Introduction

Quantitative MRI techniques such as diffusion-weighted imaging (DWI) have diagnostic and prognostic potential in a variety of renal diseases such as chronic kidney disease, polycystic kidney disease, and transplant malfunction [1]. The diffusion-weighted signal is sensitive to multiple aspects of renal function and microstructure, including tubular and vascular flow/volume, and renal interstitium [2, 3].

Within the DWI literature, the kidney microstructure is less explored compared to other organs such as the brain or prostate [4]. This is in part due to image quality challenges from echo-planar imaging artifacts [5] in the presence of respiratory and cardiac motion [6], in addition to the effect of cardiac cycle pulsatility on the diffusion signal [3, 7,8,9,10,11,12]. By mitigating these artifacts exploration of the kidney microstructure might become progressively more feasible.

Macroscopically, the kidney anatomy is composed of the medulla and cortex, which are tissue types with distinct microstructures and microcirculation. Nephrons progressively perform blood filtration and reabsorption, beginning at vascular glomeruli, proceeding through the proximal convoluted tubules of the cortex, through aligned loops of Henle in the medulla, back through the distal convoluted tubules, and into the collecting duct [13]. Each of these aspects possesses features of flow, microstructural diffusion restricting features [14], anisotropy, and permeability [15], that impact water diffusion as measured in diffusion MRI.

Many studies employed variants of diffusion-weighted imaging (DWI) in vivo using intravoxel incoherent motion (IVIM), diffusion tensor imaging (DTI), and multi-component methods to characterize renal function [3, 6,7,8,9,10,11,12, 16,17,18,19,20,21,22,23,24,25,26,27,28]. IVIM [16] separates and approximates microcirculation (vascular and tubular flow) and microstructure. DTI [29] quantifies fractional anisotropy (FA) which is more prevalent in the oriented medullary bundles of tubules and collecting ducts than in the randomly oriented tubules and vasculature in the “cortical labyrinth” [19,20,21]. Several recent studies [24,25,26] also observed anisotropy of the flow component. Time-dependent diffusion MR methods similar to prior work in the prostate [30] allowed estimations of tubular diameters in combination with Monte Carlo diffusion simulations [3]. Finally, a multiple encoded FC-IVIM method [27] including a combination of conventional and flow-compensated diffusion gradient waveforms, multiple diffusion times, and Monte Carlo simulations model might enable a more detailed characterization of flow in the renal tissue [31].

Macroscopic segmentation strategies are evolving to better assess intra-organ functional variations. Piskunowicz et al. [32] performed semi-automated layered segmentations to measure ΔT2* values for blood oxygen level dependent (BOLD) imaging for chronic kidney disease (CKD) characterization. This was followed by a more automated 12-layer concentric object method (TLCO) [33], which enabled better differentiation of CKD kidneys from controls compared to the classical cortex and medulla segmentations [34]. Recently, the TLCO was applied to the analysis of apparent diffusion coefficient (ADC) [35] and arterial spin labeling (ASL) parameters [36].

In this work, we extended the concentrically layered segmentation to DTI and IVIM parameters of the healthy human kidney, as well as the multiple encoded FC-IVIM approach [27] to investigate their spatial dependency and hypothetically interpret tissue features at each concentric layer.

Methods

Imaging and preprocessing

In this HIPAA-compliant and IRB-approved prospective study, seven healthy volunteers (three male, age 26.8 ± 2.9 years, body mass index 24.2 ± 2.0) provided written informed consent prior to imaging. The volunteers underwent abdominal imaging in a 3 T MRI system (MAGNETOM Prisma; Siemens Healthcare, Erlangen, Germany) in supine position with posterior spine array (4–6 elements activated) and anterior 18 channel body matrix array receive RF coils, 4 ECG chest leads (Siemens Healthineers) for cardiac gating and scanner body coil RF transmission. Table 1 summarizes pulse sequence parameters employed in this study. Coronal oblique T2-weighted HASTE images were collected for anatomical reference. Sagittal phase-contrast (PC) MRI images through the left renal artery were collected at multiple cardiac phases to estimate the systolic cardiac phase for kidney tissue. A research application vendor-provided work-in-progress single shot echo-planar imaging sequence with dynamic field correction, and cardiac triggered oblique coronal was used for DWI acquisitions. Images were aligned to the prior HASTE imaging, and acquisitions were at multiple echo times for each of bipolar and flow compensated pulse sequences with TR/TE1/TE2 2800/81/120 ms, matrix 192/192/1, resolution 2.2/2.2/5 mm, GRAPPA acceleration factor 2, bandwidth 2170 Hz/pixel, and b-values of 0, 10, 30, 50, 70, 80, 100, 120, 200, 400, 600, and 800 s/mm2 in 12 directions. The four DWI acquisitions (two echo times for each of bipolar and flow-compensated acquisitions) were each performed in 6 min totaling 24 min. Additionally, to correct for motion and field inhomogeneity, 16 right-to-left and 16 left-to-right phase-encoding b = 0 images were acquired sampling the full range of motion for each kidney. Marchenko-Pasteur principal component analysis (MPPCA) [37] was performed for denoising. The kidneys were registered retrospectively to correct for breathing and cardiac motion. The processing for the left and right kidney was performed independently to better mitigate asynchronous left and right kidney motion and left-sided cardiac signal drop-out. In order to correct for field-inhomogeneity artifacts, the images were inputted into FSL TOPUP [38]. The processing flowchart was a replica of the one in Gilani et al. [6]. Image by image inspections were performed to exclude corrupted images for DTI and IVIM analysis. 3 left kidneys were excluded due to substantial signal loss [39] or unsuccessful FSL TOPUP correction, resulting in the inclusion of 11 out of 14 kidneys in the analysis.

Table 1 MR acquisition parameters of this study
Full size table

Layered segmentation

The four DWI acquisitions were registered together with a rigid body mutual information-based algorithm and underwent a six-layer segmentation implemented as multiple zones concentric objects (MZCO) generation; both of these steps used the freely available software package FireVoxel (build 421, https://firevoxel.org/). The inner and outer borders of the kidney were manually prescribed and the MZCO algorithm mapped layers of equal thickness. The contours for these layers smoothly varied parallel to the prescribed inner and outer borders for segmentation of the kidney images containing both the medulla and cortex. Additionally, another layered segmentation mask was generated by the overlap of the layers derived above with a highly inclusive medulla segmentation based on hyperintense regions of the FA map.

Extraction of diffusion parameters

The bipolar diffusion-weighted images were processed in custom code written in MATLAB and Statistics Toolbox (Release 2022a, The MathWorks, Inc., Natick, Massachusetts, United States).

First, average IVIM maps were generated from biexponential fits of the directionally averaged bipolar gradient diffusion-weighted image sets at TE = 81 ms:

$$\frac{S}{{S_{0} }} = f_{{\text{p}}} \exp \left( { - b \cdot D_{{\text{p}}} } \right) + \left( {1 - f_{{\text{p}}} } \right)\exp \left( { - b \cdot D_{{\text{t}}} } \right)$$
(1.1)

where Dt is defined as diffusion coefficient of water in the tissue, fp is the fraction of DWI signal that is affected by perfusion, and Dp is pseudo-diffusion coefficient which is sensitive to flow speed and architecture [2]. Dt and fp values were determined in a first fit to high b-values (b > 200 s/mm2) and provided as first estimates. A second fit on all b-values with constrained Dt was performed to estimate fp and Dp. This approach is defined as segmented biexponential fit.

Secondly, the directional DWI signals from the bipolar acquisition at TE = 81 ms were processed analogously to the Renal Flow and Microstructure AnisotroPy (REFMAP) approach [3] to extract DTI, IVIM, and directional IVIM parameters: FA, mean (MD), axial (D∥), radial (D⊥) diffusivities, scalar f with the same definition as fp, and mean (D*), axial (D*∥), radial (D*⊥) pseudodiffusion coefficients. First, the b-value dependence of each voxel and direction was fit to a biexponential model in a segmented fashion as above. The structural diffusivities (Dt) were fit to a standard diffusion tensor model to derive D∥, D⊥, and MD as well as FA [40, 41]. Next, at each voxel, pseudo-diffusion coefficient (Dp) were projected along the already derived axial and radial eigenvectors for Dt to derive its axial (D*∥), radial (D*⊥), and mean (D*) versions. The final parameter set included DTI metrics (MD, FA, D|| , D⊥) from the Dt, as well as the scalar f, in addition to D*, D*∥, and D*⊥.

Finally, both bipolar and flow-compensated diffusion signals at both echo times, were averaged over all directions and used to generate signal intensity curves from each layer for input into the multiple encoded FC-IVIM method to estimate flow velocity v, vessel segment length l, and f [27] for each layer. In this process, high b-values from the bipolar TE 81 ms acquisition were again fitted to generate estimates of Dt, and fp which initialized their values in a combined fit of the FC-IVIM expression to all b-values. A laminar flow description was adopted for the distribution of velocities in each segment of the circulatory network, and the bulk diffusivity value for flowing spins was set to 2.15 µm2/ms. This higher value was chosen over that of pure blood (1.6 µm2/ms [42]) to reflect the significant contribution from tubular fluid. This selection, while approximate, is also deemed acceptable given the typically much larger pseudodiffusion coefficients (by a factor of 10) to which this term serves as a background.

Statistical analysis

Spearman correlations were used to assess the association of the parameters with layers. All statistical tests in this study were conducted at the two-sided 5% significance level using SAS 9.4 software (SAS Institute, Cary, NC).

Results

Figure 1 shows a sample kidney divided into six layers using the multiple zones concentric objects generation method in both full parenchymal and medullary only segmentations. Figure 2 shows a b = 0 image, and maps of IVIM (Dt, fp, and Dp) and DTI (MD, FA, D∥, and D⊥) parameters corresponding to the same subject. Similar maps for another kidney are shown in Supplementary Fig 1. Figure 3 shows boxplots of the group distribution of values of DTI (a), IVIM (b), and FC- IVIM (c) parameters against cortex and medulla inclusive parenchymal kidney layers (increasing layer # corresponding to outward layer progression). Figure 4 shows variation of the same parameters within medulla segmentations (derived from the FA maps) on the same kidney layers.

Fig. 1
figure 1

(a) A sample kidney divided into six layers using the multiple zones concentric objects generation method of FireVoxel. (b) The same kidney and layers limited to a medullary region of interest highlighted based on fractional anisotropy (FA) map

Full size image
Fig. 2
figure 2

A sample b = 0 image (a), IVIM maps Dt (b), Dp (c), and fp (d), and DTI maps FA (e), MD (f), D∥ (g), D⊥ (h), D*∥ (i), and D.*⊥ (j), all corresponding to the same subject as in Fig. 1

Full size image
Fig. 3
figure 3

Directional diffusion and flow parameters MD, FA, D∥, D⊥, D*∥, D*⊥ (a), IVIM parameters Dt, fp, Dp (b), and FC-IVIM parameters D, f, l, v (c) vs. kidney layers containing both medulla and cortex

Full size image
Fig. 4
figure 4

Directional diffusion and flow parameters MD, FA, D∥, D⊥, D*∥, D*⊥ (a), IVIM parameters Dt, fp, Dp (b), and FC-IVIM parameters D, f, l, v (c) vs. kidney layers in the medulla segmentations

Full size image

Table 2 summarizes Spearman correlations of the parameters separately in parenchyma (including both medulla and cortex) and pure medulla against the six layers. Most of the parameters showed significant changes from the inner to outer layers with a few exceptions. These trends can be summarized as microstructural parameters (D, MD, Dt, D∥, D⊥, FA, l) or dynamic flow parameters (fp, f, Dp, D*∥, D*⊥, and v). The structurally sensitive diffusivities D, MD, Dt, and most pronouncedly perpendicular diffusivity D⊥, all increased from the inner layers to the outer layers of the kidney, while FA decreased from the inner to outer layers. These trends were similar in the whole parenchymal and medullary segmentations, but the structural diffusivity correlations (MD, Dt, D, and especially D⊥) were stronger, and the FA correlations weaker, in the medullary segmentation case. Conversely, in both the parenchymal and medullar layers, D|| and l were not significantly associated with layers. Dynamic IVIM parameters fp and Dp, and FC-IVIM parameters f and flow velocity v had decreasing trends from the inner to outer layers.

Table 2 The Spearman correlations (r) and p-values characterize the association of parameters with layer index (1 = inner medulla; 6 = outer cortex)
Full size table

Figure 5 summarizes the values of Spearman correlation coefficients (r) for each parameter for both the parenchymal) and the medullary segmentations. The figure highlights that Dp derived from IVIM, its directional variants D*∥ and D*⊥, and v derived from the FC-IVIM model have the highest correlations ranging from 0.46 to 0.55 in the category of dynamic parameters [2]. FA in both the parenchymal and medullary segmentations are highly layer dependent [3].

Fig. 5
figure 5

Radar plot of Spearman correlation coefficients for each parameter, both the parenchymal (including both cortex and medulla) and the medulla only layers (i.e. an intersection of FA map and the layer masks). If the correlation coefficient was negative the parameter was marked with a negative meaning it would decrease from the inner to outer layers. Radius of each symbol depicts the magnitude of the spearman correlation coefficient r for parameter vs. layer

Full size image

Discussion

There is a well-developed literature on preclinical murine and excised human kidney microstructure using in vivo or ex vivo modalities and their combinations such as MRI [43,44,45,46], 3D x-ray [47], and light sheet microscopy [48]. A potential transfer of this knowledge to in vivo human kidney imaging and extrapolations to its microstructure could be highly impactful in the study of renal dysfunction. However, diffusion-weighted MR parameters are generally not specific [49, 50] due to the dependency of the diffusion signal on different physiological parameters and restriction, as well as inter-species variations of microstructural features.

Accordingly, in the present study, we have scrutinized the dependence of renal DWI parameters from conventional representations (DTI, IVIM), an advanced hybrid DTI-IVIM approach [3, 51], and a multiple encoded FC-IVIM model [27] on concentric layers in vivo in humans. It is worthwhile to first summarize the correlation coefficients numerically and statistically to determine the strongest associations of DWI parameters with kidney layer. Secondly, we may hypothesize biophysical/microstructural features that are consistent with these trends, informed by known anatomy and histopathology. Regarding interpretation of the layer trends, we distinguish two different sources of layer dependence, as typically required in DWI multi-component modeling: (1) variations in diffusion coefficients and (2) variations in signal fraction of these components. For our purposes and with guidance from literature, the dominant components are (a) vascular, (b) tubular, and (c) interstitium and pathologically, fibrotic fraction. It is also important to note that the nephron path (through the glomerulus, proximal convoluted tubule, tubular loops of Henle, distal convoluted tubule, and collecting duct) proceeds back and forth between cortex and medulla rather than directly along the radial direction considered here. Thus, there is unavoidable averaging of multiple elements of the nephron within each layer.

As before, we summarize results in categories of structurally sensitive parameters and dynamic parameters. Assuming the tubular fraction dominates the structurally sensitive parameters’ (D, MD, Dt, D⊥) behavior, the decreasing trend from outer to inner layers may originate from a combination of factors. It is known that the tubular volume fraction decreases from 85% in cortex to 60% in the medulla [52, 53] and that in deeper layers a larger preponderance of thinner tubules (such as the loops of Henle) was found. Thus, larger average tubule diameters in outer cortical layers induce reduced diffusion restrictions. Finally, tubule and duct orientation become increasingly aligned in the medulla, driving the canonical feature of anisotropy (elevated FA and reduced D⊥). D∥ and multiple encoded FC-IVIM branching length l follow a relatively constant trend. A relatively constant D∥, along with the known presence of tubules in all kidney layers, suggests diffusion is similarly hindered in the direction along the tubule axis (i.e. similar ‘mean free path’) in all layers. However, future studies similar to Scott et al. [54] might enable better quantification of the diffusion hindrance and branching length (l) using alternative modalities to test whether there are relatively fixed parameters across all kidney layers.

Intramedullary correlations were weaker for FA and l than for the full parenchyma; this may originate from a more homogeneous tissue sampling in the former case, in particular with regard to the diameter, degree of alignment, and vessel segment length of the tubular structures. However, it is notable that FA correlates strongly with layer in both parenchymal and medullary segmentations; the latter results suggest the limits of treating medulla as uniform entity such as in classical cortex and medulla segmentations.

The layered approach is analogous to the laminar analysis of the cortical gray matter [55] and superficial white matter [56]. However, within human kidneys, there are less distinct histological layers compared to cerebral tissue which complicates interpretations. Nevertheless, spatial profiling of DWI metrics throughout the kidney may ultimately yield more specific biomarkers of renal tissue microstructure. Further studies using in vivo and ex vivo DWI are required to establish their connection with human renal function and histology.

Regarding dynamics, a first-order physiological intuition about the dynamic IVIM parameters fp and Dp, and FC-IVIM parameters f and flow velocity v would suggest that dynamics are slow in the inner layers since more exchange of water and other ions occurs between vascular and tubular spaces. However, the averaging of all dynamic flow systems (vasa recta, tubules, collecting ducts) must again be considered when interpreting the net higher flow rates evident in the inner radial layers. Nevertheless, one interesting result of this pilot application of the FC-IVIM model is that it prescribes the majority of this change to dynamics (velocity v) and not to architecture (branching length l). Also, the slope of these variations of dynamic parameters (fp, Dp, f, v) are more significant in the first three inner layers implying these flow describing parameters might be affected by the renal artery, vein, and pelvis. Conversely, in the outer layers which are mostly cortical tissue, there are networks of microvasculature and convoluted tubules with decreased velocities.

The FC-IVIM model is one of the few treatments of IVIM signal beyond the biexponential representation (e.g. [57,58,59]) and represents potential for biologic specificity. However, as with all diffusion MR models, it has its limitations, some of which might be better addressable in the future. First, the FC-IVIM model [27] assumes only one flow compartment with a single branching length and characteristic distribution of velocities, both of which are approximations. There are also no studies yet investigating numerical correlated errors between the model parameters velocity and branching length. These issues might result in biased outputs [50]. Again, further microstructural cross-validations are required for its features similar to the work by Scott et al. [54] where the branching length derivations of the model are compared with µCT measurements. Also, the analytical expressions and numerical phase distributions underlying the FC-IVIM model provide possibilities for the estimation of microstructural parameters from in vivo data. However, a study of the model estimates from simulated MR signals from Monte Carlo and/or particle trajectory simulation of water transport in relevant microstructural/microcirculatory networks (such as anisotropic flow networks as in renal medulla) might further illuminate its range of applicability. Interpretations of IVIM and DTI parameters would also benefit from such studies. Also, the multiple encodings required for the model should be considered in light of the balance between accuracy, precision, and clinically feasible scan times. A recent work has proposed optimized acquisitions with this in mind [31]. We did not employ these exact optimized acquisitions recommended in [31] but our acquisitions were somehow similar and near optimal in the case of b-values, diffusion times and pulse sequences used.

There are further limitations to this study. First, we have thus far collected the layer trends in young healthy subjects only, aiming to improve our understanding of the sources of diffusion MR signal in the kidneys. Hence, a natural future direction would be to test the layered analysis on DTI or IVIM parameters in chronic kidney disease patients or subjects with a wider range of ages [60]. However, a redefinition of the layers might be necessary in lesioned or morphologically abnormal kidneys. Finally, we did not perform a systematic variation of the number of layers chosen in our ROI sampling, nor did we perform an interreader comparison on their prescription.

In conclusion, we have performed a preliminary study on the dependence of a set of conventional and more advanced DWI derived parameters on concentric layers in healthy human kidneys. The most significant layer dependence was observed for pseudodiffusion parameters and structural fractional anisotropy, with weaker dependences observed for structural diffusivity parameters. Further validation of these trends in comparison with histologic reference, as well as correlation with measures of renal function, is required to improve our understanding of the sources of the diffusion signal in the kidneys. This knowledge might find translations into the clinics to optimize acquisition and better understand the pathophysiology of kidney diseases in the future.