Initial Feasibility of Cardiac Diffusion-Relaxometry Spectroscopic Imaging (DRSI) for Subvoxel Tissue Characterization

Background: Due to the limited spatial resolution of conventional MRI, a single voxel often contains mixed signals from multiple tissue microenvironments.¹ Combined diffusion-relaxometry imaging allows probing of subvoxel compartments by varying multiple MR contrast encodings such as b-value, TI, and TE in a multidimensional acquisition.² When paired with appropriate modeling³, this enables correlation analysis between multiple MR parameters, such as diffusivity (D) and relaxation (R1 and R2). In this study, we implemented cardiac diffusion-relaxometry spectroscopic imaging (DRSI) and estimated voxel-wise tissue 3D D-R₁-R₂ spectra with a maximum entropy (MaxEnt) approach.

Methods: Two datasets were acquired on a 3T scanner (MAGNETOM Cima.X, Siemens Healthineers, Forchheim, Germany) from an ex-vivo swine heart with myocardial infarction (MI) (simulated RR=500 ms) and one healthy human subject (RR=894 ± 113 ms) under local IRB approval. Each dataset included 333 volumes across b ∈ [0, 50, 250, 500] s/mm² (0 or 6 directions, 1-3 averages), TE ∈ [40, 60, 80] ms , and five TIs ranging from 40 to 4160 ms (heart-rate dependent) (Figure 1). A slice-interleaved inversion-recovery⁴ single-shot M2 SE-EPI⁵ sequence was used with the following parameters: FOV = 300×300 mm², resolution = 2.7×2.7×8 mm³, BW = 2298 Hz/pixel, GRAPPA = 2, total scan time ≈ 47 min with RR=1000 ms. For in-vivo data, deformable motion correction was used to correct for in-plane motion.⁶
For each voxel, the directional-averaged signal was modeled as a sum of components with distinct R1, R2, and D values:
s(b,TI,TE)=∑_k=1 |(1-2e^(-TI×R_1,k⁾)×e^(-b×D_k^-TE×R_2,k⁾|,

where TI, TE, and b correspond to inversion time, echo time, and b-value, respectively. The number of components is denoted by K. To estimate the underlying spectrum, a maximum entropy (MaxEnt) method³ was used. This method introduced a dual formulation and measurement noise term for the spectral estimation by solving efficient convex optimization problems in a low-dimensional space.

Results: Figure 2 shows DR-CSI results from an ex-vivo MI swine heart. Two peaks were observed in the remote region, with one dominant spectral peak matching the conventional T1, T2, and mean diffusivity (MD) maps. In the infarcted region, multiple peaks were detected, with the dominant peak showing decreased R1 and R2 and elevated D, consistent with MI pathology. Figure 3 shows DR-CSI results in a healthy human subject. Three distinct components were identified in the LV, potentially attributed to intra-cellular (zero D), extra-cellular (moderate D, R1 and R2), and perfusion-related (D higher than free water) spaces.

Conclusion: This study demonstrates the feasibility of cardiac DRSI for imaging subvoxel microstructure using multidimensional MRI contrast. The technique captures heterogeneous tissue environments and may reveal novel imaging biomarkers. Further work is needed to optimize sampling schemes, reduce scan time, and validate findings across larger cohorts.

Figure 1: (A) Sequence diagram for inversion-recovery spin echo EPI sequence with M2 compensated diffusion gradients. (B) Slice-interleave acquisition for T1, T2 and diffusion joint encoding under free-breathing, 5 images at different locations (Slice #1 to #5) were acquired after application of a single nonselective inversion pulse with a delay time TIinit =40 ms for the first acquired slice. RR = duration of one heartbeat. A pause for 4 seconds was applied for longitudinal recovery. (C) Representative line profiles with the extracted position marked by the red dotted line after motion correction, showing robust motion correction. (D) Acquisition parameters for cardiac diffusion-relaxometry correlation spectroscopic imaging, with different TE, TI and b values. (E) Representative intensity profile from pixel marked as blue dot in (C) is shown.
Figure 2: (A) Traditional T1, T2, and MD maps acquired using the previously proposed sequence and acquisition protocol.4 Spatially averaged histograms of these parameters are shown for two regions of interest—scar and remote—as labeled on the T1 map. Clear shifts in the distributions were observed between scar and remote tissue, with the scar region showing elevated T1, T2, and MD values. (B) Spatially averaged DRSI results in the same two regions. In the remote (non-infarcted) region, two distinct well-resolved peaks were identified, and the dominant peak have D, R1, and R2 values comparable to those in the traditional maps. In contrast, the enhanced (scar) region exhibited multiple peaks in the DRSI spectra, with the dominant component showing reduced R1 and R2 and elevated D, consistent with MI pathology. Similar characteristics to remote tissue were observed, though it’s hard to distinguish from the conventional maps.
Figure 3: Spatially averaged 3D diffusion-relaxation (D–R1–R2) distributions estimated using the MaxEnt method for the left ventricle (LV) from a representative mid-ventricular slice for a healthy volunteer. Three distinct peaks were observed: extra-cellular (orange, moderate D, R1, and R2), intra-cellular (blue, zero D), and perfusion-related (red, very high D). (A) 3D plots of the distributions in D–R1–R2 space, with arrows indicating three primary tissue components. (B) 2D projections onto the D–R2, R1–R2, and D–R1 planes. The extra-cellular component resembles those observed in the remote zone of the ex-vivo heart and shows values comparable to T1, T2, and MD values in the healthy normal range. Elongation of extra-cellular component along the D-axis in the D–R1 projection suggests that the current diffusion sampling scheme is sparse and requires further optimization. The intra-cellular component observed in-vivo might be missing in the ex-vivo setting due to ultra-high R1, and perfusion component is unique for in-vivo scan only.