Quantitative Diffusion MRI of the Abdomen and Pelvis

Abstract

Diffusion MRI has enormous potential and utility in the evaluation of various abdominal and pelvic disease processes, including cancer and non-cancer imaging of the liver, prostate, and other organs. Quantitative diffusion MRI is based on acquisitions with multiple diffusion encodings followed by quantitative mapping of diffusion parameters that are sensitive to tissue microstructure. Compared to qualitative diffusion-weighted MRI, quantitative diffusion MRI can improve standardization of tissue characterization as needed for disease detection, staging, and treatment monitoring. However, similar to many other quantitative MRI methods, diffusion MRI faces multiple challenges, including acquisition artifacts, signal modeling limitations, and biological variability. In abdominal and pelvic diffusion MRI, technical acquisition challenges include physiologic motion (respiratory, peristaltic, and pulsatile), image distortions, and low signal-to-noise ratio. If unaddressed, these challenges lead to poor technical performance (bias and precision) and clinical outcomes of quantitative diffusion MRI. Emerging and novel technical developments seek to address these challenges and may enable reliable quantitative diffusion MRI of the abdomen and pelvis. Through systematic validation in phantoms, volunteers, and patients, including multi-center studies to assess reproducibility, these emerging techniques may finally demonstrate the potential of quantitative diffusion MRI for abdominal and pelvic imaging applications.

Introduction to diffusion MRI

Diffusion MRI Mechanism and Contrast:

Diffusion MRI has a unique ability to probe tissue microstructure with acquisitions that are sensitive to the Brownian motion of water molecules within the body ^1,2. As water molecules move randomly and interact with their surroundings (i.e., within the intra- and extracellular space), the diffusion-weighted signal reflects the degree of molecular motion: tissues where water molecules move freely with minimal barriers generate lower signal intensities than tissues where this motion is more restricted. For instance, the motion of water molecules in relatively highly-cellular structures, such as tumor, is generally more restricted (and leads to higher diffusion MRI signals) compared to normal liver or cerebrospinal fluid ³. Diffusion MRI, albeit with variations, effectively suppresses the signal of water molecules in normal tissues and preferentially preserves the signal from tissues with a more restricted diffusion pattern, which is often seen in abnormal tissue (e.g. tumor, acute infarct, abscess). The restriction of diffusion can be evaluated using a variety of quantitative measures (e.g., the apparent diffusion coefficient, ADC), which can aid in characterization of the finding of interest ⁴. For instance, an increase in ADC of a treated tumor on a post-treatment exam relative to the pre-treatment exam can indicate the disruption of the tumor cellular environment and hence a favorable response to treatment. Importantly, such tissue changes may occur in-advance of macroscopically measurable changes, such as the tumor size. In addition, the extent of diffusion restriction may indicate tumor grade or predict how the tumor will respond to certain treatments ^5,6. Importantly, the scale of molecular motion probed with diffusion MRI is on the order of microns to tens of microns, which is much smaller than the voxel size that is realistically attainable with in vivo MRI in the clinical setting (on the order of millimeters).

Diffusion MRI Pulse Sequences:

Diffusion MRI is typically based on spin-echo pulse sequences, although alternatives such as dual spin-echo or stimulated-echo may be preferable for certain specific scenarios ^1,7. The inclusion of diffusion-weighting (DW) gradient waveforms within an MRI pulse sequence sensitizes the acquisition to the Brownian motion of the imaged molecules. These DW gradients lead to motion-dependent phase accumulation in moving nuclei, which in turn leads to signal dephasing for an ensemble of nuclei (e.g., a voxel). By controlling the shape (including amplitude and timing) of the DW gradient waveforms, the degree of diffusion weighing can be modulated and is usually quantified in terms of the b-value (s/mm²), which is defined as:

$$b={\gamma }^2{\int }_0^{T_{Diff}}{\left({\int }_0^{t}G(\tau )d\tau \right)}^{2}dt$$ (1)

where G(t) is the gradient waveform, γ is the gyromagnetic ratio, and T_Diff is the total diffusion encoding time (i.e., the duration of the diffusion gradient waveform). For simplicity, the definition in Equation 1 assumes no refocusing pulses in the pulse sequence. The presence of refocusing pulses can be readily incorporated through the appropriate sign changes in G(t). Typically, Stejskal-Tanner diffusion gradients are used in Spin-Echo DW acquisitions, where a monopolar trapezoidal gradient is applied before and after the 180° refocusing pulse, respectively ⁸. With Stejskal-Tanner diffusion gradients, and under the approximation of rectangular waveforms (i.e., negligible rise time), Equation 1 can be simplified as:

$$b=(\gamma \cdot G\cdot \delta {)}^2\cdot \left(\Delta -\frac{\delta }{3}\right)$$ (2)

where G is the amplitude of monopolar gradients during the rectangular waveform, δ is the duration, and Δ is the interval between the beginning of the first gradient lobe to the beginning of the second lobe.

Importantly, the high motion sensitivity inherent in diffusion MRI leads to substantial imaging challenges ^4,7. For example, severe artifacts may appear in standard multi-shot acquisitions where different portions of k-space are obtained following separate signal excitations, due to the presence of inconsistent motion-related phase offsets across shots. For this reason, diffusion MRI is typically obtained using rapid single-shot echo-planar imaging (ssEPI) acquisitions, where the entire 2D image is sampled following a single excitation and diffusion encoding. The use of ssEPI improves the reliability of diffusion MRI, however, at the expense of B₀ heterogeneity-related image distortions that arise due to the long EPI echo trains. For this reason, the use of ssEPI methods places restrictions on the spatial resolution achievable in diffusion MRI, particularly when imaging in the presence of a complex magnetic susceptibility environment (which leads to a heterogeneous B₀ field), such as in abdominal and pelvic MRI. Further, the need for DW gradient waveforms and long echo trains in diffusion MRI typically requires long echo times (TEs), which limits the achievable SNR when imaging tissues with short T₂ relaxation times, such as the liver ⁹. Despite these technical challenges, hardware and software advances over the past several decades have enabled the widespread use of diffusion MRI for clinical and research applications in the abdomen and pelvis ¹⁰, as discussed next.

Overview of this Roadmap Review:

This manuscript summarizes the current status, challenges, and opportunities for quantitative diffusion MRI of the abdomen and pelvis. The focus of the paper is on the translational development, validation, and application of quantitative diffusion MRI. A non-exhaustive overview of advanced methods, eg: based on biophysical models, is also provided. The remaining sections describe a roadmap for this rapidly evolving field, including methods for quantitative diffusion MRI, followed by technical and biological challenges, current status and recent trends for development, validation and dissemination, and finally opportunities and remaining work.

Quantitative Diffusion MRI:

Quantitative diffusion MRI methods seek to extract objective information that is sensitive to tissue microstructure, by acquiring multiple DW images and applying a signal model that enables the estimation of relevant parameters. These methods are based on the acquisition of multiple images that probe the desired tissue microstructure by varying one or more relevant parameters (particularly the b-value), while keeping others fixed (e.g., the TE). Subsequently, the acquired images are processed using a signal fitting algorithm based on a specific tissue model or signal representation.

Acquisitions for Quantitative Diffusion MRI:

In order to enable subsequent quantification, acquisitions for quantitative diffusion MRI obtain multiple images with different DW parameters. The requirements for the choice of DW parameters depend on the type of signal model to be applied (see details below). The simplest case of estimation of the ADC requires a minimum of two images with different b-values (often, one zero and one non-zero b-value). More sophisticated signal models with a higher number of free parameters require multiple b-values and/or multiple DW orientations, and may require further considerations for sampling of diffusion-encoding and additional parametric (e.g., signal relaxation) dimensions. Beyond the strict requirements to enable estimation with each signal model, optimization of the acquisition parameters (e.g., choice of b-values) is highly dependent on the expected tissue characteristics, the hardware limitations (e.g., gradient performance), as well as the presence of challenges such as physiological motion.

Diffusion Signal Models and Connection to Microstructure:

In order to quantify the sensitivity of DW signals to tissue microstructure in abdominal and pelvic applications (Figure 1), a variety of signal models have been proposed, including the following:

Figure 1.

Schematic representation of quantitative diffusion MRI. This pipeline includes: 1) the underlying tissue microstructure, with intracellular and extracellular space, microperfusion, etc., 2) the acquired signal (e.g., as a function of b-value), which is obtained by probing the tissue microstructure using diffusion MRI, and 3) the resulting quantitative parameters (e.g., apparent diffusion coefficient or intra-voxel incoherent motion parameters), which are obtained by fitting the acquired diffusion-weighted signals using an appropriate signal model.

Apparent Diffusion Coefficient (mono-exponential model).

The simplest diffusion signal model describes an exponential signal decay with increasing b-value, parameterized by the ADC ¹¹:

$$s(b)=A{e}^{-bADC}$$ (3)

where A represents the signal intensity at b=0 s/mm². The term “apparent” reflects the fact that ADC quantification does not explicitly account for various complexities in tissue microstructure and intra-voxel heterogeneity. Indeed, the ADC model, which accurately describes unrestricted Gaussian diffusion (e.g., in a water phantom) is not based on any explicit tissue assumptions. Importantly, ADC measurements from tissue will be affected by diffusion restrictions even if these restrictions are not explicitly described in the signal model, and even if the mono-exponential model closely fits the acquired signal. Nevertheless, ADC measurements provide an overall measure of diffusion signal decay rate, with multiple clinical applications as well as limitations that are discussed throughout the remainder of this manuscript.

Intra-voxel incoherent motion (IVIM).

In many organs (e.g., liver, kidneys), the diffusion signal observed with increasing b-value is clearly not mono-exponential. Specifically, low b-values reveal a rapidly decaying signal component, attributed to a pseudo-diffusion or perfusion compartment ¹². This rapidly decaying signal can be parameterized as an additive exponential decay with rapid decay rate (often denoted D*) and signal fraction f (i.e., leading to a bi-exponential overall signal model):

$$s(b)=A[(1-f)e^{-bD}+f{e}^{-bD\ast }]$$ (4)

where D* >> D. IVIM parameters are highly application-specific. As an illustration, typical measured values for liver IVIM parameters¹³ are D ≈ 1 × 10⁻³ mm²/s, D* ≈ 50 × 10⁻³ mm²/s (although there is high variability in D* measurements), and f ≈ 0.19. Alternatively, some IVIM methods attribute the rapidly decaying signal component to blood water molecules moving in random directions with velocity v_b (which can be estimated from the acquired data instead of D*) ^13-20.

Non-exponential models.

Multiple studies have demonstrated non-mono-exponential signal decay at high b-values in various organs, particularly at relatively long diffusion times ^21-24. This signal behavior is typically reflected in the ‘bending’ of the logarithm of the acquired signal as a function of b-value (see Figure 2). Various signal models have been shown to capture this signal behavior, including kurtosis, stretched exponentials, as well as multi-component models. These multi-component models may include several decaying signal components. In addition, multi-component models such as the restriction spectrum imaging (RSI) model ²⁵ may include anisotropic components, as discussed in the next section.

Figure 2.

Diffusion of water molecules in most tissues leads to non-mono-exponential signal decay with increasing b-value. However, multiple diffusion signal models may closely fit the acquired non-mono-exponential signal. This challenge leads to difficulty in model selection as well as limited specificity of most quantitative diffusion MRI methods. Hence, interpretation of quantitative diffusion MRI measurements in terms of specific microstructural tissue properties (e.g., intracellular vs. extracellular space) should be approached with caution.

Anisotropic diffusion.

The majority of quantitative diffusion MRI in the abdomen and pelvis relies on isotropic models (where diffusion signals are assumed to be independent of the DW orientation). However, anisotropic models have been applied in several abdominal and pelvic scenarios (e.g., kidneys, prostate, etc.), typically using a diffusion tensor imaging (DTI) approach ^26-30. In DTI processing, six parameters (describing a symmetric 3x3 diffusion tensor) are estimated, enabling the assessment of directional diffusion information. Although more sophisticated descriptions of anisotropic diffusion exist (e.g., diffusion spectrum imaging) ³¹, these are not widely used in abdominal and pelvic applications due to scan time limitations and low SNR.

Potential for Improved Value through Quantification:

Compared to qualitative diffusion MRI (where images with various contrasts are visualized, but no quantification is performed), quantitative diffusion MRI may enable improved standardization of tissue characterization, as needed for the detection, staging, and treatment monitoring of disease. For example, early ADC changes during therapy have been shown to occur before other effects (e.g., changes in the size of a lesion), and therefore early ADC changes may be predictive of response to treatment ³². Indeed, by removing intra- and inter-reader variability in image interpretation, quantitative diffusion MRI (similarly to other quantitative imaging methods) may provide added value in basic and clinical research (e.g., clinical trials with improved power due to reduced measurement variability) ³³, as well as clinical care ³⁴. Finally, quantitative diffusion MRI enables 'computed' diffusion MRI, i.e., synthesizing arbitrary high b-values beyond those acquired, with several promising applications ^35-37.

Performance of Quantitative Imaging:

Technical performance.

The technical performance of quantitative imaging methods can be evaluated in terms of bias, linearity, and precision ³⁸. The bias describes systematic measurement error relative to the ground truth, whereas linearity describes the ability to produce measurements that are proportional to the ground truth. Evaluation of bias and linearity requires the existence of an accepted reference measurement, which is often complicated in the case of in vivo diffusion MRI (see below). The precision describes the tendency of repeated measurements to provide similar values, regardless of the ground truth. Precision can be defined for repeated measurements obtained under identical experimental conditions (repeatability), or under different experimental conditions (reproducibility).

Clinical performance.

The clinical performance of quantitative imaging biomarkers describes their ability to perform clinically relevant tasks, such as detect a specific condition. Clinical performance is typically quantified in terms of diagnostic accuracy (including sensitivity, specificity, and positive and negative predictive values).

Tradeoffs in the Performance of Quantitative Diffusion MRI.

The choices of image acquisition parameters and signal modeling in diffusion MRI lead to frequent tradeoffs between bias and precision (e.g., more sophisticated models have the potential to reduce bias, but often at the cost of reduced precision). In addition, signal representations used in quantitative diffusion MRI are often not based on specific models of tissue microstructure, which can be viewed as attaining sensitivity at the cost of specificity ^14,39.

Challenges to Quantitative Diffusion MRI of the Abdomen and Pelvis:

Quantitative MRI methods face multiple challenges, including acquisition artifacts and limitations, signal modeling limitations, and biological variability ¹⁰. These challenges, which may lead to poor technical and/or clinical performance, are summarized next.

Acquisition artifacts and limitations:

Physiological motion.

A central challenge of diffusion MRI, particularly for abdominal applications, is its sensitivity to physiological motion ^3,40,41. Importantly, diffusion MRI seeks to achieve high sensitivity to microscopic molecular motion (on the order of microns to tens of microns) while remaining insensitive to macroscopic physiological motion from respiration, cardiovascular motion, peristalsis, and bulk patient motion (on the order of millimeters to centimeters) ⁴. However, physiological motion complicates quantitative diffusion MRI in several ways. Rapid compressive motion that occurs during the application of DW gradients (e.g., “intra-repetition” motion on a time scale of tens of ms, as introduced by cardiac motion and pulsation on the left lobe of the liver and other abdominal organs) leads to artifactual signal dropout in DW images. This signal dropout introduces bias in quantitative diffusion parameters (e.g., overestimation of ADC) as well as poor precision. Sometimes the affected tissues completely disappear from high b-value images (Figure 3). Slower motion that occurs during the entire DW acquisition (e.g., “inter-repetition” motion on a time scale of seconds to minutes, as caused by respiration or bulk patient motion) leads to mis-registration across repetitions and DW encodings. This mis-registration can introduce blurring and bias in subsequent quantitative parameter mapping, particularly in the setting of quantification of focal disease (Figure 4).

Figure 3.

Diffusion MRI has high potential for the detection and characterization of lesions in the liver, but also faces technical challenges. (Left) Axial T₁-weighted fat-suppressed delayed gadoxetate-enhanced image in a patient with metastatic breast cancer demonstrates lesions in both lobes of the liver (yellow arrows in the right lobe and red arrows in the left lobe). Diffusion-weighted imaging (DWI) and ADC map (middle and right, respectively), depict the right lobe lesions with excellent conspicuity (yellow arrows). However, the left lobe lesions are not evaluable on DWI and ADC map due to severe cardiac motion, which results in signal loss in the DWI image (white dashed circle), and thus extreme overestimation in the ADC map (red dashed circle). Furthermore, DWI and ADC suffer from artifact in the right lobe due to incomplete suppression of subcutaneous fat signal combined with chemical shift artifacts along the anterior-posterior phase encoding direction (white arrowheads).

Figure 4.

Inter-repetition motion induced by respiration can lead to mis-registration between individual repetitions and b-values, as well as bias and blurring in quantitative ADC mapping of the abdomen. Axial diffusion-weighted images (DWI) at b = 50 and 500 s/mm² as well as the ADC map (A, B, and C, respectively) in a patient with metastatic urothelial carcinoma who underwent an MRI abdomen for staging. Direct averaging over multiple repetitions results in blurring of organ boundaries (A, blue arrow). Image misregistration between different b-values results in different components of normal structures being seen on different b-value images, (e.g. middle hepatic vein in A and B, yellow arrows), or visualization of a lesion only on one image (e.g. a small lesion is only seen in A and not B, red arrows). Finally, misregistration results in artifactually heterogeneous ADC of lesions (C, white arrow).

Image distortions.

Diffusion MRI is generally acquired using single-shot EPI. Due to the long readout times of single-shot EPI, B₀ field heterogeneities (which are unavoidable in abdominal and pelvic MRI due to the complex magnetic susceptibility environment with multiple tissue/air interfaces) ⁴² lead to considerable artifactual phase accumulation during data collection ⁷. This time-dependent phase accumulation confounds the intended k-space encoding and leads to substantial distortions along the phase encoding dimension. For example, with a typical effective bandwidth of 25 Hz/pixel along the phase encoding dimension, a B₀-related off-resonance of 100 Hz (common in abdominal imaging) will lead to a 4-pixel shift. In contrast, the frequency encoding dimension, i.e., the “rapid” EPI dimension, acquired with a typical bandwidth of 3 kHz/pixel, generally does not lead to noticeable B₀-related distortions. Distortions can be extreme in the presence of metal implants, e.g., spine hardware and hip prostheses (Figure 5). Generally, the distortions introduced by static B₀ effects (e.g., susceptibility-related) are constant across diffusion encodings (b-values) as long as the EPI readout is the same. Other MRI artifacts, including eddy currents induced by the strong DW gradient waveforms, may also lead to distortions that increase with the diffusion weighting ⁴³.

Figure 5.

Left: Axial 3D T₁-weighted fat-suppressed 20-minute delayed gadoxetate-enhanced image in a patient with metastatic neuroendocrine tumor demonstrates a subcentimeter liver lesion (yellow arrow) that is consistent with a metastasis as verified by somatostatin-receptor PET imaging (not shown). Although the lesion is detectable on this image, its identification is challenging due to a combination of small size and proximity to the middle hepatic vein. Right: Diffusion-weighted imaging (DWI) demonstrates the lesion with excellent conspicuity (yellow arrow) and can aid the radiologist to look for and confirm the finding. Of note, extensive metal-associated artifact on DWI due to presence hardware in the spine (red arrow) results in nondiagnostic evaluation for tumor in this area, which is a well-recognized limitation of this technique.

Signal to Noise.

Diffusion MRI is typically a low-SNR technique, due to several reasons. The exquisite sensitivity of diffusion MRI to tissue microstructure emerges at relatively high diffusion weighting (b-value), which leads to substantial signal decay. In addition, the need for DW gradient waveforms and long EPI echo trains in the pulse sequence leads to long echo times, which result in low SNR particularly in tissues with short T₂ (e.g., liver). This challenge is especially acute in patients with iron overload, which results in substantially shorter T₂ and T₂^* with the corresponding loss in SNR for diffusion MRI acquisitions (Figure 6). Additional factors contribute to the challenging noise properties of diffusion MRI, including the prevalent application of parallel imaging acceleration (with corresponding SNR penalty), as needed to minimize image distortion in EPI, and the fact that quantitative diffusion MRI typically relies on magnitude images (after rectification), leading to substantial noise floor effects. For example, noise floor effects, which occur at higher b-values, appear as a bending of the logarithm of the signal. This effect may confound the measurement of quantitative diffusion parameters, e.g., by introducing bias (underestimation) in ADC ^27,44-46 (Figure 7). In more advanced non-mono-exponential signal models, noise floor effects may lead to an artifactual estimation of restricted diffusion (e.g., overestimation of the kurtosis parameter K) ⁴⁷. A variety of image reconstruction, filtering and post-processing techniques have been proposed to ameliorate noise floor effects in quantitative diffusion MRI ^27,47-49.

Figure 6.

Quantitative diffusion MRI is confounded by liver iron overload, which leads to very short relaxation times (T₂ and T₂*). In a patient with no iron overload (top row), liver diffusion MR signals can be adequately sampled to enable ADC mapping. However, in a patient with iron overload (bottom row), due to the short T₂ and T₂*, the signal has decayed away in the liver (dashed yellow contour) even at a low b-value (b = 50 s/mm²). This decay is due to nearly complete T₂ decay of the signal at the relatively long echo-times used in diffusion MRI (e.g., TE = 72 ms). This paucity of signal precludes the reliable estimation of quantitative diffusion parameters, such as ADC.

Figure 7.

Noise effects can lead to bias (underestimation) in ADC measurements. Although the noise distribution in complex MR images is well approximated as zero-mean Gaussian, the signal magnitude operation that is routinely performed in diffusion MRI (in order to avoid phase inconsistencies across acquisitions) alters the noise distribution. Indeed, quantitative diffusion MRI is based on magnitude signals, where the noise distribution is not zero-mean at low SNR (e.g., at high b-values). The presence of a non-zero noise floor leads to underestimation in ADC measurements, particularly for acquisitions with high b values or low SNR. For example, prostate diffusion MRI is often acquired using two separate series with b = (0,800) and b = (0,1500), respectively, where the low b-value series provides higher SNR and quantitative reliability, and the high b-value series provides improved contrast but leads to ADC underestimation. Note that this simulation does not include TE-dependent SNR differences between the two series, or diffusion time effects on ADC, in order to illustrate the noise bias effects. Importantly, this noise-dependent bias can lead to poor reproducibility across acquisition protocols with different SNR (e.g., using different systems, receive coils, pulse sequences, b-values, or spatial resolutions). In addition, noise effects confound the evaluation of more sophisticated signal models (not shown in the illustration), particularly when acquiring high b-values.

Other artifacts and limitations include fat suppression errors, gradient nonlinearities, partial volume effects and general imaging artifacts.

Fat suppression is critical in diffusion MRI, particularly in areas with large fat depots such as the abdomen and pelvis ^4,50,51. Fat has a long T₂ relaxation time and extremely low diffusion coefficient, which makes it appear bright in diffusion MRI. Further, fat signals have a large chemical shift (e.g., 3.4 ppm shift between the main methylene peak and the water peak), which leads to large spatial shifts (chemical shift artifact) along the phase encoding dimension in EPI acquisitions. Therefore, unsuppressed fat signals, e.g. from visceral or subcutaneous depots, may appear artifactually shifted to overlap with the organs of interest (Figure 3). Suppression of fat signals is typically performed in MRI based on the chemical shift effect (i.e., by avoiding excitation of signals that resonate at the expected fat frequency, while preserving signals that resonate at the expected water frequency). However, in the presence of B₀-induced frequency offsets as commonly observed in the abdomen, chemical shift-based fat suppression may fail, leading to bright fat signals that overlap with the organs of interest ^52,53. Alternative fat suppression methods based on inversion recovery fat nulling, which rely on the short and predictable T₁ relaxation time of fat signals (approximately 300-400 ms at clinical field strengths), have the potential for improved robustness at the cost of SNR ^54-56. Hybrid methods that rely on both chemical shift and T₁ are also available ^56,57.

Gradient nonlinearities, if present, lead to non-uniform b-values throughout the image (particularly away from isocenter), and introduce spatially-dependent quantification bias ⁵⁸.

Partial volume effects introduce bias in quantitative diffusion MRI, particularly in the assessment of small anatomical structures or small focal lesions. Partial volume effects constitute a challenge in diffusion MRI, which is obtained with limited in-plane spatial resolution to minimize distortions in the context of long-readout EPI acquisitions, and with relatively thick slices to preserve SNR.

General imaging artifacts, including parallel imaging artifacts or EPI ghosting, which deteriorate the image quality in diffusion MRI, will generally propagate into the subsequent quantitative maps and may lead to bias and poor precision.

Modeling limitations:

Incomplete signal model.

The use of excessively simple models (e.g., ADC in the presence of IVIM effects at low b-values, or not accounting for restricted diffusion effects at high b-values, or diffusion time dependence ⁵⁹) leads to poor data fitting and/or poor reproducibility (i.e., strong dependence of the estimated parameters on the acquisition parameters such as b-values).

Unstable signal estimation.

The use of sophisticated models can overcome some of the limitations of simple signal models such as ADC. However, increasing the model complexity and the number of free parameters can also lead to instability and noise amplification in model fitting ⁶⁰. This challenge is particularly acute in abdominal diffusion MRI applications (e.g., liver), where the presence of low SNR, motion artifacts, and other challenges, complicates the application of sophisticated models with many free parameters.

Choice of signal model

Multiple different non-mono-exponential models generally enable close fitting of the acquired data, and therefore selecting the optimal model is challenging (Figure 2). The choice is sometimes made based on studies of clinical accuracy (i.e., the ability of the model’s estimated parameters to detect a certain condition) ⁶¹, since technical bias can generally not be evaluated in vivo for signal models (representations) that are not based on specific tissue models. In some studies, the choice of model is based on statistical approaches such as the Akaike information criterion (AIC) ⁶², which seeks a balance between the model’s ability to closely fit the data, and the complexity (e.g., number of free parameters) in the model.

Lack of tissue model.

Fundamentally, many of the models applied in diffusion MRI (including ADC, DTI, and others) are better described as signal representations, rather than models of tissue microstructure ³⁹. Signal representations are mathematical expressions that empirically enable fitting of the acquired diffusion signal. Indeed, signal representations may have important clinical applications (e.g., in quantifying the difference between healthy and diseased tissue). However, because of the lack of a tissue modeling foundation, it may not be possible to evaluate the bias of methods based on signal representations as there is no ground truth for the measured parameters. Further, these methods often suffer from inherently poor reproducibility across different acquisition parameters ^40,63, which requires highly standardized acquisitions and quality control in order to enable reproducible quantification.

Biological variability:

In addition to the technical challenges described above, biological variability may confound the ability of quantitative diffusion measurements to evaluate disease ^3,64,65. For example, ADC in the prostate of healthy subjects may increase with age ⁶⁶ and ADC in various abdominal organs may vary with sex and age ⁶⁷. Further, different tumors have been shown to lead to different correlations between ADC and cellularity, possibly as a result of various other microstructural features that affect ADC (e.g., extracellular matrix, nucleic areas, ratio of stroma to parenchyma, or microvessel density) ^64,68.

Recent Technical Developments

Addressing acquisition artifacts and limitations:

Physiological motion

“Intra-repetition” motion artifacts in MRI, including those caused by cardiovascular motion and pulsation, can lead to localized signal dropouts in the DW signal ^69,70. These signal dropouts worsen with increasing b-value, and generally vary across repetitions. Post-processing methods to combine multiple repetitions based on statistical modeling ameliorate the signal dropouts ^69,71. However, these post-processing methods generally only provide a partial solution due to the unpredictable nature of the signal dropouts. Instead, this artifact can be largely avoided through the application of motion-compensated DW waveforms ^72-76. These DW waveforms rely on moment nulling, by setting to zero the first order motion moment M₁, and occasionally also the second moment M₂, where:

$$M_n={\int }_0^{T}dtG(t)t^n$$ (5)

Importantly, nulling the n^th moment leads to zero phase accumulation in hydrogen nuclei within molecules with n^th moment motion. For instance, first order moment nulling (M₁=0) leads to zero phase accumulation by molecules with constant velocity, whereas second order moment nulling (M₂=0) leads to zero phase accumulation by molecules with constant acceleration. Motion-compensated DW waveforms have been used for decades ⁷⁷, however classic DW waveforms based on simple analytical moment-nulled solutions result in substantial echo time penalties (lower SNR), which limits their utility in organs such as the liver where the T₂ is relatively short ⁷⁷. In recent years, optimized motion-compensated diffusion MRI methods have been proposed ^72,74,78. These methods rely on waveforms that are designed by solving an optimization problem (minimizing the TE) subject to a set of constraints that include the desired b-value, hardware constraints, and pulse sequence timing constraints. Optimized motion-compensated diffusion MRI methods enable motion-robust diffusion MRI with minimized TE penalty, i.e., optimized SNR, and have recently been shown to provide improved ADC quantification (e.g., better repeatability) compared to standard monopolar DW acquisitions ⁷⁹. These methods may enable improved diffusion MRI in organs such as the liver ^72,74,80 and pancreas ⁸¹.

Importantly, motion compensated diffusion MRI methods provide motion robustness by avoiding signal dropouts in the presence of compressive tissue motion. However, motion compensation (e.g., M₁=0) also leads to poor suppression of blood signals in highly perfused organs such as the liver. Residual signals from blood vessels appear bright in motion-compensated diffusion MRI and therefore may confound interpretation and quantification. In order to address this challenge, a DW method with non-zero but small (and constant across b-values) M₁ was introduced. This method, termed M₁-optimized diffusion imaging (MODI), has the potential to enable motion-robust, blood-suppressed, minimum-TE diffusion MRI of the liver ⁷³. In addition, a mixed waveform protocol for both reduction of cardiac motion artifact and blood-suppressed diffusion MRI of the liver was also proposed to address this problem ⁸⁰. Figure 8 shows an example of MODI acquisitions to enable motion-robust, blood-suppressed diffusion MRI of the liver.

Figure 8.

Two central challenges to diffusion MRI in the abdomen, distortion and motion-induced signal dropout, can be addressed by multi-shot (ms)EPI acquisitions and motion-compensated gradient waveforms, respectively. The images show example diffusion MRI acquisitions in a volunteer, comparing monopolar and motion-compensated waveforms, each with both single-shot (ss)EPI and msEPI acquisitions. The orange contours depict the expected visualization of liver based on T₂-weighted imaging as a reference. The misalignment between the contour and the liver anatomy due to distortion on diffusion MR image is marked by yellow arrows. Red arrows indicate motion-induced signal dropout and ADC bias in the left lobe, which are significantly reduced by the motion-compensated acquisition.

In diffusion MRI of the abdomen, inter-repetition motion due to respiration is typically addressed using breath-held or respiratory-triggered acquisitions ^82-86. Free-breathing acquisitions without gating or triggering are typically used in the pelvis, and sometimes used in the abdomen. However, different quantitative values have been reported for BH vs RT vs FB acquisitions in liver imaging ⁸⁶.

Distortion

The presence of severe image distortion is common in diffusion MRI, due to the use of EPI with long readouts in the presence of substantial B₀ magnetic field heterogeneities as observed in the abdomen and pelvis ⁷. For this reason, considerable research efforts have been devoted to obtaining diffusion MRI with reduced distortions, based on a variety of approaches:

Correcting distortions via post-processing, e.g., based on a measured B₀ field map ^87,88, or based on two acquisitions obtained with reversed k-space acquisition order ⁸⁹.

Shortening the EPI readout.

This can be achieved using parallel imaging acceleration (as routinely implemented in clinical imaging), reduced field-of-view acquisitions (which require shorter EPI readouts) ^90-93, or multi-shot acquisitions ^94-98 (where each shot covers a subset of the k-space, segmented either along the phase encoding or frequency encoding dimension). Figure 8 demonstrates an example where combining MODI motion-robust acquisitions with multi-shot EPI reduces image artifacts and distortion.

Non-EPI diffusion MRI.

By using non-EPI acquisitions, the presence of EPI-related distortions can be avoided, even in the presence of nearby metal implants. Techniques to mitigate this challenge have been proposed. Example non-EPI acquisitions include fast spin echo (FSE) and Periodically Rotated Overlapping ParallEL Lines with Enhanced Reconstruction (PROPELLER) ^99,100 (Figure 9) or MultiVane with short-tau inversion-recovery fat suppression ¹⁰¹. Another non-EPI approach is based on diffusion-preparation pulse sequences ^102,103. In these acquisitions, diffusion contrast is first encoded into the transverse magnetization signal, which is then flipped back onto the longitudinal axis. Subsequently, a variety of image acquisitions can be used, while maintaining inherent diffusion-weighting. Although these methods completely avoid EPI-related image distortions, they introduce additional challenges, including lower SNR efficiency, high SAR for FSE acquisitions, or motion-related inconsistency across shots in diffusion-preparation methods.

Figure 9.

Radial k-space sampling via PROPELLER (Periodically Rotated Overlapping ParallEL Lines with Enhanced Reconstruction) acquisitions enable distortion-free diffusion MRI, and the low SNR efficiency of PROPELLER may be ameliorated by denoising (e.g., DL-based) methods. The images show multishot DW-EPI and PROPELLER DW prostate images of a healthy volunteer. PROPELLER b=0 and b=600 images were reconstructed using conventional reconstruction and DL-based methods. Susceptibility artifacts (red arrow) are observed in EPI images, but not in PROPELLER images. While the PROPELLER images with conventional reconstruction showed low SNR, DL-based reconstruction enabled significantly increased SNR and good image sharpness, also enhancing the appearance of the ADC map.

Signal to Noise.

A variety of techniques enable improved SNR in diffusion MRI of the abdomen and pelvis. These developments include novel receive coil arrays, or specialized coils (e.g., endorectal coils, or wearable coils ¹⁰⁴ for pelvic MRI), improved gradient systems which enable shorter echo times, novel pulse sequences ¹⁰⁵, as well as post-processing based denoising methods ^106-109.

Other image acquisition developments

Additional technical improvements continue to emerge in the field, including improved fat suppression ⁵⁶, and eddy current suppression with optimized gradients ^110,111. These methods continue to advance diffusion MRI toward the goal of enabling reliable, artifact-free imaging, as required for precise quantification.

Signal modeling:

Multi-dimensional diffusion-relaxation methods have gained substantial interest in recent years ^112-116. In addition to sampling the b-value and (potentially) the diffusion orientation dimensions, additional dimensions may be sampled in diffusion MRI, in order to further probe the underlying tissue microstructure. Several diffusion-relaxation sampling methods have been developed, where diffusion properties (b-value), T₂ relaxation, and T₁ relaxation can be simultaneously examined in a multi-dimensional experiment (see Figure 10 for an example of ADC-T₁-T₂ mapping in the prostate) ¹¹⁷. Further, sampling of the diffusion time dimension in addition to the b-values enables enhanced probing of restricted diffusion (as diffusion restrictions become more apparent at long diffusion times) ^8,23,24. Finally, additional properties of the DW waveform can be explicitly sampled, including the first motion moment M₁, in order to enable improved assessment of diffusion and perfusion properties of the tissue ¹¹⁸. Generally, diffusion in tissues leads to signals that depend on the overall DW waveform rather than simply on the b-value, so careful design of DW waveforms may enable new biomarkers of tissue microstructure.

Figure 10.

Top row: A prostate MRI in a 69-year-old man demonstrates a highly suspicious lesion (red arrows) in the right peripheral zone (PI-RADS v2.1 score 5). Bottom row: T₁, T₂ and ADC maps are estimated simultaneously using the rapid STEM method ¹¹⁷. T2 and ADC measurements in the lesion are labeled in the figure. Corresponding histopathology demonstrates Grade Group 4 prostate adenocarcinoma.

Further, advanced diffusion MRI methods (particularly in brain imaging) often focus on tissue models rather than signal representations ¹¹⁹. Tissue models rely on an explicit description of tissue microstructure (e.g., the number and shape of tissue compartments, and assumptions on the molecular diffusion within each compartment), which lead to analytical or computational expressions for diffusion MRI signals and enable estimation of tissue parameters. In addition to their basic scientific value, tissue models are compelling for several practical reasons, including the potential to enable improved reproducibility, and the potential for improved validation of bias due to the explicit tissue microstructure assumptions. However, explicitly modeling the relevant tissue features is challenging and often requires high-SNR, artifact-free, high b-value signals. For these reasons, such advanced methods are often developed for brain imaging, and may subsequently be translated to other applications, including abdominal and pelvic imaging.

Some recent efforts in abdominopelvic diffusion MRI include the use of specialized diffusion-weighted acquisitions, followed by advanced signal models. For example, tensor-valued (non-linear) diffusion encoding has been used to estimate microscopic diffusion anisotropy in the prostate ¹²⁰. A combination of (conventional) pulsed gradient and oscillating gradient acquisitions followed by application of a two-compartment model (intra- and extra-cellular) has been proposed to estimate tissue parameters (including cell size and intra-cellular volume fraction, among others) and evaluated in colon cancer models^121,122. A multi-b-value acquisition followed by three-compartment modeling has been proposed for the evaluation of prostate cancer ^123,124. These biophysical-based methods have shown promise to overcome the limitations (e.g., limited specificity) of previous signal representations such as ADC or kurtosis models.

Current status of validation and dissemination of diffusion MRI

Technical validation

Phantom validation.

Evaluation of bias and precision of quantitative diffusion MRI methods benefits from access to phantoms with highly controlled properties ¹²⁵. Phantoms that provide reliable diffusion properties have been used in various studies, e.g., to assess bias ¹²⁶ and multi-center reproducibility ¹²⁷. Due to the high dependence of molecular diffusion on temperature, diffusion phantom MRI experiments require careful temperature control and/or temperature measurement. Temperature control is typically achieved using an ice-water bath to ensure a temperature of 0°C during scanning. Alternatively, temperature measurement while scanning is achievable using a variety of MR-compatible temperature sensors. For phantom construction, pure substance phantoms based on water, acetone, or various oils, are compelling due to their simplicity and the well-known diffusion properties of these substances across temperatures ¹²⁸. However, pure substance phantoms provide a limited number of diffusion coefficients. This limitation is overcome by using solution-based phantoms, where the inclusion of a known concentration of solute leads to a predictable decrease in the diffusion coefficient of the solvent. For example, solutions of polyvinylpyrrolidone (PVP) in water ¹²⁹ have been used in multiple studies and are commercially available (CaliberMRI, Boulder, CO).

To enable validation of advanced signal models, phantoms that enable diffusion restriction have been proposed. Although phantoms based on various easily obtainable substances (including cream and asparagus stems ¹³⁰) provide restricted diffusion, they offer a limited range of diffusion properties. This limitation has been addressed through a variety of synthetic phantoms (e.g., particle suspensions ¹³⁰, fibers ¹³¹, and liquid crystal systems ¹²⁷), which provide various diffusion restrictions, but also suffer from a limited range of diffusion properties.

In vivo validation.

In vivo technical validation studies of quantitative diffusion MRI methods are often focused on the evaluation of precision (including repeatability and reproducibility). Precision of ADC mapping varies widely across organs. A recent prospective multi-center precision study on healthy volunteers estimated the limits above which a change in ADC can be considered as significant, on a per-organ basis: ¹³² 16% for paraspinal muscle, 22% for renal cortex, 26% for prostate, 29% for renal medulla, 35% for the liver, and 45% for the spleen. One can see that depending on the pathology of interest and the organ in which it occurs, the poor precision may preclude meaningful and prospective quantitative utility of current diffusion MRI methods. One factor affecting the quantitative precision of ADC measurement is the region-of-interest (ROI) measurement strategy; for instance, large 3D ROIs provide improved repeatability compared to small 2D ROIs ¹²⁶. Reproducibility of ADC quantification across different types of acquisition pulse sequences has also been evaluated, including different respiratory mitigation strategies in the liver ^82,84-86 and different distortion reduction strategies in the prostate ⁹⁵. Beyond ADC mapping, the evaluation of precision of more advanced diffusion signal models has demonstrated mixed results. For example, IVIM of the liver generally enables repeatable and reproducible measurements of the diffusion coefficient D, but poorer repeatability and reproducibility of the perfusion fraction (f) and especially the pseudo-diffusion coefficient (D*) ^13,133,134.

As mentioned previously, there is often substantial heterogeneity in quantitative diffusion measurements across studies, which is due in part to the technical challenges in diffusion MRI of the abdomen and pelvis as well as the lack of standardization of acquisition parameters and reconstruction/processing methods. Importantly, this heterogeneity highlights the need for both technical developments that address technical challenges as well as improved standardization across imaging sites.

Current state of clinical dissemination, applications, and challenges of abdominal and pelvic diffusion MRI

Diffusion MRI is routinely acquired in nearly all abdominal and pelvic clinical MRI exams obtained for evaluation of various disease processes, whether neoplastic or infectious/inflammatory. Diffusion MRI is predominantly utilized for two purposes: (1) detection and (2) characterization of abnormalities. Among the most common indications for clinical abdominal MRI is the detection, characterization, and follow up of liver lesions ^{23,26,40,82,83,135-142}. In the setting of cancer, detection of liver lesions often changes the disease stage and hence how the patient is treated. Not only presence or absence of liver lesions is important, but often the number of liver lesions is of pivotal clinical significance. For instance, a limited number of metastases can be treated by surgical or ablative approach (often with a curative intent), whereas more widespread and/or presence of bilobar lesions often requires systemic treatment. Hence, successful detection of small liver lesions can make a difference in patient management. Gadoxetate-enhanced liver MRI is highly sensitive for the detection of liver metastases in the setting of colorectal cancer and neuroendocrine tumors due its high spatial resolution and retention of contrast by normal hepatocytes ^143-153. However, detection of lesions that are sub-centimeter and/or adjacent to vasculature on gadoxetate-enhanced images remain a challenge for the interpreting radiologist. Diffusion MRI has a unique ability to detect such lesions by suppressing the signal from background liver parenchyma and major vessels, and therefore serves as an invaluable tool for the detection of metastatic liver lesions (Figure 5).

Besides detection, diffusion MRI also improves characterization of liver metastases ^135,154. Importantly, quantitative diffusion measurements reflect a number of tissue properties (including cellularity, cell size, necrosis, and perfusion) that are relevant in the context of treatment monitoring. Further, these tissue effects often appear before macroscopic morphologic changes (e.g., alterations in lesion size). Therefore, quantitative diffusion MRI has also sparked interest for the early assessment of response to treatment in various cancer imaging applications ⁵. However, the potential impact of diffusion MRI for guiding personalized treatment remains largely unknown. For the purpose of characterization, malignant lesions generally have a lower ADC than benign lesions, and several thresholds have been proposed ⁸³. However, there is substantial overlap in ADC values between benign and malignant solid liver lesions ^40,83. This overlap, which likely arises from technical challenges as well as biological variability within tumors, currently limits the practical utility of ADC for lesion characterization. In order to overcome ADC variability, some studies use an ADC ratio (where ADC in the tissue of interest is normalized to a reference, such as the spleen) ^136,155-157, yet this is not utilized routinely in clinical practice. It must be noted that full characterization of lesions is made in conjunction with other pulse sequences and possibly tissue sampling, but diffusion MRI often serves as an integral component of this multiparametric evaluation. This approach is also utilized for non-liver abdominal applications. For instance, the pancreatic tail is a common place for accessory splenic tissue (a splenule), a benign asymptomatic finding present in a large portion of the population that is often found incidentally ^158-161. However, a splenule is often difficult to differentiate from a solid tumor at the tail of the pancreas, such as a neuroendocrine tumor ^162-166. Considering the high prevalence of splenules, tissue sampling of all observed lesions is difficult to justify due to the inherent risks of biopsy. Hence, MRI is often employed to evaluate pancreatic tail lesions: if the lesion demonstrates signal characteristics identical to the adjacent splenic parenchyma on every acquisition, then it is considered as a splenule; otherwise it will need further characterization, such as biopsy. Diffusion MRI can play an important role in this setting, as some lesions demonstrate similar T₁, T₂, and enhancement characteristics to the spleen, but have a non-spleen-like diffusion MRI appearance. These lesions are therefore considered as suspicious for tumor (Figure 11).

Figure 11.

Diffusion MRI assists in the characterization of pancreatic tail lesions. (Top row) A lesion in the tail of the pancreas does not appear to follow the signal intensity of the spleen on low and high b-value diffusion MR images, which is further confirmed by quantitative assessment on ADC. Hence, this was considered as suspicious for malignancy, and biopsy confirmed a neuroendocrine tumor. (Bottom row) Conversely, a lesion in the tail of the pancreas of a different patient appeared identical to spleen on the low and high b-value diffusion MR image and demonstrated an ADC value very similar to that of the spleen. Considering its spleen-like appearance on all other MRI acquisitions (not shown here), this finding was considered as probably a benign splenule and was further confirmed by ferumoxytol MRI.

One of the most well-defined roles for diffusion MRI in the pelvis is detection of prostate cancer. Per the current guidelines of the Prostate Imaging Reporting and Data System (PI-RADS v2.1) ¹⁶⁷, single-shot echo-planar diffusion MRI is the key pulse sequence for detection of clinically significant cancer in the prostate gland peripheral zone, where the majority of prostate cancers arise ¹⁶⁸. Furthermore, diffusion MRI plays a complementary and important role to T₂-weighted imaging for detection of cancerous lesions in the transition zone ¹⁶⁷. The role of quantitative diffusion MRI in prostate cancer has been investigated, where a meta-analysis reported a moderate correlation (r = −0.48) in the peripheral zone and a weak correlation (r = −0.22) in the transition zone between tumor ADC and histologic grade ⁶⁴. Despite this correlation, utilization of a cut-off value to differentiate between clinically significant cancerous lesions and other lesions has not been established due to the overlap between ADC values seen in both findings.

The role of diffusion MRI in evaluation of other abdominopelvic malignancies, such as rectal and gynecologic malignancies remains less well-defined. Diffusion MRI is often utilized by the radiologist as a problem-solving tool or to help point the radiologist to the area of abnormality, which is then correlated with other pulse sequences. The technical challenges associated with diffusion MRI, which have been discussed throughout this work, constitute a major barrier to more extensive and reliable utilization of this technique as a quantitative tool in the setting of pelvic malignancies.

The potential of qualitative and quantitative diffusion MRI methods has also been investigated in the setting of radiotherapy. Applications of diffusion MRI in the setting of radiotherapy include prediction of treatment response, delineation of the treatment area, and assessment of treatment response. For instance, studies have demonstrated a correlation between pretreatment ADC and/or changes in ADC, and chemoradiation therapy response in patients with rectal cancer ^169,170 and cervical cancer ^171-174. However, the significant overlap of the quantitative diffusion parameters between the responder and non-responder cohorts limits the prospective, individual-based utility of quantitative diffusion MRI as a decision-making tool. In a group of patients who underwent chemoradiation treatment for locally advanced cervical cancer, post-treatment tumor 50^th percentile ADC on a histogram analysis (but not the mean ADC) was higher in the responder group compared to the non-responder group¹⁷⁵. However, in these same patients, no pretreatment features of diffusion MRI were found to be predictive of treatment response¹⁷⁵. This further supports how the variability within the quantitative parameters of diffusion MRI hinders its widespread utilization. For these reasons, the incorporation of quantitative diffusion MRI into clinical practice for radiotherapy applications remains limited and would require further evaluation in well-designed prospective studies¹⁷⁶.

Diffusion MRI has also demonstrated a promising potential in the assessment of non-oncologic disease processes. One area of utility for diffusion MRI is the evaluation and surveillance of inflammatory bowel disease, an incurable disease that affects a young population. In this setting, MRI is the preferred imaging modality due to lack of ionizing radiation and the favorable safety profile of gadolinium-based contrast agents ^177,178. In addition, in the urgent care/emergency setting, such as for diagnosis of acute appendicitis in young children and during pregnancy, CT is not ideal due to the potential risks of ionizing radiation. Further, gadolinium-based contrast is often avoided during pregnancy, in order to minimize risk to the fetus. In this setting, diffusion MRI can provide additional and arguably significant tissue contrast by pointing the radiologist to the area of inflammation and enabling an accurate diagnosis ^179-182.

In addition to determining the presence or absence of pathology in abdominopelvic structures, quantitative diffusion MRI techniques have been employed to evaluate disease processes that exist over a spectrum between “normal” and “abnormal.” For instance, liver fibrosis, which is the sequalae of chronic liver disease of various etiologies (e.g., alcoholic and non-alcoholic fatty liver disease, iron overload, and viral hepatitis) is observed over a spectrum between mild fibrosis to end-stage fibrosis/cirrhosis. Hence, noninvasive quantitative evaluation of disease status is of high clinical significance and impact. Multiple studies have assessed the ability of quantitative diffusion MRI measurements (including ADC, IVIM, and specialized quantitative measurements) to quantify fibrosis, but only with mixed results ^{16,136,137,142,183-188}. A major and arguably understated challenge that essentially prohibits the utilization of ADC as a prospective marker for liver fibrosis is the significant overlap of ADC values between various stages of the disease. Furthermore, the wide disparity of ADC measurements across different studies ^16,18,189 (likely due to technical challenges such as motion artifacts, partial volume effects, modeling limitations, and variability in acquisition parameters) suggests that caution should be exercised when interpreting ADC measurements for this purpose.

Overall, despite the broad utilization of qualitative diffusion MRI, quantitative diffusion MRI is currently not widely used in the clinic. Even though an ADC map is typically produced and reviewed by the radiologist (e.g., in prostate imaging as described above), the ADC map is used largely for its ability to provide diffusion contrast while avoiding T₂ shine-through effects (where lesions appear darker than nearby healthy tissue in diffusion-weighted images due to T₂ contrast). However, the quantitative measurements from the ADC map are rarely used in a systematic manner for lesion detection or characterization in the clinic. Further technical development, validation and standardization are needed in order for quantitative diffusion MRI to fulfill its considerable potential and achieve broad dissemination in the clinical setting. This remaining work and research opportunities are summarized below.

Opportunities and Remaining work

Improved systems.

Continuous development of MR systems and hardware has enabled improved SNR in diffusion MRI, with shorter TEs and higher spatial resolution. However, these state-of-the-art advances have traditionally been focused on high field systems (e.g., 3T or higher). The development of modern MRI systems (e.g., with high performance gradients) operating at lower fields may have advantages for diffusion MRI applications, particularly in the abdomen. Despite the inherent baseline SNR loss, diffusion MRI of the abdomen at lower field strengths enables longer readouts with reduced T₂ and T₂* decay, reduced B₀-induced phase accumulation (i.e., reduced distortions), and reduced chemical shift artifact for fat signals. For these reasons, the current emergence of modern low-field systems (e.g., 0.55T or lower) offers intriguing possibilities ¹⁹⁰. Other hardware innovations, such as localized non-linear gradient systems ¹⁹¹, also offer fascinating opportunities.

Acquiring reliable signals.

The presence of motion artifacts, noise, distortions and limited spatial resolution, among other challenges, currently complicate the subsequent quantification of diffusion properties from acquired signals. Ongoing and emerging efforts to develop artifact-free, high-SNR diffusion MRI acquisitions are likely a necessary condition for the establishment of reliable and reproducible diffusion MRI methods in the abdomen and pelvis. These improvements in reliability may enable the application of advanced signal models with multiple free parameters.

Establishing new signal models.

In order to enable reproducible quantification of tissue properties, signal models that capture the relevant properties of the underlying tissues are highly desirable. In neuroimaging applications of diffusion MRI, emerging signal models are increasingly focused on explicit tissue modeling assumptions (rather than simply signal representations) ^192,193. Such developments may ultimately be successful in the abdomen and pelvis as well, particularly once reliable signals can be acquired as described above. Importantly, application of advanced models often requires acquisitions with very high b-values or diffusion times, which in turn requires additional technical development to enable reliable acquisitions.

Standardizing acquisition and processing.

Although advanced acquisition techniques and signal models are expected to enable improved quantitative reproducibility across a wider range of experimental conditions, these efforts will likely need to be matched by enhanced standardization and multi-center optimization of the acquisition and processing methods ^194,195. Recent developments in artificial intelligence may help address challenges with SNR, artifact correction, and data harmonization.

Clinical validation.

Promising diffusion MRI methods in various abdominal and pelvic applications require rigorous clinical validation, including their predictive ability and multi-center reproducibility in a clinical setting. In addition, there is a need to characterize biological variability by conducting adequate studies in patients ³. Further, diffusion MRI has proved useful in many studies in combination with other techniques (e.g., contrast enhanced). However, the ability of diffusion MRI to fully replace these techniques remains unclear.

Dissemination.

Widespread dissemination of quantitative diffusion MRI in clinical research as well as in the clinic will require demonstration of value. For example, questions of improved clinical outcomes and cost effectiveness will need to be addressed. These challenging questions highlight the importance of multi-disciplinary collaboration across multiple stakeholders, including MRI physicists, other basic scientists, radiologists and other clinicians (both academic and non-academic), regulatory agencies, and industry partners.