. In a separate line of argument, Tao et al. [14] applied a group-wise registration method based on PCA[15] to T1 and ECV mapping. Considering the T1 relaxation pattern across different T1-weighted images, they used the PCA eigenvalue spectrum to detect and correct motion-induced misalignments in the images. All pre- and post-contrast images were registered using a non-rigid B-spline transformation and PCA-based similarity matrix. Robust-PCA-based registration reduces contrast variability by modeling frame stacks as low-rank and sparse; however, in cardiac mapping with multiple acquisitions, short frame counts, and cross-acquisition/dose intensity shifts can challenge the low-rank assumption. Our joint, physiology-weighted separation uses myocardium and blood-pool temporal references to guide the removal of contrast-variant components across all frames at once, yielding a more uniform appearance for MI-based alignment. We observed that de-contrasting reduced MI ambiguities and improved interface sharpness. Our method differs from earlier approaches by eliminating contrast variation and then performing groupwise registration of the entire image dataset. Recently, Li et al.[16] used contrast separation but only addressed intra-acquisition motion in T1 mapping. Our work extends the concept proposed by Li et al. to parametric maps such as ECV and fMBV, where inter-acquisition motion is also present. Unlike methods that perform contrast separation within each acquisition and then attempt cross-acquisition alignment, our strategy performs a joint groupwise contrast separation across all frames spanning pre-contrast, post-contrast, and multiple doses. In practice, this unified contrast separation reduced overall image contrast variation within the entire image series and yielded a single, consistent target for image registration, which was found to reduce performance variation in challenging cases.

The observed improvements in quantitative cardiac MRI registration using our proposed approach may help classify myocardial diseases by enhancing precision. More precise ECV and fMBV maps with reduced registration artifacts can increase diagnostic accuracy. These improvements are particularly beneficial for conditions such as hypertrophic cardiomyopathy, heart failure with a preserved ejection fraction, cardiac amyloidosis, myocarditis, and coronary artery disease. The enhanced precision and reliability of the measurements allow for better disease monitoring, earlier detection of pathological changes, and a more decisive assessment of treatment efficacy.

This study had some limitations. First, it was conducted at a single center, which could introduce bias related to local imaging protocols and patient populations, and at a single field strength. Multicenter studies can help establish the broader applicability of the proposed method. Second, the de-contrasting approach showed improvements in image registration and parametric mapping compared with the baseline registration method. However, its performance has not been directly compared with that of other advanced registration techniques, such as PCA-based groupwise registration approaches. The proposed method incorporates the specific T1 signal model of the sequence and contrast dose kinetics of the experiment, which may outperform the PCA-based method that is completely data-driven. Future studies are warranted to confirm this hypothesis. Finally, this study focused on specific contrast agents and MRI sequences. Further investigation is needed to determine whether these benefits extend to images of other contrast mechanisms (such as CEST, T2 maps, or T1rho maps) and acquisition protocols commonly used in cardiac MRI. Future research could explore several promising avenues. Optimizing the computational efficiency of the de-contrasting algorithm could be pursued to facilitate its integration into routine clinical workflows. Furthermore, exploring the potential of this technique in other organ systems or imaging modalities could open new avenues for improving medical imaging analyses and diagnoses.

CONCLUSION

This study demonstrated significant improvements in the precision of ECV and fMBV myocardial parametric maps using the proposed de-contrast group-wise image registration. The results suggest that de-contrasting is an essential pre-processing step for multi-contrast image datasets, allowing for more accurate imaging registration.

Abbreviation List

fMBV

Fractional Myocardial Blood Volume

ECV

Extracellular Volume

CMR

Cardiac Magnetic Resonance

MOLLI

Modified Look-Locker Inversion

ROI

Region of Interest

SD

Standard Deviation

IQR

Interquartile Range

CV

Coefficient of Variance

PSIR

Phase-sanstive Inversion Recovery

PCA

Principal Component Analysis

Data Availability Statement

Data supporting the findings of this study are available from the corresponding author upon reasonable request. The implementation code can be accessed via our repository https://github.com/fredy1215/de-contrast-FMBV-ECV.git.