Key words
abdomen - MR-diffusion/perfusion - MR-functional imaging - physiological studies -
kidney - renal imaging
I. Introduction
All over the world people are suffering from chronic kidney diseases with a high number
of unreported incidents [1]. The evaluation of kidney function has remained a huge challenge for modern medicine
and nephrologists are repeatedly faced with complex kidney pathologies and pathophysiologies
that the diagnostic methods used in the clinical routine, such as laboratory tests,
ultrasound, and renal scintigraphy, sometimes do not have sufficient sensitivity or
specificity to solve [2]
[3]
[4]. How can we improve diagnostic methods using recent developments in imaging techniques
and data analysis to help clinicians in the early detection, monitoring, and treatment
of kidney diseases?
In the past decades abdominal imaging with MRI has become a well-established method
for detecting various pathologies. Nonetheless, the range of MR imaging techniques
used in the clinical routine only represents a portion of the diagnostic potential
of MRI. Functional MR imaging (fMRI) is an emerging field with great potential to
take diagnostic imaging to the next level. Without the application of contrast agents,
multiparametric fMRI provides non-invasive methods to measure organ perfusion, diffusion,
oxygenation and to characterize changes in tissue composition. Especially the assessment
of kidney function could become a major area of application, since the complex interaction
of blood flow, perfusion, and oxygenation in the renal physiology and pathophysiology
of various kidney diseases is still the subject of numerous studies. Additionally,
recent advances in the field of deep learning in medical imaging could substantially
contribute to coping with data postprocessing and to helping to integrate functional
information into the clinical routine for more efficient and feasible diagnostics.
This article outlines the most widely used fMRI techniques and gives an overview of
the current state of clinical applications, recent studies, and new developments in
data postprocessing with regard to emerging deep learning techniques.
II. Functional MRI Parameters
II. Functional MRI Parameters
We can already look back at decades of research on functional imaging with MRI. Since
the implementation of MRI scanners for medical diagnostics, MRI has gone far beyond
visual imaging [5]. A variety of fMRI techniques have been developed and to some extent applied in
the clinical routine. Many promising techniques, however, have remained restricted
to research, which also applies to functional imaging of the kidneys [6].
One of the major barriers for broader application and clinical use of renal fMRI is
the confusing diversity of fMRI techniques and variations in acquisition, post-processing,
and analysis approaches [7]. In order to merge the developments and results of individual studies around the
world and standardize clinical research, a multinational group funded by the European
Union COST (European Cooperation in Science and Technology) Action called ‘PARENCHIMA’
was formed [8]
[9]. Inspired by the recommendations of this network, this article focuses on the main
techniques of fMRI including perfusion, diffusion, and BOLD (blood oxygen level-dependent)
imaging added by further MRI techniques such as T1 / T2 mapping and dynamic contrast-enhanced
(DCE) MRI, which might also contribute to a more comprehensive understanding of pathological
changes in renal structure and function.
Measuring perfusion with ASL
Renal perfusion is expected to be one of the main elements of renal function, and
changes in kidney perfusion might significantly impact the progression of acute kidney
injury and chronic kidney diseases. Beside different methods measuring renal blood
flow, the assessment of renal perfusion is far more complicated and difficult to perform
non-invasively. Arterial spin labeling (ASL) is a technique for the non-invasive imaging
and quantification of perfusion without the administration of contrast media using
the water molecules of the blood as endogenous tracers. Therefore, the blood has to
be prepared and “labeled” magnetically before entering the imaging plane, which is
performed by implementing radiofrequency pulses to change the longitudinal magnetization
of the water protons. After generating two images, a “control” image without labeled
blood and a “label” or “tag” image with labeled blood, the relative perfusion can
be calculated by subtraction. The result depicts a perfusion-weighted image, where
signal intensity is proportional to the perfusion. Using kinetic models, quantitative
perfusion maps can be computed and the perfusion can be measured in mL/100 g/min [10].
There are different schemes for labeling the blood with flow-sensitive alternating
inversion recovery (FAIR) pulsed ASL and pseudocontinous ASL being the most widely
used [11]. In recent years, faster readout strategies such as 3 D gradient and spin echo (GRASE)
sequences contributed to faster image acquisition and improvement of the intrinsic
low signal-to-noise ratio (SNR) of ASL. Motion correction tools and suppression of
background tissue compartments can additionally help to optimize image quality [12].
A shortcoming of renal ASL is the problem of validation due to the lack of a gold-standard
technique for perfusion measurement. ASL has been validated against para-aminohippurate
clearance, ultrasound flowmetry, microspheres, and scintigraphy [13]. There were also numerous studies testing the reproducibility of renal ASL, though
with different technical implementations [13]. The clinical applications of ASL encompass diagnostics for renal transplant recipients,
living kidney donors, acute kidney injury, and chronic kidney diseases, showing a
correlation of decreasing GFR with a reduced perfusion signal [14]. ASL was also tested under the influence of various drugs affecting renal perfusion
[15].
Assessing renal microstructure with DWI
In the clinical routine, the only way to assess changes in the renal interstitium
to date is renal biopsy. But there is a promising non-invasive alternative to make
microstructural changes visible and measurable with diffusion-weighted imaging (DWI).
DWI is sensitized to the Brownian motion of water molecules, making it possible to
draw conclusions concerning renal microstructure and the degree of renal fibrosis.
Strong additional bipolar gradients are applied with varying gradient strengths and
durations constituting different diffusion-weighting, which is summarized in b-values.
As a quantitative parameter for diffusion, the apparent diffusion coefficient (ADC)
is calculated from DW images to indicate the degree of water displacement [16]. Beside the diffusion of water molecules, an effect called pseudodiffusion can be
observed at lower b-values, which results from water motion in preformed structures,
i. e., perfusion. With the intra-voxel incoherent motion (IVIM) approach, pseudodiffusion
and “real” diffusion can be distinguished by using a biexponential model [17]. In order to investigate the spatial dependence of diffusion and quantify the degree
of the well-known spatial anisotropy of the renal medulla, a larger number of non-collinear
diffusion directions than the three directions usually used for the calculation of
ADC can be acquired. This technique called diffusion tensor imaging (DTI), however,
is accompanied by a high expenditure of time [17]. The majority of studies applying renal DWI simply use ADC, despite the advantage
of additional information from IVIM and DTI [18].
Similar to renal ASL, validation studies with renal DWI also lack a standardized acquisition
and analysis protocol. For biological validation, again no gold standard is available.
Renal DWI has been technically validated in numerous reproducibility studies and applied
in various clinical studies, including acute graft dysfunction, acute pyelonephritis,
polycystic disease, and chronic kidney disease [19].
BOLD MRI to reflect renal tissue oxygenation
Blood oxygenation level-dependent (BOLD) MRI is the first method to non-invasively
estimate the oxygenation status of blood and to evaluate tissue oxygenation. This
is made possible by the change of the magnetic properties of hemoglobin (HB) with
oxygen saturation, which leads to a decrease or increase in free induction signal
decay time constant T2* resulting in a measurable contrast in BOLD imaging [20]. A decrease in tissue oxygenation is assumed to play a major role in the progress
of chronic kidney disease and acute kidney injury. The information about the ratio
of oxy- and desoxy-HB in correlation to perfusion could help to gain more comprehensive
insight into pathophysiological processes and to determine between causes and consequences
of renal hypoxia [21].
Since no gold-standard for non-invasive measurement of tissue oxygenation is available,
BOLD MRI has only been validated against micropuncture techniques in animals [22]. Various studies have demonstrated the reproducibility of this technique in humans
as well as its potential to detect changes in human renal tissue oxygenation in patients
with chronic kidney diseases, renal artery stenosis, and transplant kidneys. Moreover,
BOLD-MRI has been proven to be a useful tool for evaluating drug effects on kidney
function and for estimating its potential nephrotoxicity. [20]
Similar to ASL and DWI, BOLD MRI lacks standardization in image acquisition and analysis.
It is most frequently performed with multiple gradient echo (GRE) sequences at 3 T
as the preferred field strength. It is recommended to standardize the physiological
status of the patient regarding hydration and salt intake, since these exogenous factors
have been shown to significantly influence tissue oxygenation [9]. In research settings, regions of interest (ROIs) have been a widely used technique
for image analysis, whereas for clinical settings a whole kidney analysis encompassing
the cortex and medulla should be performed. The gain of information could be increased
using BOLD in combination with other fMRI techniques such as ASL and DWI [23].
Tissue characterization with T1 and T2 Mapping
Tissue changes caused by inflammation, edema, fibrosis, or necrosis lead to changes
in relaxation times which can be imaged and quantified by T1 or T2 mapping [24]. For T1 mapping, the established method is the inversion recovery (IR) technique. The IR preparation
is repeated several times while increasing inversion time to acquire multiple data
points using a single-shot imaging module (e. g., echo planar imaging (EPI), steady
state free precession (SSFP)) with long repetition time (TR). Recently published techniques
such as the variable flip-angle (VFA) or modified Look-Locker imaging (MOLLI) are
possible alternatives. [25]. Due to differences in the content of free water between the medulla and the cortex,
T1 mapping allows pronounced corticomedullary differentiation (CMD) [26]. In renal transplants, however, cortical and medullary T1 relaxation times seem
to be higher than in healthy subjects [24]. Various studies also depicted a correlation of T1 relaxation times and the degree
of renal impairment and GFR in patients with renal insufficiency [26]
[27].
T2 mapping is usually performed with multi-echo(fast) spin-echo sequences and enables
full kidney coverage within a short period of time. Image quality can be influenced
by susceptibility artifacts due to blood flow, tissue diffusivity, magnetic field
inhomogeneities, and imperfect slice selection pulse profiles [25]. To mitigate these problems, special T2 preparation pulses (e. g., CPMG) can be
used in combination with a fast single-shot readout (similar to the application of
an inversion pulse in T1 mapping) [28]. In cardiac imaging, T2 mapping is an established technique to detect edema after
myocardial infarction or inflammation [29]. Beside several animal studies examining T2 sensitivity to ischemia-reperfusion
injury [24], T2 mapping might have the potential to detect early stages of autosomal dominant
polycystic kidney disease (ADPKD) in the future [30]. Larger studies evaluating the value of T2 mapping of the kidneys in humans are
still missing.
Dynamic contrast-enhanced MRI
Dynamic contrast-enhanced (DCE) MRI implies the administration of gadolinium-containing
contrast agents. They enter the kidney’s capillary system and undergo glomerular filtration
and tubular excretion before being eliminated in the urine. By analyzing the contrast
enhancement in the tissue of interest, DCE offers the possibility to describe the
glomerular filtration rate (GFR) by measuring tubular flow and tubular transit time
as well as vascular parameters such as permeability, renal blood flow, and blood volume
[31]. DCE MRI has been validated in several studies against different methods and in
patients with acute kidney injury, renal tumors, and kidney transplants [32]
[33]
[34], where it has been found valuable for assessing renal perfusion and filtration.
The main drawback of DCE is the application of a gadolinium-containing contrast medium,
which entails increased risk for developing NSF [35] and gadolinium retention in the brain [36], especially in patients with impaired kidney function.
Multiparametric fMRI protocols and applications
Apart from ASL, DWI, BOLD MRI, T1 and T2 mapping, and DCE MRI, there are further techniques
providing valuable information about pathological processes in the kidneys. According
to a position paper from the COST Action PARENCHIMA, further commonly available MRI
biomarkers are renal volumetry, phase-contrast MRI, and magnetization transfer (MT)
[37]. Most recent studies implementing multiparametric fMRI techniques comprise a combination
of ASL, DWI, BOLD, and T1 and T2 mapping. Sometimes volumetric analysis and phase-contrast
MRI are added [38]. A few include DCE or fat quantification with Dixon MRI [39]. Obviously, the choice of MRI biomarkers depends on the clinical question. Multiparametric
fMRI offers the possibility to select from a range of MRI biomarkers that address
the study hypothesis and kidney pathology. Especially the examination of transplant
kidneys with multiparametric fMRI protocols is of great interest, since there are
a lack of sensitive monitoring techniques for the evaluation of kidney function and
early detection of kidney failure. The combination of functional parameters assessing
changes in perfusion, diffusion, oxygenation, kidney volume, and tissue structure
is a promising approach to gain deeper insight into the biological process of acute
rejection of kidney transplants and prediction of long-term complications [40]. Multiparametric fMRI has recently been applied to examine patients with chronic
kidney diseases [41]
[42], acute kidney injury [38], IgA nephropathy [43], polycystic kidney disease [44], and interstitial renal fibrosis [45]. [Fig. 1] shows examples of a multiparametric fMRI protocol for kidney examination comprising
T1 mapping, BOLD, DWI, and ASL.
Fig. 1 Multiparametric renal functional imaging of a volunteer showing multiple slices of
different functional parameters. A Tissue characterization with T1 mapping. B BOLD MRI with T2* mapping. C Assessing microstructure with ADC mapping. D Perfusion imaging with ASL (renal blood flow (RBF) mapping).
III. Data postprocessing and analysis: new strategies with deep learning
III. Data postprocessing and analysis: new strategies with deep learning
With multiparametric fMRI a vast amount of data can be generated. How should this
quantity of data be handled? How can this information be directly extracted?
Data postprocessing and analysis is a topic rarely described in detail in most clinical
studies regarding functional renal imaging. Nonetheless, it is one of the main limiting
factors for use in clinical practice, since it can have a significant impact on data
interpretation. For successful application of multiparametric fMRI protocols in the
clinical routine and studies of large patient cohorts in the future, standardized
data postprocessing and data analysis workflows are needed.
Over the last few years impressive developments in deep learning (DL) have aroused
great interest across a variety of areas including healthcare applications [46]. Recent advances in artificial neural networks show great potential for medical
imaging technology, data analysis, and diagnostics and might substantially impact
the future of healthcare. MRI could take advantage of these advances by implementing
DL techniques for the optimization of data acquisition, postprocessing, and analysis
[47]. There are also efforts by the European Union COST to form a database for machine
learning applications on renal MRI and for “bridging the gap between data and technology”
[48].
The term “deep learning” refers to a deep network of multilayer neural networks to
analyze data. The main difference of the DL approach compared to other machine learning
techniques is its ability to learn feature representation and classification in the
same process, thus optimizing both simultaneously [49]. DL techniques with application in medical imaging are mostly based on convolutional
neural networks (CNNs) to learn useful representations of images and other structured
data [50]. The multitude of options to implement DL in the process of data analysis ranges
from data acquisition to computer-aided diagnosis. [Fig. 2] gives an overview over the different application steps for deep learning approaches.
Fig. 2 Image processing pipeline depicting different application steps for deep learning
approaches in multiparametric renal MRI.
Starting from MR image acquisition, DL can be implemented in the process of image
reconstruction to significantly improve robustness, accuracy, and image quality from
undersampled k-space data as well as to optimize speed compared to conventional reconstruction
approaches [51]
[52]
[53]
[54]
[55]
[56]. Impressive results have been seen for instance in dynamic MR image reconstruction
of cardiac MRI, where real-time image reconstruction [57] and 4 D DL reconstruction networks have been developed [58]. Also, CNN-based methods can assist in the detection of artifacts [59], prospective motion correction [60], and image denoising [61]
[62]
[63]. In image super-resolution, deep learning techniques are implemented for reconstruction
of higher-resolution images or image sequences from low-resolution images [64]
[65]
[66]. Further areas of application include image synthesis to derive new parametric images
of tissue contrast from a collection of MR acquisitions [67]
[68], quantitative susceptibility mapping (QSM) to noninvasively estimate magnetic susceptibility
of biological tissue [69]
[70], and MR fingerprinting (MRF) [71].
Multiparametric renal fMRI poses several challenges for data analysis regarding registration
and segmentation, where DL seems especially promising to boost further development
and the path to the clinical routine. First, renal parenchyma is difficult to differentiate
from the surrounding organs and structures only by signal intensity. Secondly, multiparametric
fMRI encompasses heterogenous signal contrasts and image qualities. Furthermore, kidneys
can vary dramatically in their anatomical position, size, and features, such as cysts.
Last but not least, motion due to breathing leads to a considerable variation in the
position of the kidneys not only between but also within measurements. Image registration
and segmentation, therefore, are a precondition for efficient analysis of multiparametric
fMRI data.
Image registration implies spatial alignment of intra- and inter-subject kidney images
to enable further processing steps. There is a range of strategies for image registration
with different approaches, which can be grouped into image acquisition techniques
and post-processing methods [72]. The emerging use of DL shows the greatest potential to contribute to more efficient
image registration, thereby surpassing standard deformable registration algorithms
in accuracy and speed. Still, application to renal MRI is pending, which might also
be attributed to the lack of public datasets and validation protocols. The application
of newly developed methods to other imaging modalities and organs, however, seems
promising for the transfer to renal MRI [50].
For quantitative analysis of multiparametric fMRI in the kidneys, organ segmentation
for the assessment of total kidney volume (TKV) but also kidney compartments including
the cortex and medulla is an essential step. Due to the challenges of renal multiparametric
fMRI as described above, manual segmentation has been the prevalent segmentation technique
in renal MRI studies. However, for clinical use of renal fMRI, this time-consuming
and laborious method needs to be replaced by more efficient segmentation techniques.
Beside other semi-automatic and automatic segmentation techniques such as image processing
and model-based image segmentation, machine learning and especially deep learning
approaches have again been shown to be the most promising to deal with the more complex
multiparametric datasets [73]. DL has already been applied in a few studies for segmentation of the kidneys to
estimate TKV [74]
[75]
[76]
[77]
[78].
Going even further beyond the tasks of image pre- and postprocessing, DL can be implemented
for computer-aided diagnosis. By combining the ability to analyze imaging data and
retrieve clinical information, automated classification systems can be developed to
assist clinical diagnostics, as has been successfully shown in the diagnosis of prostate
cancer, for example [79]
[80]. In kidney diagnostics, DL methods have been demonstrated to assist in the diagnosis
of transplant rejection with fMRI and the integration of clinical data, thereby forming
a computer-assisted diagnostic (CAD) system for the assessment of kidney function
including DWI, BOLD, and creatinine clearance [81]
[82]
[83]. Another application of DL is the differentiation of renal cell carcinoma [84]
[85].
IV. Discussion
Renal MRI is an emerging technique that has not yet been established in the clinical
routine. Several other more established imaging techniques with different strengths
and weaknesses are usually preferred by clinicians.
The most commonly applied method for renal imaging is ultrasonography. It is also
a non-invasive, non-ionizing technique, which provides the ability to dynamically
image morphological abnormalities with high resolution, to measure blood flow with
the Doppler method, and to apply a safe contrast agent to visualize perfusion without
harming the kidney [86]
[87]
[88]
[89]. In contrast to MRI, it is widely available and cost-efficient [90]. However, image quality is operator-dependent and can be considerably reduced by
gas between the transducer and the organ of interest or by the obesity of subjects.
Measurements and images are more difficult to reproduce and quantification is feasible
only to a limited extent [2]
[91]
[92].
Like MRI, CT is a tomographical imaging technique that uses radiation to obtain images.
Even though MRI techniques have become a lot more time-efficient in the past decades,
CT is still much faster than MRI and is more cost-effective [93]
[94]. Apart from morphological and angiographical imaging, CT has the ability to measure
renal blood flow and perfusion as well as GFR and tubular function [2]. The major drawback of CT is the need for nephrotoxic contrast agent for most tasks
besides detecting renal obstruction, which limits its use for kidney diseases [95].
Functional renal imaging in the clinical routine also includes renal scintigraphy.
It is the gold standard for measuring glomerular filtration and tubular function [92] and enables accurate evaluation of split function and renal obstruction. Nonetheless,
image resolution and quality are very poor in comparison to other imaging modalities
and diagnostic value is limited.
Most clinicians are not aware of the potential of MRI to offer functional imaging
of the kidneys even without the administration of contrast media. But there are also
several disadvantages to using MRI that might present an obstacle to broader application.
The leading drawback of MRI is the limited availability, especially in smaller hospitals,
and the cost expenditure associated with purchase and maintenance [90]. Renal MRI can be performed with 1.5 and 3 Tesla, although studies have shown the
benefits of 3 Tesla for SNR, examination time, and spatial resolution [9]. Moreover, the use of MRI needs experienced operators. When using multiparametric
fMRI protocols to examine kidneys, there is still a need for standardized procedures,
protocols, and postprocessing as well as more offers by medical technology companies
[6]. Contrary to general belief, the examination time of MRI has been reduced significantly
in the past decades and single parameters can be measured in a few minutes. Furthermore,
breathing strategies such as breath-hold, respiratory-triggered, or free-breathing
imaging can be adapted to patient condition for most sequences [9]. Last but not least, relative contraindications such as pacemakers and cochlear
implants have to be considered.
Nonetheless, MRI has a lot to offer and might help to even reduce the costs of multiple
sometimes even invasive diagnostic examinations. Beside high-resolution anatomical
imaging, the range of functional parameters offered by MRI is exceptional and can
be achieved mainly without the use of contrast agent. Examination protocols can be
adapted to the clinical issue, kidney disease, and patient condition and are suited
for both short-term and long-term monitoring. Eventually, the aim of developing multiparametric
fMRI for the kidneys is not to replace established techniques such as ultrasonography
and renal scintigraphy, but to broaden and improve renal imaging and help clinicians
and patients in the treatment of kidney diseases.
V. Summary
Multiparametric functional MRI of the kidneys is a promising approach for assessing
renal function and pathophysiology. Combinations of perfusion, diffusion, and BOLD
imaging together with techniques for tissue characterization such as T1 and T2 mapping
and further MR biomarkers can be selected depending on the clinical question and the
kidney pathology to gain more comprehensive insight into the cause and consequences
of diseases and the effect of therapeutic interventions. However, several obstacles
have to be overcome before the method can be implemented in the clinical routine.
On the one hand, there is a need for standardization of fMRI protocols to enable comparability
of studies and to facilitate clinical application. On the other hand, new strategies
for handling the emergence of huge amounts of data are necessary. Recent advances
in DL techniques open new possibilities for data postprocessing and analysis and might
give clinical use of renal fMRI a decisive push forward. More studies examining different
kidney pathologies with larger cohorts and longitudinal design with implementation
of a standardized workflow for data acquisition, postprocessing, and analysis are
needed to further improve and at the same time demonstrate the ability of multiparametric
fMRI to enhance diagnostic imaging of the kidneys.
