Technical Advancements in Abdominal Diffusion-weighted Imaging

Abstract

Since its first observation in the 18th century, the diffusion phenomenon has been actively studied by many researchers. Diffusion-weighted imaging (DWI) is a technique to probe the diffusion of water molecules and create a MR image with contrast based on the local diffusion properties. The DWI pixel intensity is modulated by the hindrance the diffusing water molecules experience. This hindrance is caused by structures in the tissue and reflects the state of the tissue. This characteristic makes DWI a unique and effective tool to gain more insight into the tissue’s pathophysiological condition. In the past decades, DWI has made dramatic technical progress, leading to greater acceptance in clinical practice. In the abdominal region, however, acquiring DWI with good quality is challenging because of several reasons, such as large imaging volume, respiratory and other types of motion, and difficulty in achieving homogeneous fat suppression. In this review, we discuss technical advancements from the past decades that help mitigate these problems common in abdominal imaging. We describe the use of scan acceleration techniques such as parallel imaging and compressed sensing to reduce image distortion in echo planar imaging. Then we compare techniques developed to mitigate issues due to respiratory motion, such as free-breathing, respiratory-triggering, and navigator-based approaches. Commonly used fat suppression techniques are also introduced, and their effectiveness is discussed. Additionally, the influence of the abovementioned techniques on image quality is demonstrated. Finally, we discuss the current and future clinical applications of abdominal DWI, such as whole-body DWI, simultaneous multiple-slice excitation, intravoxel incoherent motion, and the use of artificial intelligence. Abdominal DWI has the potential to develop further in the future, thanks to scan acceleration and image quality improvement driven by technological advancements. The accumulation of clinical proof will further drive clinical acceptance.

Introduction

Diffusion has been studied by researchers for quite some time now, though during most of its history it was not well understood. It may have been the first time that mankind witnessed a physical phenomenon depending on processes at the molecular level. Indeed, the random motion observed is caused by the never-ending motion of molecules in the solvent, driven by thermal agitation.

This fact was certainly not known to the first experimenters who noted the phenomenon. At the end of the 18th century, it was IngenHousz who first described this motion while looking at tiny bits of pounded charcoal in a drop of wine spirit.¹ The observation was made possible thanks to the microscope, at that time a hot new tool for observations on the (relatively) small scale. He described what he saw as “a confused, continuous and violent motion”. IngenHousz may not be well known, but he is also the discoverer of photosynthesis, and he was a pioneer of vaccination.²

Following this early observation, the same phenomenon was studied by many other experimenters. Best known is the paper by Brown, a botanist, who looked at grains of pollen immersed in water and published a detailed report.³ The motion came to be known as Brownian motion, named after him.

A mathematical-physical basis for what was observed was presented by Einstein,⁴ who also made the link with diffusion, a physical phenomenon described earlier by Fick.^5,6 On top of that, Einstein’s theory was based on the idea that matter had to be granular at the atomic/molecular level. Using this theory, the physical reality of atomism, first hypothesized by Demokritos, a Greek philosopher who lived about 2400 years ago, could be proven for the first time. Even more, the theory allowed estimating the size of the molecules.⁷ Perrin did confirm Einstein’s theory experimentally, incidentally also thanks to a newly developed type of microscope, the ultramicroscope.⁸ This work earned Perrin a Nobel prize. Of note is that in the same period, with an article published in the same year, Einstein used the same idea of granularity in his thinking about light. He postulated that light consisted of finite packets of energy (we now call them photons), and with this concept he could satisfactorily explain the photo-electric effect.^9,10 This earned him a well-deserved Nobel prize too.

The importance of diffusion itself for a living organism should not be underestimated. It is one of the main mechanisms needed to bring nutrients at the end of their journey all the way to cells, in the absence of bulk flow, and to move inside the cell, for instance toward the mitochondria. It is thus, simply, a necessary mechanism to sustain life. Note, however, that with MRI, we make images based on the diffusion of water molecules, not of the nutrients.

Diffusion properties are not determined by nuclear MR physics, as, e.g. relaxation times are. Diffusion is determined by the micro-environment of the diffusing molecules, which for MR in almost all cases will be water molecules. Therefore, the diffusion properties are not dependent on the magnetic field strength and, theoretically, if measured with a given set of acquisition settings, their values would be the same at all field strengths. This makes diffusion-weighted imaging (DWI)-based parameters excellent candidates as biomarkers, also because the acquisition is non-invasive, fairly straightforward, and does not necessitate contrast agents. It is well known, though, that the measured diffusion coefficient is strongly dependent on, for example, the exact timing of the sequence, and standardization of the acquisition and post-processing has not yet been achieved.^11–13

Another important consideration is that the diffusion process in a living entity is not entirely the same as the diffusion process that Einstein described. His theory describes diffusion in a boundary-free homogeneous solvent, while it is well established that water in and around cells will encounter barriers that will influence the diffusion. Consequently, diffusion in a tissue will not follow the same rules as free diffusion. Hence, one needs to be cautious when applying Einstein’s concepts to the in vivo process. Luckily, thanks to the size of the voxels used in MRI and the inherent averaging of all the diffusion processes ongoing in a voxel, it is (in most cases) still possible to characterize the diffusion by one coefficient. This is now called the apparent diffusion coefficient (ADC), as introduced by Le Bihan.¹⁴ “Apparent” is added to distinguish it from the original diffusion coefficient linked with free diffusion.

On the other hand, it is, of course, entirely thanks to the very fact that diffusion is hampered that we can extract information about the cellular environment from diffusion measurements, and that diffusion can help clinicians with their diagnosis. The obstacles in the way of the diffusing water molecules are biological structures, and the properties of these structures will determine the diffusion. Hence, diffusion can give us information about tissues, and a change in the physiology, for instance due to disease, is often reflected in a change of the diffusion properties that can be visualized using a DWI sequence.¹¹

An important problem, and one that still baffles quite a few people, is how it is even possible to use the small movements of water caused by diffusion, over a distance in the order of micrometers (10⁻⁶m), to create images from the body, given that a patient is constantly moving, with respiration being the most important source of motion to be reckoned with in abdominal and whole-body (WB) imaging. Luckily, the influence on the diffusion-weighted MR signal of the mostly translational bulk motion caused by respiration, over a distance in the order of centimeters (10⁻²m), can be separated from the influence of the random motion of the diffusing water molecules.^15–18

Indeed, the bulk motion will cause a phase change which will be constant for a whole voxel, assuming that all spins in that voxel experience the same rigid-body motion. On the other hand, diffusion will result in a phase accumulation that will be different for all spins as they move in a random way through the gradients used for diffusion sensitizing and, thus, will accumulate phase differently. Because of the averaging within one voxel, this results in an overall attenuation of the MR signal.

Therefore, coherent motion will not influence the pixel intensity of magnitude images, while diffusion will. If a single average image can be acquired, this resolves the issue. If averaging is needed, and this is typically the case as DWI inherently has a relatively low SNR, this preferably needs to be done in the magnitude image domain. It is also important to note that the measured ADC is not affected by linear translational bulk motion.¹⁹

Another important question in DWI is whether we can construct a model that predicts the diffusion properties, and specifically, the exact value of the ADC, of a tissue. This question remains unanswered, though one can identify many parameters that influence the diffusion properties. The interplay between these properties is, in general, rather complicated, making it (hitherto) impossible to produce a model to calculate the diffusion coefficient.

A few observations help in understanding diffusion properties better, though. First: in general, tissues that are haematopoietic or lymphopoietic or part of the reproductive system are hyperintense in diffusion-weighted images. Bone marrow, the spleen, lymph nodes, testes, and ovaries are among the tissues that are hyperintense, even if they are normal. Radiologists must be fully aware of this when interpreting body diffusion-weighted images. Second: the degree to which water molecules are hindered in their motion can be parametrized by the cell perimeter length as measured on the 2D images, at least in a model system.²⁰ This cell perimeter length represents the amount of intact membrane in a voxel. Water molecules are hindered on and near the membranes and their hydration layer and, therefore, their diffusion length is reduced and the ADC becomes smaller. This could be the dominant mechanism for many biological tissues. Third: the cellularity of the tissue certainly plays a role, but this could be understood by considering the second observation. Higher cellularity means more membranes (and other obstacles) in a voxel, and hence, a lower ADC. Fourth: let us not forget that physical parameters like temperature and viscosity play a role. Indeed, the diffusion constant of water has a well-known correlation with temperature, and it could be used to measure the absolute temperature. Viscosity also directly influences diffusion. Higher viscosity means lower diffusion, affecting, for instance, the signal intensity in a cyst. Apart from the ones mentioned above, other features are known to influence diffusion characteristics, such as tortuosity (in brain), connective tissue (in liver), nucleus/cytoplasm ratio, lipids in cells, and both the ratio of and exchange between intracellular and extracellular compartments. It is clear that it is not easy, perhaps not even possible, to model the diffusion mechanism in such a way that a workable equation can be derived to calculate the ADC. We can, however, obtain an understanding for why, for instance, cancerous tissue shows a high signal in general, as cancerous tissues tend to have high cellularity and larger cells with irregular shapes and, therefore, a larger membrane area as well as an abnormal nucleus/cytoplasm ratio. All these contribute to the hyperintensity.

The move toward using DWI in clinical applications started with its early conceptualization and first realization in clinical research by Le Bihan.¹⁴ Their implementation is mainly based on the work by Stejskal and Tanner,²¹ who relied on the theoretical framework developed by Abragam²² and Torrey, Carr, Purcell, and Hahn.^23–25

An important breakthrough for DWI as a clinical tool came with the discovery that DWI was an excellent imaging tool to diagnose stroke, with unique capabilities especially in the early phase.²⁶ At first, it was mainly used in the brain but DWI was soon developed for application outside the brain.^13,27,28 It then quickly gained prominence in oncological imaging such that, thanks to the demonstrated clinical advantage, many routine MR examinations now include DWI.^11,29 Primary tumors and metastases generally show as hyperintense. The ease of the examination for both patients and clinicians, compared to other modalities, together with the absence of radiation and contrast agents contributed to the acceptance of DWI in the clinical setting. Researchers are looking further into the usefulness of DWI for oncological screening and therapy follow-up and are also exploring its use for more general screening, as DWI is sensitive to many types of abnormalities, while not being very specific.^30–35

In this review, we discuss how DWI has been used in abdominal and WB imaging and what optimizations are needed to make it succeed. Techniques like parallel imaging (PI), in this case used mostly for distortion reduction, and improved fat suppression were crucial in bringing DWI outside the brain and further into clinical daily practice. Recently, we have also seen further improvement through the advent of artificial intelligence (AI), which will enhance reconstruction, acquisition, and data processing alike. We believe this will further improve the usefulness of DWI to support radiologists and clinicians in their clinical decisions, to the benefit of an ever-greater pool of patients who can be helped with this imaging technique, which provides information about the microscopic environment in a macroscopic image.

PI Advantages in DWI Acquisition

DWI has been designed to investigate the microscopic random motion of water in tissues.^14,26,36 All DWI methods contain diffusion-weighting gradients, or motion-probing gradients (MPGs), that render sequences sensitive to tissue diffusion.²¹ Because strong gradients are usually used for MPGs,^16,21 even a small level of subject motion (such as movements associated with the cardiac cycle or random body motion) can generate large phase variations between k-space profiles.^37–39 This can cause significant ghost artifacts in the reconstructed images, unless the phase variations are properly taken care of.³⁸ Hence, one of the main challenges in DWI, especially for abdominal DWI, is to reduce the effect of subject motion while preserving the sensitivity to intravoxel incoherent motion (IVIM) due to diffusion in tissue.

The simplest way to minimize the effect of motion is to acquire all echo signals after a single excitation, i.e., single-shot echo-planar imaging (ssh-EPI), which results in the same motion-related phase error for all k-space profiles. The effect of the motion will then be reflected as an additional phase, which is constant per voxel and is discarded in the reconstruction of the magnitude image.^17,18

Typically, averaging of ssh-EPI DWI acquisition is often needed in order to obtain a clinically acceptable SNR. Even in the case of using averaging in ssh-EPI, artifacts due to phase variations will not occur if averaging is performed on magnitude images.^15–18 Therefore, ssh-EPI is capable of fast acquisition while suppressing the image degradation due to physiologic or involuntary subject motion. For these reasons, ssh-EPI has been highly successful in abdominal DWI as well as brain DWI.^40–44

However, EPI, especially ssh-EPI, has the disadvantage of being susceptible to chemical shift and, similarly, the off-resonance effect resulting from variations in local magnetic susceptibility. The corresponding artifacts are called chemical shift and susceptibility artifacts, respectively.^16,45

In conventional spin-echo or gradient-echo imaging, chemical shift artifacts can appear in the readout direction and not along the phase-encoding direction. This is because transverse magnetization is spoiled between phase-encoding steps and phase information is destroyed; thus, phase shift accumulation does not occur in non-EPI sequences. On the contrary, in EPI, the readout bandwidth (BW_read) is sufficiently wide to effectively reduce the chemical shift artifacts in the readout direction. However, due to the way phase-encoding is implemented in EPI, the phase-encoding effectively behaves as a frequency encoding.⁴⁵ The bandwidth in the phase-encoding direction (BW_phase [Hz]) is expressed with the following equation⁴⁵:

$$B{W}_{phase}=\frac{\text{\#}shot}{T_{esp}}$$ (1)

where #shot is the number of shots for EPI and T_esp (s) is the echo spacing time. The T_esp is equivalent to the sampling interval in the readout direction but is far longer. The BW_phase, inversely proportional to T_esp, is then typically of the order of 1 kHz, while BW_read is typically two orders of magnitude larger.⁴⁵ Therefore, chemical shift as well as susceptibility artifacts can be substantial in the phase-encoding direction.

The chemical shift ΔP_cs (m) in the phase-encoding direction can then be calculated from the following equation using BW_phase⁴⁵:

$$\mathrm{\Delta }{P}_{CS}=\frac{\mathrm{\Delta }{f}_{CS}\cdot FO{V}_{PE}}{B{W}_{phase}}$$ (2)

where Δf_cs (Hz) is the chemical shift frequency difference and FOV_PE (m) is the FOV in the phase-encoding direction. It can be understood from equations (1) and (2) that chemical shift becomes larger with narrow BW and because #shot is 1 in the case of ssh-EPI, the ΔP_cs is maximal.

In general, abdominal imaging requires a larger FOV compared with brain imaging, and significantly stronger variations of susceptibility are present. In addition, abdominal anatomy contains more fat tissue than brain. Consequently, dealing with the chemical shift or susceptibility artifacts in abdominal imaging is more challenging compared with brain imaging when ssh-EPI DWI techniques are applied, particularly at higher field strengths of 3T or more.⁴⁶

PI is a breakthrough technology that can alleviate the above ssh-EPI-related challenges.^12,47–49 The basis of PI methods is that the scan time is shortened by reducing the number of phase-encoding steps in k-space by a reduction factor (R). If the spatial resolution is kept the same, the same coverage in k-space is required. Therefore, reducing phase-encoding steps results in an increase in the distance between phase-encoding lines with a factor of R. In other words, the acquisition FOV defined by the Nyquist theorem decreases by a factor of R. If the acquisition FOV is smaller than the object, aliasing artifacts occur. PI is the algorithm that can be used to unwrap aliasing by using phased array coils for reconstructing the final FOV, utilizing the coils’ sensitivity profiles.

PI is generally used to shorten the scan time. But, in ssh-EPI, the impact of using PI on scan acceleration is relatively limited (only around 50 ms reduction per slice) because the readout time for completely filling the k-space is just 100 ms or less.⁴⁵ However, because the acquisition FOV is decreased by R, as is clear from equation (2), the use of PI largely results in a reduction of the chemical shift and susceptibility artifacts. When R values such as 2 or 3 are used, the acquisition FOV decreases to one-half or one-third and, thus, the chemical shift artifact would proportionally decrease. In Fig. 1, a comparison of DWI images taken with and without PI is shown. Here, R  =  2 and R  =  3 are used for PI, and significant image quality improvement in terms of image distortion and susceptibility artifacts is demonstrated as R increases.

Fig. 1.

Comparison of free-breathing liver DWI with b-value of 1000 s/mm² images without PI (a) and with PI using reduction factor (R) of 2.0 (b) and 3.0 (c). Image distortion and susceptibility artifacts are improved as R increases. DWI, diffusion-weighted imaging; PI, parallel imaging.

As shown in equations (1) and (2), shortened T_esp also helps reduce the artifacts. The T_esp can be reduced by applying a stronger readout gradient.⁴⁵ However, since the maximum gradient strength is already used in most ssh-EPI cases, major investment in hardware would be required. In the case of PI, the chemical shift can be reduced by R without any hardware changes, which means it is easy to implement. As a result, with the advent of PI, the abdominal DWI application has evolved dramatically.^28,50–59 In Fig. 2, the applicability of ssh-EPI DWI with PI for various abdominal regions is demonstrated.

Fig. 2.

Examples of abdominal and pelvic DWI with b-value of 1000 s/mm² images acquired on a 3T system using ssh-EPI DWI with PI. R =  2.5 was used for all imaging. Liver (a) and pancreas (b) with axial acquisition, kidney (c) with coronal acquisition, and rectum (d) with sagittal acquisition were obtained without severe distortion, thanks to PI. DWI, diffusion-weighted imaging; PI, parallel imaging; R, reduction factor; ssh-EPI, single-shot echo-planar imaging.

On the other hand, PI comes at the cost of decreasing the SNR when increasing R.^12,47–49 PI accelerates the scan by skipping k-space lines, which causes a decrease in the SNR. In addition, the SNR in PI is spatially dependent and is determined by parameters such as the number of coil elements and the sensitivity profile of each element, which strongly depends on element size and on the geometric arrangement of the elements. These are all expressed by a parameter called geometry factor (g-factor). Note that the g-factor depends on the geometry and thus varies from position to position. The g-factor expresses the capability of the PI algorithms to disentangle signals that are overlapping because of the skipping of k-space lines. The SNR in PI is expressed using the g-factor (g) as follows⁴⁸:

$$SN{R}_{PI}=\frac{SN{R}_{full}}{g\sqrt{R}}$$ (3)

Here, SNR_PI is the SNR in the PI scan and SNR_full is the SNR in the full sampling scan.

Since PI was introduced, it has been shown that the image quality of PI depends on the number and geometrical arrangement of the individual coil elements with respect to the object of interest, while optimizing each coil element size.^60–62 Coil development has been actively carried out to reduce the g-factor in order to improve its performance. Until the mid-2000s, the number of independent receiver channels was typically limited to six to eight. However, thanks to the availability of new spectrometer hardware, there has been a fundamental trend toward considerably increasing the number of coil elements used for reception. Since the late 2000s, coils with up to 32 elements^63–66 have been introduced to boost the SNR, hence allowing for an increasing R. Even more complex coil arrays, consisting of up to 128 individual elements, have been proposed and realized.^67–70 Also, as the coil coverage increases with the increasing number of coil elements, the positioning of the coil with respect to the object of interest becomes less critical, thereby simplifying the setup and patient preparation process – but only if done in combination with an automatic coil selection algorithm.⁷¹

A method of incorporating SNR optimization into the image reconstruction algorithm of PI has also been reported. Sensitivity encoding (SENSE), one of the PI methods reported in 1999, operates in the image space and uses prior knowledge of the coil sensitivities of the receiver coil elements to disentangle the folded image data.⁴⁸ This prior knowledge feeds the SENSE inversion problem, which can be optimally solved by a least squares algorithm. SENSE is a solution to an overdetermined problem. In the well-known SENSE solution, by utilizing the receiver noise matrix, this overdetermination of the reconstruction inversion problem is exploited to optimize the SNR in the image.⁷² Here, the noise matrix describes the levels and correlation of noise in the receiver channels, as determined by averaging reasonably large sets of simultaneously acquired samples.⁴⁸ Further optimization of the SNR is also achieved by minimizing the signal in the background using regularization.

Once the patient and coil positioning are determined, the SNR optimization can be simulated prior to the actual scan by utilizing the noise matrix as well as coil sensitivity information. Automatic selection of the optimal combination of coil elements according to the imaging position under the wide sensitivity coverage of a large array has been developed to enhance the SNR.⁷¹ Also, a coil compression technique has been developed in which the data acquired with the selected coil subset can be reduced to alleviate the image reconstruction speed and memory constraints.⁷³ These techniques have led to improvements in workflow as well as consistent image quality in routine clinical examinations.

PI has been a favored choice for ssh-EPI DWI. However, even ssh-EPI with most recent coils and PI algorithms can present significant artifacts in some applications or certain patient conditions. For further reduction of these artifacts, further increase of R is required, but at the cost of SNR degradation as mentioned above. In particular, PI with excessive R usually suffers from a central noise band-like artifact because of the high g-factor from the coil due to typical coil geometry characteristics.^74,75 In recent years, compressed sensing (CS) image acquisition and reconstruction techniques have been used in MRI.⁷⁶ CS is an acceleration technique that requires random undersampling of k-space data, high sparsity of the image signal, and nonlinear reconstruction with iterative reconstruction. The PI problem can also be solved using iterative reconstruction,⁴⁸ an implementation which was originally introduced as the adaption of PI for use with data acquired with arbitrary k-space sampling. Therefore, PI and CS can be combined in a common iterative reconstruction loop. The clinical use of the combination of PI and CS has already been demonstrated; for instance, an iterative L1-regularized denoising filter, such as a wavelet-based one, is used to achieve an optimal balance between noise reduction and data consistency.^77,78

It is straightforward to combine PI and CS only to non-single-shot EPI scans because random sampling can be implemented easily. On the other hand, single-shot EPI sequence requires regular sampling in general as the steps in k-space are at an equal distance, corresponding to the blip gradient in the phase-encoding direction. However, although regular sampling may not be ideal, it has been shown that the integration of PI and L1-regularized denoising is beneficial also in this case.^79–84 Compressed SENSE, the integration of SENSE and L1-regularized denoising in an iterative reconstruction loop, is normally used with random sampling. Recently, however, it has also been applied to single-shot EPI (EPI with compressed SENSE [EPICS]).^79–82,84 In this platform, the SENSE problem is solved iteratively combined with L1 regularization for sparse wavelet transformed images. This approach resulted in a reduction of g-factor-related noise and improvement of image quality without further optimization of the single-shot EPI sampling scheme.^79–82,84 The image quality improvement and clinical usefulness of DWI with EPICS (EPICS-DWI) in the upper abdomen has also been demonstrated,^85,86 showing EPICS-DWI’s feasibility in DW imaging of the upper abdomen and significantly improved image quality compared with SENSE-DWI in an aggressive setting using R = 3. In Fig. 3, a comparison of SENSE and EPICS is shown, in which EPICS demonstrates improved image quality in terms of the noise in the central part of the abdomen.

Fig. 3.

Representative images of respiratory-triggered abdominal DWI with b-value of 800 s/mm² with SENSE (a) and EPICS using a reduction factor of 4. EPICS improves image quality by reducing g-factor-related noise in the central part of the image, which overlaps with the liver and pancreas. DWI, diffusion-weighted imaging; EPI, echo-planar imaging; EPICS, EPI with compressed SENSE; SENSE, sensitivity encoding.

Motion Correction for DWI

DWI in the body is highly sensitive to physiological motion such as heartbeat and respiration. Motion can cause ghosting and blurring artifacts, resulting in poor image quality as well as in inaccurate ADC quantification. In DWI, three types of motion effect need to be addressed. The first is signal attenuation due to phase dispersions within the voxel. This can happen if significant motion occurs while an MPG is applied. Second, ghost artifacts due to phase variations in k-space can occur due to significant motion in the intervals between acquisition shots. Third are misregistration artifacts due to the position shift by (mainly) respiratory motion between the single-shot acquisitions used for averaging. The first is, in general, not prominent. Indeed, motion can mostly be regarded as rigid-body bulk motion during MPG, which is less than 100 ms long in most cases. The second can be solved by using ssh-EPI as discussed in the previous section. Therefore, the main issue that needs to be addressed with body DWI and movement is mitigating misregistration artifacts.

To suppress the influence of misregistration artifacts, the simplest approach is to use ssh-EPI with a single breath-hold.^40,87–89 However, since a higher SNR and higher spatial resolution are desired in most body DWI applications, multiple breath-holds for higher averaging and the resulting possible misregistration are often unavoidable.⁹⁰ Respiratory-triggering^91–94 and navigator-based techniques^95–97 have been used to mitigate misregistration artifacts while allowing thinner slices or higher-resolution imaging.

Respiratory triggering uses an air-filled pressure sensor with a respiration belt or a contact-less camera-based respiratory monitoring sensor to constantly track the respiratory phase.^91,92 This technique allows acquisition of the image in the same respiratory phase, which can be averaged to obtain a high SNR image.^93,94 However, respiratory triggering increases scan time.

The use of a navigator is another approach to mitigate motion artifacts. In a navigator-based technique, the location of the diaphragm is constantly measured by the navigator and used to trigger the scan.^95,96 This approach was reported to enable a more precise ADC measurement than the breath-hold approach.⁹⁷

The free-breathing approach is a relatively time-efficient technique compared to the respiratory-triggering approach, allowing thinner slices or higher-resolution imaging compared to the breath-hold approach.^98,99 This approach is based on the fact that, in general, the expiratory period of respiration is longer than the inspiratory period, so even if imaging is performed without triggering, the averaged image of multiple acquisitions is predominantly weighted by signal obtained in the expiratory period.¹⁰⁰ Also, the scan time or acquisition interval does not depend on the variations of the respiratory cycle during acquisition. Due to these advantages, the free-breathing approach is commonly used for diffusion-weighted whole body imaging with background body signal suppression (DWIBS) although image sharpness may not be as good as in the respiratory-triggered approach.²⁸ Further, several reports have suggested its effectiveness in terms of image quality or ADC quantification in liver.^98,99

More studies need to be done to determine which methods are best for mitigating respiratory motion in abdominal DWI and acquiring reproducible ADC values. It has to be noted that averaging of multiple acquisitions must be done in the image-space domain because even coherent motion across multiple acquisitions can result in phase jumps, and thus artifacts, when they are averaged in the k-space domain.^15–18,28 Figure 4 presents a comparison of liver DWI with a b-value of 1000 s/mm² with respiratory-triggering (scan time 5:20), free-breathing (scan time 3:49), and breath-holding (scan time 0:16) acquisitions. Respiratory triggering showed the best image quality with high SNR and sharpness. The free-breathing image is also of good quality, but image blurring is more obvious compared to the others. The breath-holding image suffers from low SNR due to the limitation of scan time.

Fig. 4.

Comparison of liver DWI with b-value of 1000 s/mm² with RT (a), FB (b), and BH (c) acquisitions. Actual scan time for RT was 5:20, for FB was 3:49, and for BH was 0:16. RT shows the best image quality with a high SNR and sharpness. The FB image also has good quality, but blurring is more obvious. The BH image suffers from low SNR due to the limitation of (short) scan time. BH, breath-holding; DWI, diffusion-weighted imaging; FB, free-breathing; RT, respiratory-triggering.

Since, in general, the motion during the short period of diffusion sensitizing can be regarded as rigid bulk motion, there is no tissue signal loss. However, cardiac motion is known to result in signal loss particularly in the left liver lobe, as cardiac motion causes a compression of this part of the liver and the motion is no longer rigid-body motion. This leads to an overestimation of the ADC value.^101,102 Several approaches have been reported to tackle this problem, such as modeling stochastic attenuation bulk motion,¹⁰³ optimizing cardiac trigger delay,¹⁰⁴ modifying the signal averaging weights,^100,105 or correcting the signal dropout by deep learning.¹⁰⁶ The use of motion-compensated (MoCo) diffusion-encoding gradient waveforms may play a major role here. MoCo allows the nulling of first- and/or second-order gradient moments, making the imaging less susceptible to motion. Although MoCo tends to prolong TE and reduce SNR, several groups have investigated the optimization of diffusion-encoding waveform to shorten the TE by formulating this as constrained optimization problems.^107,108 Suppressing the influence of cardiac motion is an active area of research that will continue to be explored.

Fat Suppression

Homogenous fat suppression is key to good image quality in body DWI. Generally, achieving a uniform magnetic field becomes technologically more challenging at higher magnetic fields. Thus, as 3T systems started to be widely used in daily clinical practice, homogenous fat suppression became a major challenge. EPI acquisition, commonly used in body DWI, is sensitive to magnetic field inhomogeneities and can also produce severe chemical shift artifacts.¹⁰⁹ A chemical shift artifact is due to the displacement of water and fat signals, which originates from the difference in Larmor frequency between the two. This effect is prominent in the phase-encoding direction in EPI.¹⁶ Fat molecules are heavier than water molecules and thus diffuse more slowly.¹¹⁰ In sequences used in clinical studies, the attenuation of fat signal due to diffusion is so small that it can safely be ignored. Therefore, ADC of fat can be approximated as zero for sequences used in clinical cases.¹¹¹ When not suppressed adequately, the fat signal will contaminate the ADC value calculation in voxels where water and fat signals coexist as well as produce artifactual signal in the image.^111,112 Hence, it is extremely important to have optimal fat suppression. A few methods have been developed to suppress the fat signal, particularly at higher fields where the chemical shift between water and fat is larger.^113–117 In the following, some of these techniques, namely spectral inversion recovery (SPIR), spectral adiabatic inversion recovery (SPAIR), short TI inversion recovery (STIR), and slice-selective gradient reversal (SSGR), are described.

SPIR and SPAIR are a hybrid of chemical shift-selective and inversion-recovery techniques. The flip angle for the inversion pulse is generally lower than 180° for SPIR, but SPAIR uses 180° adiabatic pulses.¹¹⁴ These techniques work particularly well at 3T, where the chemical shift between water and fat is larger. Because of the use of adiabatic pulses, SPAIR is less sensitive to B1 inhomogeneity than SPIR. Thus, especially at higher fields (because B1 inhomogeneity can become an issue, especially for large FOV), SPAIR is more suited to achieve homogeneous fat suppression if the delay time is adjusted appropriately.^99,118 STIR consists of a broadband inversion pulse and is reported to work better than other techniques at larger volume and higher field.²⁸ Due to this feature, STIR is often used for DWIBS.²⁸

SSGR uses slice-selection gradients of opposing polarity for the excitation and the refocusing pulse.¹¹⁵ Due to the chemical shift, fat protons excited by the excitation pulse are not refocused and thus do not form a spin-echo signal.^116,117 SSGR has several advantages compared to other methods: SSGR is specific absorption rate (SAR)-efficient and can also be combined with other fat suppression techniques, such as SPIR or STIR, without any additional time penalty.^119–121 However, SSGR requires a large chemical shift between water and fat; therefore, it is more effective at 3T. The usefulness of SSGR has been clinically demonstrated for breast,¹²¹ liver,¹¹⁷ and WB-DWI.¹²²

Despite the many fat suppression techniques being developed, challenges remain. Chemical shift-selective fat suppression methods cannot eliminate the fat signal originating from fatty acids that have a chemical shift close to water (4.7 ppm). For example, olefinic protons, which make up approximately 5%–10% of all fat protons, have chemical shifts of around 5.4 ppm.^123,124 Complete suppression of this fat signal is challenging, and further work needs to be done. Figure 5 shows a comparison of liver DWI with b-value of 1000 s/mm² with different fat suppression techniques (no fat suppression, SPAIR, SSGR, and combination of SPAIR and SSGR) acquired with 3T MRI. Generally, in SPAIR images, certain parts of subcutaneous fat still exhibit high signal. In SSGR images, on the other hand, a ghost artifact can appear. Combined use of SPAIR and SSGR showed the best fat suppression quality, without residual high-intensity subcutaneous fat signals and ghosting artifacts.

Fig. 5.

Comparison of liver DWI with b-value of 1000 s/mm² with different fat suppression techniques acquired with 3T MRI. Without fat suppression (a), with SPAIR (b), with SSGR (c), and combined use of SPAIR and SSGR (d) are shown. In the SPAIR image, a certain part of subcutaneous fat still exhibits high signal (yellow arrows). In the SSGR image, a ghost artifact exists (light blue arrowheads). Combined use of SPAIR and SSGR show the best fat suppression without residual high intensity subcutaneous fat signal and ghosting artifacts. DWI, diffusion-weighted imaging; SPAIR, spectral attenuated inversion recovery; SSGR, slice-selective gradient reversal.

Current and Future Clinical Applications of Abdominal DWI

Current applications

There is growing evidence supporting abdominal DWI in lesion detection and follow-up, thanks to high lesion conspicuity and its non-invasive nature without contrast enhancement.^28,125–127 The clinical usefulness of DWI in the abdomen has been demonstrated,^28,125–127 and DWI has been incorporated into clinical guidelines such as the Liver Imaging Reporting and Data System (LI-RADS),¹²⁸ Prostate Imaging Reporting and Data System (PI-RADS),¹²⁹ Vesical Imaging Reporting and Data System (VI-RADS),¹³⁰ METastasis Reporting and Data System for Prostate Cancer (MET-RADS-P),¹³¹ and Myeloma Response Assessment and Diagnosis System (MY-RADS).¹³² Furthermore, in a review article by Summers et al., DWI is suggested as part of the WB-MRI protocols for breast and ovarian cancer, lymphoma, and screening.¹³³ It is currently used in both oncological^125,127 and non-oncological¹³⁴ applications. In this section, we describe the usefulness and achievements of abdominal DWI in clinical practice for both oncological and non-oncological applications, and future applications will be addressed.

For the liver, DWI is useful in the detection of small lesions around vessels and in the periphery of the liver that can be challenging to detect in routine T2-weighted imaging.^135,136 A comprehensive meta-analysis of 39 articles indicated that DWI can improve diagnostic performance (sensitivity: 90.6%–95.5%) for the detection of liver metastases when combined with gadoxetic acid-enhanced MRI.¹²⁵ Therefore, DWI is currently recommended as an essential protocol for the detection of liver metastases from colorectal, pancreatic, and neuroendocrine primaries in clinical practice.^125,137,138 DWI has also been found useful in detecting primary hepatic malignancies such as hepatocellular carcinoma (HCC) and cholangiocarcinoma. A combination of hyperintensity in diffusion-weighted images and arterial hyperenhancement with contrast media injection results in increased sensitivity for diagnosis of HCC, particularly for small HCC of < 20 mm.^139,140

Recently, it has been reported that diffusion kurtosis imaging (DKI)-derived K, the apparent kurtosis coefficient,¹⁴¹ can help the assessment of post-therapeutic response in HCC.¹⁴² DKI could provide more information on tissue structure than does conventional monoexponential analysis. In addition to lesion detection, several studies^89,143 have attempted characterization of liver lesions by using DWI and showed a significant difference in ADC between benign and malignant liver lesions. However, it has been reported that ADC needs to be used in combination with other morphological imaging, as there is a large overlap of ADC between benign and malignant lesions.^89,143,144

For non-oncological applications, several studies have shown that ADC in cirrhotic liver is lower than in normal liver and higher grades of fibrosis are associated with lower ADC values.^{87,88,145,146} However, there is still overlap of ADC among fibrosis stages.^145,146 The decrease in water diffusivity associated with hepatic fibrosis is thought to be multifactorial and may be due to increased connective tissue in the liver, resulting in diffusion restricted more than in normal liver parenchyma.^93,146–151 K indicated significant differences among normal and early hepatic fibrosis (F0 and F1), substantial stages (F2 and F3), and advanced hepatic fibrosis (F4).¹⁵² DKI may be useful for the classification of fibrosis stage.

WB-DWI²⁸ was introduced in 2004 as DWIBS. DWIBS is mainly used in oncology for overall screening^34,153,154 and follow-up.^34,155 It has been shown to be advantageous for evaluating myeloma,^31,132 lymphoma,^33,156 and bone diseases such as metastatic prostate^131,157,158 and breast cancer.^159,160 Technically, WB-DWI has two main features. First, free-breathing acquisition is used.^28,51,57 As mentioned in the previous section, the shorter scan time and patient comfort in free-breathing acquisition are an advantage compared to respiratory-triggered acquisition. Because WB-DWI requires multiple image stacks (usually three or four), scan time reduction by free-breathing is an important benefit. Second, STIR is used for reliable fat suppression. Uniform fat suppression is essential for lesion detection in WB-DWI, as an unsuppressed fat signal can lead to false positive diagnosis and erroneously calculated ADC values.

The diffusion contrast is mainly determined by the b-value. A higher b-value can increase lesion conspicuity on DWI. However, using a high b-value decreases the SNR. Optimal b-values depend on the organs of interest and the lesions and are still under investigation. Taouli et al. reported optimal b-values for several abdominal organs in a review article from an ISMRM-sponsored workshop.¹³ For liver, the use of a b-value of zero for ADC quantification is not recommended, in order to decrease the influence of perfusion. At least two or three b-values should be used for clinical purposes. When using two b-values, minimum b-value of < 100 s/mm² and maximum b-value between 500 and 1000 s/mm² are suggested. For three b-values, an additional intermediate b-value between 400 and 500 s/mm² is proposed in addition to the previously mentioned two b-values. For pancreas, the use of three b-values (for example, 0, 150, and 1000 s/mm²) to detect focal pancreatic lesions and extract ADC to improve characterization should be used. For kidney, it is mentioned that ADC calculated at b-values of 0 and 800 s/mm² is more effective in characterization of some lesions compared to ADC calculated at b-values of 0 and 500 s/mm². Goshima et al. reported that low b-value such as 100 s/mm² and high b-value such as 800 s/mm² are recommended in the detection and characterization of benign and malignant hepatic lesions.¹⁶¹

Future applications

Simultaneous multiple-slice excitation

The PI concept can be applied to the simultaneous multiple slice-excitation technique (multi-band [MB]) referred to as MB-PI.¹⁶² MB-PI allows multiple-slice acquisition at a single excitation RF pulse by unfolding the image data from multiple slices using a PI technique like SENSE or generalized autocalibrating partially parallel acquisitions (GRAPPA).¹⁶³ Several studies have reported the clinical utility of abdominal DWI using MB-PI for scan time reduction.^164–166

The RF power needed increases with the number of excited slices. One solution is to increase RF pulse duration. However, extended RF pulse duration also leads to decreased RF bandwidth, increased TE, and signal decay during the RF pulse.^167–169 Another solution is to increase the peak B1 amplitude, but it can reach the hardware limit of a clinical system, especially at higher magnetic field systems.^167–169 A better solution is to combine MB with variable-rate selective excitation (VERSE) technology which can reduce RF pulse duration subject to constraint of the peak B1 amplitude, using time-variable gradient waveforms.^170,171 Several in vivo validations have already been reported for upper abdominal DWI using MB-PI with VERSE, showing reduction of scan time in the respiratory-triggering approach¹⁷² or breath-holding time.¹⁷³ Clinical feasibility has also been shown for detecting HCC.¹⁷³ A comparison of conventional SENSE-DWI, MB-SENSE-DWI, and MB-VERSE-DWI is presented in Fig. 6. Both MB cases applied an MB factor of 2, resulting in a scan time that is halved compared with the conventional SENSE scan. Still, MB VERSE images showed nearly the same image quality as conventional SENSE images, and better image quality compared with MB SENSE images.

Fig. 6.

Comparison of respiratory-triggered liver DWI with b-value of 800 s/mm² with conventional SENSE (a), with MB SENSE (b), and with MB SENSE with VERSE for excitation and refocusing pulses of single-shot EPI (c). Both MB cases applied an MB factor of 2, resulting in halved scan time (1:08) compared with the conventional SENSE scan (2:18). MB VERSE images show nearly the same image quality as conventional SENSE images and better image quality compared with MB SENSE images. DWI, diffusion-weighted imaging; EPI, echo-planar imaging; MB, multi-band; SENSE, sensitivity encoding; VERSE, variable-rate selective excitation.

PI, CS, and deep learning

EPICS, which combines EPI with compressed SENSE, can reduce g-factor-related noise and improve image quality.^{80–82,85,86} Furthermore, a novel approach that replaces the sparse transform with deep-learning in the CS-PI reconstruction framework has been introduced. This new approach, based on Adaptive-CS-Net,¹⁷⁴ is referred to as compressed SENSE-AI.¹⁷⁵ Preliminary results show that WB-DWI with compressed SENSE-AI reduces g-factor-related noise and increases the SNR compared to conventional DWI with SENSE.¹⁷⁵ In Fig. 7, a comparison of WB-DWI images acquired with SENSE, EPICS, and compressed SENSE-AI is shown. EPICS reduces noise, especially g-factor-related noise, but compressed SENSE-AI reduces noise even further, resulting in a noticeable improvement in the SNR. This technique has great potential to improve diagnostic performance for WB-DWI.

Fig. 7.

Direct coronal whole-body DWI acquired with SENSE (a), EPICS (b), and CS-AI (c). These images were acquired with the following parameters: voxel size =  4.5 × 4.5 × 4.5 mm³, b-value = 1000 s/mm², acceleration factor =  5.5, and scan time =  3:15 (per station). C-SENSE reduces noise compared to SENSE, but CS-AI reduces noise even more. CS-AI further improves the image contrast and shows better visualization of the spinal cord, spine vertebrae, and lymph nodes. AI, artificial intelligence; CS, compressed SENSE; DWI, diffusion-weighted imaging; EPICS, echo-planar imaging with compressed SENSE; SENSE, sensitivity encoding.

IVIM

In clinical practice, the ADC is often used to reflect diffusion. However, ADC can be influenced by perfusion effects, which makes it difficult to assess the pure diffusion phenomenon. In general, the orientation of capillary vessels within a voxel can be assumed to be random. Flow through these capillaries then mimics diffusion. Therefore, blood microcirculation as well as water diffusion lead to signal attenuation in the presence of MPGs.¹⁷⁶ To disentangle perfusion and diffusion effects, Le Bihan et al. introduced the concept of IVIM, which evaluates the quantitative parameters that separately reflect tissue diffusivity and tissue microcapillary perfusion, based on a biexponential model.^36,177–179 Some quantitative parameters related to IVIM are calculated using the following equation:

$$S_b=S_0\cdot \left(F\cdot \exp \left[-b{D}^{*}\right]+\left[1-F\right]\cdot \exp \left[-bD\right]\right),$$ (4)

where D* and F are the perfusion-related pseudo diffusion coefficient and perfusion fraction, respectively, and D is the ADC. S_b is the signal intensity at a given b-value, and S₀ represents the signal intensity measured at a b-value of zero. A key feature of IVIM diffusion MR imaging is that it does not involve contrast agents, and it may serve as an interesting alternative to perfusion MRI in patients with contraindications to contrast agents or patients with renal failure at risk of nephrogenic systemic fibrosis¹⁸⁰ or for gadolinium deposits in brain basal ganglia.¹⁸¹ Klauß et al. reported an excellent correlation between perfusion fraction, F, and microvessel density in pancreas tumors (r = 0.85).¹⁷⁷ Therefore, F could be used as a non-invasive and quantitative imaging measure to directly assess the volume fraction of capillaries. Guiu et al.¹⁷⁸ also found decreased parenchymal perfusion in liver steatosis with IVIM. However, there are a couple of problems that need to be addressed. First, D* is not stable. Several studies have suggested that D* is less reproducible than D^179,182,183 and could be strongly affected by the cardiac cycle.^183,184 Furthermore, there is no established clinical significance for this parameter.^{177,185–187} Second, no consensus has been reached regarding the best selection of b-values. This is important, as scanning with multiple b-values to assess IVIM obviously leads to a longer scan time. Of interest is that only three b-values suffice to calculate F and D,¹⁸⁸ thus reducing the scan time as often there is no need to calculate D*.¹⁸⁹ This approach could address the main two drawbacks of IVIM. Its applicability to clinical practice requires further investigation.

Virtual MR elastography

Virtual MR elastography (VMRE) was introduced by Le Bihan et al.¹⁹⁰ VMRE involves the calculation of ADC based on DWI with b-values of 200 and 1500 s/mm², called shifted ADC (sADC). These specific b-values were chosen to optimally reflect the increased connective tissue in fibrotic liver parenchyma, causing an increased hindrance in water diffusion.¹²⁷ The sADC values can then be converted to a diffusion-based shear modulus by using a linear equation. In a recent study reported by Kromrey et al.,¹⁹¹ DWI-derived liver tissue elasticity strongly agreed with MRE. Further studies are needed to assess the usefulness of DWI-derived elasticity in clinical practice.

Oscillating gradient spin echo

Oscillating gradient spin echo (OGSE) was introduced to allow DWI scanning with much shorter diffusion times than conventional pulsed gradient spin echo (PGSE) sequences.^192–194 Mapping hepatocyte size in the liver using a broad range of diffusion times of approximately 2.5–70 ms achieved with OGSE and PGSE has been reported.¹⁹⁵ Cell size influences the ADC, especially at shorter diffusion times. A model equation then allows quantification of the cell size. DWI-derived microstructural parameters, such as hepatocyte size, agreed well with histological findings. However, at this time, clinical application of OGSE is very limited due to the achievable gradient strength and slew rate in clinical MR scanners, which result in an achievable shortest diffusion time of about 5 ms.¹⁹⁵

Conclusion

Abdominal DWI has made dramatic technical progress, leading to greater acceptance in clinical practice. It has the potential to develop further, thanks to scan acceleration and image quality improvement driven by technological advancement. The exploration of new physiological parameters will continue to be of major interest. Importantly, the accumulation of clinical proof will drive clinical acceptance.