In Vivo Cardiac Diffusion-Weighted Magnetic Resonance Imaging

Abstract

Objectives:

Diffusion-weighted imaging (DWI) and the introduction of the intravoxel incoherent motion (IVIM) model have provided a unique method for evaluating perfusion and diffusion within a tissue without the need for a contrast agent. Despite its relevance, cardiac DWI has thus far been limited by low b values because of signal loss induced by physiological motion. The goal of this study was to develop a methodology for estimating IVIM parameters of in vivo cardiac magnetic resonance imaging using an efficient DWI acquisition framework. This was achieved by investigating various acquisition strategies (principal component analysis [PCA] filtering and temporal maximum intensity projection [PCATMIP] and single trigger delay [TD]) and fitting methods.

Material and Methods:

Simulations were performed on a synthetic dataset of diffusion-weighted signal intensity (SI) to determine the fitting method that would yield IVIM parameters with the greatest accuracy. The required number of b values to correctly estimate IVIM parameters was also investigated. Breath-hold DWI scans were performed for 12 volunteers to collect several TD values during diastole. Thirteen b values ranging from 0 to 550 s/mm² were used. The IVIM parameters derived using the data from all the acquired TDs (PCATMIP technique) were compared with those derived using a single acquisition performed at an optimized diastolic time point (1TD).

Results:

The main result of this study was that PCATMIP, when combined with a fitting model that accounted for T1 and T2 relaxation, provided IVIM parameters with less variability. However, an acquisition performed with 1 optimized diastolic TD provided results that were as good as those provided using PCATMIP if the R-R variability during the acquisition was sufficiently low (±5%). Furthermore, the use of only 9 b values (that could be acquired in 2 breath-holds), instead of 13 b values (requiring 3 breath-holds), was sufficient to determine the IVIM parameters.

Conclusions:

This study demonstrates that IVIM is technically feasible invivo and reports for the first time the perfusion fraction, f, and the diffusion coefficients, D and D*, for the cardiac DWI of healthy volunteers. Motion-induced signal loss, which is the main problem associated with cardiac DWI, could be avoided with the combined use of sliding acquisition during the cardiac cycle and image postprocessing with the PCATMIP algorithm. This study provides new perspectives for perfusion imaging without a contrast agent and demonstrates that IVIM parameters can act as promising tools to further characterize micro-vascular abnormalities or dysfunction.

Diffusion-weighted imaging (DWI) has been demonstrated over the last 2 decades to be a unique method for gaining access to the structural and physiological characteristics of tissues. It has been extensively developed for applications in the brain, notably for its ability to detect stroke earlier than other conventional magnetic resonance sequences can.¹ It is also largely used in neurological pathologies, where fiber architecture is assessed using diffusion tensor imaging.² Moreover, the introduction of the intravoxel incoherent motion (IVIM) model by Le Bihan et al³ gave rise to the ability to assess tissue perfusion and diffusion parameters without the need for a contrast agent. Indeed, signal loss in DWI is affected by mobile water molecules that are present in both the intracellular and extracellular compartments, as well as in blood flowing in capillaries and vessels. This model is therefore of wide interest for nonneurologic pathologies. For example, this technique has been used to assess the differentiation state of lesions in the liver,⁴ kidney,⁵ and pancreas⁶ and also to characterize cirrhosis.^7,8 Whereas the acquisition strategy is similar for abdominal organ applications, the data modeling and fitting methods vary considerably from one study to another. Recently, Lemke et al⁹ proposed a refinement of the DWI signal model proposed by Le Bihan et al³ that includes the T1 and T2 relaxation parameters of the blood and tissue as well as the echo time (TE)/repetition time (TR) sequence parameters and thus represents a new approach to data postprocessing.

Thus far, however, very few studies have investigated cardiac DWI, and the only study that evaluated cardiac perfusion and diffusion parameters in vivo using IVIM was performed in the canine heart.¹⁰ In fact, cardiac DWI is very challenging because cardiac motion can induce a large signal loss that, until now, has limited diffusion weighting to low b values. Therefore, strategies must be developed to minimize physiological motion during acquisition and to separate signal loss due to motion from that due to diffusion. Recently, Rapacchi et al¹¹ proposed a method to obtain cardiac DWI in which motion-induced signal loss was minimized by PCA filtering and temporal maximum intensity projection (PCATMIP) techniques. These acquisitions were performed while the subject was freely breathing and therefore necessitated a coregistration of the images before any additional analysis. However, other potential strategies include the use of respiratory triggering (PACE)¹² or breath-holds. As a result, the method of choice could be aligned to the technique that provides the most accurate perfusion and diffusion parameters. In this study, we have chosen to acquire data during subject breath-holds to avoid the potential source of signal variability resulting from the coregistration step. We compared the IVIM parameters obtained at a single optimized diastolic time point in the cardiac cycle with the results obtained with the PCATMIP technique.

The goal of this study was to propose a methodology for IVIM parameter estimation of in vivo cardiac magnetic resonance imaging (MRI) by developing an efficient DWI acquisition framework that could meet the requirements of routine clinical examination and by assessing the influence of various acquisition strategies and fitting methods on the estimation of IVIM parameters.

MATERIAL AND METHODS

Diffusion Models

Le Bihan et al³ first described signal behavior as a biexponential function of the perfusion and diffusion parameters. Actually, at low b values, a signal decays more rapidly because of the perfusion of water contained in microvessels and capillaries. At higher b values, the signal decay is caused by the isotropic diffusion of water into compartments where water is more constrained.

The signal decay as a function of the b value is thus defined as follows:

$$SI=S_0\left[\left(1-f\right)e^{-bD}+f{e}^{-b\left(D+D^{*}\right)}\right].$$ (1)

Here, SI represents the signal intensity, S₀ represents the SI for b = 0, f represents the perfusion fraction, D stands for the diffusion coefficient in the “constrained” compartment, and D* represents the diffusion coefficient from the perfused compartment.

However, this model (further referred as model 1) does not take into account the relaxation rates of blood and tissue, which, in some cases, can greatly influence the signal decay. Lemke et al⁹ proposed the following modified version of the signal model (further referred to as model 2) that accounts for the T1 and T2 relaxation rates:

$$SI=S_0\left[\frac{\left(1-\text{f}\right)\times \left(1-e^{-TR/T{1}_{\mathit{\text{tissue}}}}\right)\times e^{-TE/T{2}_{\mathit{\text{tissue}}}}\times e^{-bD}+f\times \left(1-e^{-TR/T{1}_{\mathit{\text{blood}}}}\right)\times e^{-TE/T{2}_{\mathit{\text{blood}}}}\times e^{-b\left(D+D^{*}\right)}}{\left(1-f\right)\times \left(1-e^{-TR/T{1}_{\mathit{\text{tissue}}}}\right)\times e^{-TE/T{2}_{\mathit{\text{tissue}}}}+f\times \left(1-e^{-TR/T{1}_{\mathit{\text{blood}}}}\right)\times e^{-TE/T{2}_{\mathit{\text{blood}}}}}\right].$$ (2)

Here, TE and TR represent the sequence parameters and T1_tissue/blood and T2_tissue/blood are the longitudinal and transverse relaxation parameters of the tissue and blood, respectively.

Fitting Method

In the literature, several different methods have been proposed for fitting the f, D, and D* IVIM parameters.^{4,5,7–10,13} To the best of our knowledge, no study has systematically compared these different methods and models to determine the optimal approach. We performed these comparisons with simulated DWI data and found that the method giving the IVIM parameters with the best accuracy was an extension of the method proposed by Lemke et al⁹ combined with the use of model 2 (see Appendix 1, Supplemental Digital Content 1, http://links.lww.com/RLI/A66). To fit the data, the signal-to-noise ratio (SNR) of the SI curves was enhanced by averaging all of the datasets acquired (mean of the SI for all subjects was performed for each b value), and the biexponential model 1 was applied to estimate D*. Note that in this case, we averaged SI/S₀ instead of SI. Subsequently, with known D*, the biexponential model 2 was applied to each dataset to estimate S₀, f, and D.⁹ In vivo data will be fitted with both models 1 and 2 to determine the added value of taking into account the relaxation parameters of blood and tissue, even though the best accuracy was obtained with model 2 in our simulations.

Evaluation of Different Acquisition Strategies

Because it has been demonstrated that motion can cause severe signal loss in DWI,¹⁴ we adopted the strategy proposed by Rapacchi et al,¹¹ which involves acquiring the DWI at different time points of the cardiac cycle over a predefined window of time. This time window was always located during diastole and corresponded to the minimal motion (in-plane and through-plane) of the left ventricle (LV) on the long and short axis cine. The trigger time was then spread at increments of 10-millisecond steps to acquire 10 DWI series. These data were then processed using the PCATMIP algorithm¹⁵ to separate signal loss due to motion from that due to diffusion. These images will be referred to as “PCATMIP” in the following discussion.

However, this approach required a high number of breath-holds (10 series with variable trigger delays [TDs] × 3 series with variable b values = 30 breath-holds), which was not realistic for a routine clinical context. For this reason, we determined an optimized diastolic TD during the cardiac cycle (defined as the TD that provides the maximum SI for DWI and corresponds to the minimal motion time), and we acquired the DWI series during the breath-hold at a single TD (method referred to as 1TD in the following discussion).

In Vivo Acquisition and MRI Parameters

The images were acquired using a MAGNETOM Avanto1.5 T (Siemens AG, Healthcare Sector, Erlangen, Germany). Twelve healthy volunteers were enrolled in the study, including 6 men and 6 women with a mean (SD) age of 28 (5) years (25–43 years) and a mean (SD) heart rate of 55.9 (9.5) beats/min. All subjects provided informed consent for the institutional review board–approved study protocol.

The imaging was performed on a midventricular short-axis slice of the LV myocardium. Before DWI acquisition, the standard 2- and 4-chamber and short-axis cines were acquired on the slice of interest. All DWI data were acquired in 3 orthogonal directions (except for b = 0) using a prototypical single-shot echo planar imaging sequence with optimized bipolar diffusion encoding gradients,¹⁶ improved fat suppression using a gradient reversal technique, a low-bandwidth excitation radio frequency pulse,¹⁷ and a global phase correction. The bipolar diffusion scheme was preferred to the monopolar approach because it has been shown to be more robust to eddy current artifacts and its optimized implementation in our study allowed a reduced TE compared with classical bipolar scheme.¹⁶ The SI measurements were performed on the trace image. At each TD, 3 DWI series were acquired during 3 separate breath-holds with increasing b values (the set of 13 b values was chosen based on Callot et al¹⁰): first series with b = 0, 15, 30, 45, and 60; second series with b = 0, 75, 90, 105, and 120; and third series with b = 0, 250, 350, 450, and 550. The acquisitions were performed at end-expiration breath-hold, which is well known to give more reproducible slice position than end-inspiration does. Moreover, the b = 0 image acquired at each breath-hold was visually inspected and compared with the others to ensure that the slice position was the same between the different breath-holds. If a change in the slice position was observed, the acquisition was performed again. The sequence parameters used for the study are listed in Table 1.

TABLE 1.

Parameters of the Sequences Used in the Study

Cine trueFISP	FOV, 287 × 400 mm	Breath-hold
	Matrix, 184 × 256	Acquisition time, ~20 s
	Resolution, 1.56 × 1.56 × 8 mm
	TE/TR, 1.49/44.7 ms
	Flip angle, 70 degrees
	22 cardiac phases
DWI	FOV, 236 × 420 mm	Breath-hold
	Matrix, 160 × 90	Acquisition time, ~16 s
	Resolution, 2.63 × 2.63 × 6 mm	b values, 0, 15, 30, 45, and 60 for the first series; 0, 75, 90, 105, and 120 for the second series; 0, 250, 350, 450, and 550 for the third series
	TE/TR, 56/110 ms
	Partial Fourier acquisition, 6/8 Bipolar diffusion scheme
T1 map	FOV, 272 × 340 mm	Acquisition time, ~ 15 s
	Matrix, 192 × 154
	Resolution, 1.77 × 1.77 × 8 mm
	TE/TI/TR, 1.06/100/900 ms
	Flip angle, 35 degrees
T2 map	FOV, 276 × 340 mm	Acquisition time, ~ 15 s
	Matrix, 192 × 156
	Resolution, 1.77 × 1.77 × 8 mm
	TE/TR, 1.12/225 ms
	Flip angle, 70 degrees

TI indicates inversion time.

T1 and T2 mapping was performed using respectively an electrocardiogram-triggered Modified Look-Locker Inversion Recovery^18,19 sequence and a T2-prepared TrueFISP sequence.²⁰

Image Analysis

Image postprocessing was performed using MATLAB (R2010a; The Mathworks, Inc, Natick, MA). The SI was measured in the LV after the manual segmentation of the myocardium for DWI with b = 0. The region of interest (ROI) was limited to the mid-wall area, avoiding the papillary muscles and the edge pixels to minimize the impact of partial volume effects. An ROI was also manually drawn in the visible part of the liver, avoiding the large vessels. The IVIM parameters obtained for the liver were used as a validation of the method because these results could be compared with previously published values.

The previously described,¹¹ the PCATMIP algorithm was used with the set of 10 repetitions (acquired with time-shifts of 10 ms) for each b value. Briefly, a spatiotemporal block-wise filter was first applied to the images to enhance their SNR, and this was achieved using a PCA filter with a boxcar of 15 × 15 pixels. Then, the TMIP was applied to the filtered images to obtain a motion-independent DWI (see Rapacchi et al¹¹ and Pai et al¹⁵ for the detailed algorithm).

The IVIM parameters were derived in 2 steps. First, D* was obtained for the 12 volunteers using either the mean SI value measured in the LV or the liver ROI; subsequently, the maps of the f and D parameters were calculated by representing the fit results on a pixel-by-pixel basis. The final f and D parameters were taken as the mean of the pixels within the LVor liver ROI from the previously computed maps.

The images were also evaluated in SNR terms. The SNR was calculated as the mean SI in the LVor liver ROI divided by the standard deviation of the SI in the ROI. The given value was the mean obtained on all b values. We chose this approach to evaluate SNR because the reduced field of view makes it difficult to have images containing a large signal-free region. Although it has been shown that this method can provide biased SNR owing to the combined used of phased-array coils and GRAPPA reconstruction,²¹ this bias would be the same for all the measurements and would not prevent a comparison between 1TD and PCATMIP SNR for the same ROI.

Considering the data dispersion, we hypothesized that it would be possible to estimate f, D, and D* with the same precision but with fewer b values, which would lead to a shorter acquisition time. This was investigated with simulations (see Appendix 1, Supplemental Digital Content 1, http://links.lww.com/RLI/A66), and we found that an acceptable compromise would be the use of 9 b values that are achievable in 2 breath-holds instead of 3 (b = 0, 30, 60, 75, 90, 120, 250, 450, and 550). In the following analysis, the in vivo data were also fitted with this subset (n = 9) of b values.

The correlations between estimated IVIM parameters and relaxation parameters T1 and T2 as well as with heart rate were also evaluated. For more details, see Appendix 2 (Supplemental Digital Content 2, http://links.lww.com/RLI/A67).

Estimation of RR Cycle Variability

Because the SI of cardiac DWI is sensitive to cardiac motion, we estimated RR cycle variability during the DWI acquisition. The measured RR cycle was the interval between the successive triggers of the physiological monitoring unit of the MRI (sometimes different from “true” RR because of gradient interference that occasionally altered the electrocardiogram signal). However, the measured cycle corresponded to the true TR applied during the sequence (note that the TR of the pulse sequence was typically fixed before the start of the acquisition, but the true TR corresponded to the time between 2 successive trigger events). For each sequence, we measured the mean RR interval as well as the corresponding standard deviation. Two measurements were particularly important; the first concerned the RR variability during the DWI scans (intrascan variability), and the second concerned the RR variability between scans used in the measurement set (interscan variability). Because several scans were used to measure the SI curve versus the b values (3 scans for 1TD and 30 scans for PCATMIP), the intrascan variability was taken as the maximum value of the RR standard deviation, whereas the interscan variability was defined as the maximum difference of the RR interval (ΔRR) obtained from the relevant scans. The values were reported relative to the mean RR interval to express the coefficient of variation (CV) as follows:

$$\begin{gathered} {\text{CV}}_{\text{intrascan}}=\frac{{\left\{\text{SD}\left(\text{RR}\right)\right\}}_{\text{max}}}{\text{mean}\left\{\text{RR}\right\}}, \\ {\text{CV}}_{\text{intrascan}}=\frac{{\left\{\Delta \text{RR}\right\}}_{\text{max}}}{\text{mean}\left\{\text{RR}\right\}}, \end{gathered}$$ (3)

Statistics

All data are represented as the mean ±SD. All statistical analyses were performed using the Aabel 3 software (Gigawiz Ltd Co, OK). P values < 0.05 were considered statistically significant.

Because the SNR cannot describe the dispersion of SI measurement as a function of b value for a single pixel in the image, we evaluated the standard deviation of fitted data (ie, SI curve measured in LVor liver vs b value for a single pixel in the image) by using the definition of the sample standard deviation²²:

$$s=\sqrt{\frac{1}{N-m}\sum {\left(y_i-y_{\text{fit}}\right)}^2},$$ (4)

where N is the number of measurements (in our case 13, for the b values), m is the number of parameters determined from the fit (m = 3 in our case), y_i is the measured data (SI here), and y_fit is the expected value of SI as evaluated with the fit; Σ_i(y_i − y_fit)² would be the residuals of the fit. In the following results and discussions, the results will be detailed as s/S₀ (SI for b = 0) to give relative results.

RESULTS

Figure 1 shows representative DWI data that were obtained for 1TD and after PCATMIP for 1 volunteer. Table 2 shows the mean SNR measured in each volunteer in the LVand liver ROI for 1TD and PCATMIP, as well as the relative mean sample standard deviation s/S₀ for these ROIs. With PCATMIP, SNR is enhanced compared with 1TD (P = 0.002 for LV and P = 0.0009 for liver), whereas s/S₀ is significantly reduced (P = 0.0002 for LV and P = 0.0008 for liver). If we compare our in vivo measurements and the simulation results concerning the accuracy of IVIM parameters estimation in function of the data dispersion s (see Appendix 1, Supplemental Digital Content 1, http://links.lww.com/RLI/A66), we obtain the following relative errors on f, D, and D* (σ_f, σ_D, and σ_D*, respectively); for LV and 1TD, σ_f = 1.17 and σ_D = 0.40 for corresponding s/S₀ = 10% and σ_D* = 0.11 for s/S₀ = 1.5%. Note that D* is not estimated on each pixel of each volunteer like f and D are but is derived from the mean SI of all the volunteers in the entire LV ROI; because of this, the data dispersion is therefore reduced to 1.7% in our experiments. For PCATMIP, σ_f = 0.57 and σ_D = 0.22 for corresponding s/S₀ = 5% and σ_D* = 0.11 for s/S₀ = 1.5% (measured dispersion on mean SI of all volunteers was 1.37%), corresponding to a reduction of 51% and 45% of the relative error, compared with 1TD, for f and D, respectively (σ_D* being the same).

FIGURE 1.

Trace DWI data for the different b values for 1TD (A) and PCATMIP (B). The chosen TD for panel A was the one that exhibited the highest SI for high b values (corresponding to minimal motion).

TABLE 2.

Mean SNR Measured in the LV and Liver for Each Volunteer and Relative Noise Level for 1TD and PCATMIP

Volunteer	LV				Liver
	1TD		PCATMIP		1TD		PCATMIP
	SNR	s/S0, %	SNR	s/S0, %	SNR	s/S0, %	SNR	s/S0, %
1	4.20	11.85	4.67	4.60	3.75	11.87	4.86	5.27
2	3.31	9.92	3.75	5.08	7.91	6.18	10.43	2.67
3	2.62	18.55	2.97	5.35	6.36	9.71	8.77	4.08
4	3.40	12.44	3.67	5.40	9.00	7.14	14.11	2.52
5	4.97	6.17	5.57	3.04	6.37	4.09	7.10	3.00
6	3.41	7.67	4.27	4.44	6.79	6.25	11.00	2.97
7	3.98	11.22	5.54	4.14	5.71	8.32	8.18	3.24
8	4.61	5.71	5.07	3.63	5.28	3.37	5.00	3.23
9	3.09	7.62	3.10	4.35	9.83	4.15	9.25	4.18
10	3.10	8.88	3.41	6.69	6.50	6.93	9.74	5.70
11	2.39	18.05	2.98	6.47	7.26	5.34	10.59	3.65
12	2.93	11.47	4.86	6.78	7.07	4.64	9.40	2.54
Mean	3.50	10.80	4.16*	5.00†	6.82	6.50	9.04†	3.59†

SNR is measured on the LV and liver ROI in each trace-DWI for the different b values and mean is performed on all the b values. Relative noise level corresponds to the sample standard deviation s measured in the SI curve (vs b value) for each pixel of the ROI; mean is then performed on s of all the pixels of the ROI and the result is reported on the mean S₀ of the studied ROI to give a percentage of noise. Last row gives the mean results as well as the significant difference between 1TD and PCATMIP.

^*P < 0.01.

^†P < 0.001.

For the liver and 1TD, σ_f = 0.79 and σ_D = 0.29 for s/S₀ = 6.5% and σ_D* = 0.08 for s/S₀ = 1% (measured dispersion on mean SI of all volunteers on liver ROI was 0.83%), whereas for PCATMIP, σ_f = 0.38 and σ_D = 0.15 for s/S₀ = 3.5% and σ_D* = 0.05 for s/S₀ =0.7% (measured dispersion on mean SI of all volunteers was 0.65%), corresponding to a reduction of 52%, 48%, and 38% of the relative error, compared with 1TD, for f, D, and D*, respectively. These results show that PCATMIP yields a more accurate estimation of IVIM parameters relative to the 1TD approach.

Figures 2 and 3 display the SI as a function of the b values as measured in all volunteers when acquiring at 1TD and after using PCATMIP for the LV and liver, respectively. The LV data at 1TD were occasionally prone to large variability (eg, volunteers 3 and 11). In these cases, PCATMIP highly reduced the SI dispersion (ie, physiological noise) and permitted the recovery of some of the signal that was lost due to cardiac motion. Concerning these 2 volunteers, we observed that the differences in the results obtained between 1TD and PCATMIP were both correlated with an intrasequence RR interval variability that was greater than 5% for acquisitions at 1TD (CV_intrascan = 6.8% and 6.5% for volunteer 3 and 11, respectively), which supports the hypothesis of motion-related SI loss. However, intersequence RR variability seemed to be less of an issue, as volunteers with CV_interscan values of 13.7% (volunteer 8) and as high as 23.1% (volunteer 4) did not exhibit such signal variability. Moreover, even when a high RR variability (ie, increased physiological noise) was observed within the data for the PCATMIP acquisition (CV_intrascan ranges from 4.8% to 61.9% and CV_interscan ranges from 5.4% to 50.7%), this method was robust enough to be independent of the RR irregularity. The SI values measured in the liver did not demonstrate differences between the 2 acquisition methods, which confirmed that the signal variability observed in the heart was primarily caused by cardiac motion and not to other artifacts.

FIGURE 2.

DWI myocardial SI as a function of the b values (n = 13) for an ROI covering the whole myocardium. The crosses indicate the values obtained for 1TD, and the dashed lines represent the fitted data. The open circles indicate the values measured from the PCATMIP images, and the plain lines represent the fitted data. The fits were performed with model 2.

FIGURE 3.

DWI liver SI as a function of the b values (n = 13) for an ROI covering the apparent portion of liver and avoiding large vessels. The crosses represent the values obtained for 1TD, and the dashed lines represent the fitted data. The open circles represent the values measured from the PCATMIP images, and the plain lines represent the fitted data. The fits were performed with model 2.

The maps of the f and D parameters are shown in Figure 4. The results were typically less noisy for the PCATMIP method than for the 1TD method. This suggests that on a pixel-wise basis, PCATMIP also allowed for spatial recovery of the signal that was lost because of cardiac motion. Figure 5 shows the results of the IVIM parameters that were obtained using either model 1 or 2 and either the full set of b values (n = 13) or the reduced set (n = 9), where the acquisition was allowed to be performed during 2 breath-holds instead of 3. Table 3 summarizes these results for the LV myocardium and liver.

FIGURE 4.

Trace DWI (b = 15 s/mm²) for 1 volunteer (left) and maps of the perfusion fraction, f, and the diffusion coefficient,D, derived with model 2 and 13 b values. The results are shown for the 1TD and the PCATMIP methods. The D* map is not shown because a single value was estimated based on the mean SI of the LV ROI of the 12 volunteers (see details in the “Methods” section).

FIGURE 5.

IVIM parameters for data acquired with 1TD and PCATMIP for the LV heart and the liver. The results are shown for model 1 (M1) and model 2 (M2) with both the full set of b values (n = 13) and the reduced set (n = 9), which could be acquired in 2 breath-holds instead of 3. The f and D represent the mean values of the LV or liver ROI that were obtained from the calculated maps, whereas D* was obtained from the mean SI of the 12 volunteers.

TABLE 3.

IVIM Parameters, f, D, and D*, Obtained for the LV Myocardium and the Liver With Models 1 and 2 and the 2 Different Acquisition Methods

Organ	Model	No. of b Values	Method	f	D, 10−3 mm2/s	D*, 10−3 mm2/s
LV	1	13	1TD	0.162 ± 0.075	2.57 ± 1.29	106.3
			PCATMIP	0.208 ± 0.057	2.44 ± 0.99	76.3
	2	13	1TD	0.117 ± 0.061	2.57 ± 1.28	106.3
			PCATMIP	0.150 ± 0.046	2.43 ± 0.98	76.3
	2	9	1TD	0.118 ± 0.056	2.55 ± 1.13	153.8
			PCATMIP	0.164 ± 0.047	2.34 ± 0.88	46.6
Liver	1	13	1TD	0.123 ± 0.048	1.53 ± 0.27	100.3
			PCATMIP	0.121 ± 0.037	1.49 ± 0.27	104.7
	2	13	1TD	0.106 ± 0.039	1.53 ± 0.27	100.3
			PCATMIP	0.103 ± 0.029	1.49 ± 0.27	104.7
	2	9	1TD	0.107 ± 0.038	1.54 ± 0.28	75.6
			PCATMIP	0.107 ± 0.029	1.49 ± 0.28	64.5

The results of model 2 were also estimated with 9 b values (achievable in 2 breath-holds) instead of 13 b values (achievable in 3 breath-holds).

The perfusion fraction was estimated with less standard deviation when the fit was performed with model 2 rather than with model 1. Because of the correction resulting from the introduction of tissue relaxation parameters, a lower value was obtained with model 2 (P < 0.001 when either 13 or 9 b values were used). As expected, the diffusion coefficients D and D* were identical between the 2 models. The estimation of f and D was not significantly different when 9 b values were used for 1TD; however, f was slightly enhanced (P = 0.001) and D was reduced (P = 0.027) in the case of PCATMIP. Moreover, with both models, we observed a reduction in the variability of f and D with PCATMIP, as compared to 1TD.

The perfusion fraction that was obtained for the liver exhibited a similar behavior, as the deviation in the results was smaller with model 2 and with the PCATMIP method. Similar to the heart, f was reduced when it was computed using model 2, as compared with model 1, when either 13 or 9 b values were used (P < 0.001), and D was not significantly different between the models. The parameters that were obtained using 9 b values were not significantly different, although there was an increased f value when PCATMIP was used (P = 0.005). However, similar to the heart, the standard deviations of f and D were the same for both the full set of b values and the reduced set (n = 9). This is an important result, which implies that acquisition time can be shortened by one-third, such that 2 breath-holds are sufficient.

DISCUSSION

Our simulations showed that the best fitting method was an extension of the method proposed by Lemke et al⁹ and that the use of only 9 b values (which can be acquired in 2 breath-holds) was sufficient to determine the IVIM parameters. In healthy volunteers, IVIM was technically feasible in vivo. The motion-induced signal loss could be minimized with the use of sliding acquisition during the cardiac cycle and the use of the PCATMIP algorithm for the postprocessing of the data.

Acquisition Method

We chose to perform DWI acquisition during breath-holds to minimize the acquisition time. This approach indirectly reduces motion-induced artifacts in the image (combination of RR variability and through-plane motion variability due to respiratory motion). In fact, although earlier studies had shown that cardiac DWI could be performed under free-breathing with subsequent image coregistration¹¹ or with respiratory triggering such as PACE,¹² these methods cannot retrieve signal loss due to through-plane heart motion.

The b values that were selected for acquisition were adapted from those of Callot et al,¹⁰ as these provided good coverage of the SI curve. Lemke et al²³ recently demonstrated an optimal distribution of b values that was able to minimize the error associated with IVIM parameter estimation, although the proposed sets contained too many b values, which were also too large, to be applicable to the heart with the current sequence (in most cases, we did not observe any SI in the myocardium with b values >600 s/mm²). The b values that were chosen allowed for the entire acquisition to be performed in a reasonable amount of time with adequate breath-hold durations.

Although the SI curves obtained at 1TD for most volunteers were of high fidelity, there was substantial data dispersion for some volunteers originating from motion-induced signal loss and largely caused by intrasequence RR variability. The first important result was that the PCATMIP technique clearly permitted the SI curve to be corrected for the observed dispersion. In these cases, the signal loss from 1 DWI could be compensated for during the next repetition of the sequence, which was shifted by 10 milliseconds. The shifted acquisition also allowed us to determine the time during the cardiac cycle that provided DWI with the least amount of signal loss due to motion, that is, the TD at which the DWI data should be acquired. Finally, these promising results and the observed robustness of PCATMIP lead us to suggest that it may be the solution of choice for further studies involving arrhythmic patients.

For subjects who presented an RR interval variability CV_intrascan value below 5% during acquisition, the images that were acquired at 1TD presented SIs that were similar to those derived from PCATMIP. Therefore, this parameter can act as a quality control for acquisition and could be used to determine whether repeated and shifted acquisition is required. Maps of the f and D parameters were clearly improved with PCATMIP, as compared with the 1TD acquisition. This suggests that on a pixel-wise basis, PCATMIP may compensate for signal loss due to regional cardiac motion. Note that regional abnormalities (eg, paradoxical motion in left bundle branch block or ischemia-related motion abnormalities) may occur without substantial RR variability, which would prevent the selection of a correct TD for the whole myocardium. In such patients, PCATMIP would be of crucial importance because it can compensate for regional signal loss due to motion.

Perfusion Fraction and Diffusion Coefficients

As expected, the cardiac diffusion coefficient value (D =2.43 ± 0.98 × 10⁻³ mm²/s) that was obtained with PCATMIP and model 2, which was the method that provided the least variability, was smaller than the apparent diffusion coefficient reported by Rapacchi et al¹¹ (7.1 × 10⁻³ mm²/s). However, it was greater than others that have been reported (in dogs,¹⁰ D = 1.26 ± 0.10 × 10⁻³ mm²/s; and in humans,²⁴ apparent diffusion coefficient = 1.55 ± 0.28 × 10⁻³ mm²/s). The pseudo-diffusion coefficient, D*, was also greater in the present study (76.3 × 10⁻³ mm²/s) than that measured previously for dog hearts¹⁰ (D* = 12.87 ± 2.56 10⁻³ mm²/s) but was similar to values reported for liver⁷ (79.1 ± 18.1 × 10⁻³ mm²/s) and spleen²³ (143 ± 127 × 10⁻³ mm²/s). However, we must note that our simulations yielded the highest error when D* was evaluated. Finally, the perfusion fraction, f, was greater when it was evaluated using model 1, as compared with model 2, which was expected because the rapid T2 relaxation of the myocardium tends to reduce the SI in addition to the decrease that is caused by the diffusion process. Consequently, f appears greater when the tissue relaxation correction is not taken into account. The value that we obtained with model 2 and PCATMIP (f = 0.150 ± 0.046) was slightly higher than the result reported by Callot et al¹⁰ (f = 0.0824 ± 0.0127); however, there was no correction for T2 relaxation in their model. In the rat heart, the fraction of the capillary volume in the myocardial tissue has been shown to be near 20% in diastole and near 14% in systole,²⁵ and our results are also within this range of values.

One explanation for the apparent discrepancies between our results and those published previously may involve the equilibrium state of the magnetization vector. Because we acquired the data at each RR to minimize the acquisition time, the signal recovery may not have been complete between 2 consecutive images, and the dynamic equilibrium of the magnetization vector would then have depended on the individual’s RR. In contrast, Callot et al¹⁰ acquired images over several RR cycles, which allowed the magnetization to completely recover between acquisitions.

Concerning the liver, the diffusion coefficients, D and D*, that were obtained in the present study (D =1.49 ± 0.27 10⁻³ mm²/s and D* = 104.7 × 10⁻³ mm²/s with PCATMIP and model 2) were in accordance with those previously reported⁷ (D = 1.10 ± 0.7 × 10⁻³ mm²/s and D* = 79 ± 18.1 × 10⁻³ mm²/s). Moreover, if we correct the perfusion fraction according to the relaxation parameters, T1 and T2 of the blood and liver (taken from²⁶), the reported f value decreases from 0.27 (published value estimated with model 1) to 0.11 (corrected value estimated with model 2), which is in accordance with our results (f = 0.103 ± 0.029).

Conclusions and Perspectives

This study demonstrates the technical feasibility of in vivo IVIM parameter measurement for human subjects. To our knowledge, this is the first study that has accessed the IVIM parameters of the human heart in vivo. The PCATMIP technique reduces the physiological noise in the IVIM curves and thus demonstrates its ability to minimize the motion-induced signal loss, which is typically the main problem associated with cardiac DWI. The experimental results confirmed the simulation results, suggesting that a reduced set of b values was sufficient for estimating parameters with identical dispersion. Finally, we have shown that acquisition performed at a single, optimized diastolic TD (1TD) provided results that were as effective as those delivered using PCATMIP if the RR variability during the acquisition was sufficiently low (<5%).

The estimation of IVIM parameters is of great interest because it provides a new method to estimate perfusion without the need for a contrast agent. Conceptually, the question of whether IVIM imaging could be used to estimate perfusion and its relevance have been debated,²⁷ although IVIM has been shown to be favorably comparable to perfusion measures that use conventional perfusion tracers.²⁸

In fact, for the case of increased perfusion, it would potentially be possible to separate the effect of increased capillary recruitment (increase in f) and increased blood velocity (increase in D*) using IVIM imaging because these parameters are indistinguishable with classical tracers. This study therefore provides new perspectives on the use of perfusion imaging without contrast media.