Diffusion-Weighted Imaging in Cancer: Physical Foundations and Applications of Restriction Spectrum Imaging

Diffusion-weighted imaging (DWI) is increasingly used as an imaging biomarker for the detection and characterization of primary and metastatic brain tumors as well as for prognostication and monitoring treatment response in this patient population. The unique ability of this technique to probe the underlying structure of brain tissue at a cellular level makes it well poised to answer questions about tumor biology, as well as the microstructure of peritumoral white matter. The greatest enthusiasm to date has surrounded the ability of DWI to estimate tumor cellularity on the basis of quantitative images of diffusion, namely apparent diffusion coefficient (ADC) maps. In particular, numerous investigators have shown a negative correlation between tumor ADC values and cellularity (1–3), which is commonly attributed to increased restricted diffusion imposed by tumor cells. Following this logic, ADC has been used with varying success to characterize and grade primary and metastatic brain tumors (3–9), to assess tumor response to therapy (10–12), and to predict survival in patients with malignant tumors (13–17). However, confusion about the true biophysical basis of the DWI signal abounds, and awareness of the limitations of the existing approaches has led to the development of advanced DWI methods that extend beyond the tensor model. These methods include high-angular diffusion imaging techniques, such as diffusion spectrum imaging (DSI; ref. 18) and Q-ball imaging (19), as well as methods to probe non-Gaussian diffusion, including biexponential (20, 21), stretched exponential (22), and kurtosis imaging (23), and methods to study perfusion-related effects at low b-values, such as intravoxel incoherent motion (IVIM; ref. 24). Although these techniques often provide a better characterization of tissue architecture than traditional models, the relationship of these measures to the underlying pathophysiology of tumors is largely unknown.

Advances in magnetic resonance (MR) technology are now creating even more possibilities, increasing excitement in the field by broadening the potential applications of DWI within cancer diagnosis and treatment. Newer scanners with improved gradient performance allow for higher b-values with shorter diffusion and echo times. In addition, improved coil technology and higher field strength provide better signal-to-noise ratio (SNR). With these MR advancements, it has become possible to develop new, innovative DWI methods that provide more direct measures of tumor cellularity by leveraging the intrinsic contrast of tumor cells relative to other tissues. Such direct measures address a top challenge posed by the NCI to develop in vivo imaging methods that portray tumor “cytotypes,” i.e., imaging methods that can probe the identity, quantity, and location of different cells that make up a tumor and its microenvironment (NIH RFA-CA-13-020). However, the successful application of such methods requires not only advanced MR technology and an appreciation of the clinical challenges in neurooncology, but also an exquisite understanding of the physical basis of the DWI signal and its current limitations.

The purpose of this review is to provide a detailed explanation of the biophysical basis of diffusion contrast and to demonstrate what is known about how it reflects tissue microstructure in the context of key clinical dilemmas in neurooncology. We begin with a basic introduction of how diffusion measurements are derived from DWI and then follow with a review of the biophysical properties of water diffusion in tissue. We then describe how diffusion parameters in tissue are derived from the measurements via fitting of the diffusion model to observed data. For this review, we focus on one advanced DWI modeling technique called restriction spectrum imaging (RSI), which provides a general framework for estimating tissue properties from DWI data and addresses the NCI challenge for more direct, in vivo imaging of tumor cells. We also address the power of multispectral imaging, highlighting the importance of combining advanced DWI methods such as RSI with other imaging modalities (i.e., MRI perfusion, PET) to optimize the detection and monitoring of brain tumors. Although the primary focus of this review is on the application of RSI to neurooncology, the same methods are equally applicable to other applications of oncologic imaging, such as prostate, liver, and breast cancer, and we briefly present some preliminary data in prostate cancer at the end of “clinical applications.”

While early diffusion measurements using nuclear magnetic resonance (NMR) in biologic tissue were made in the 1960s and 1970s, it was not until the mid 1980s that the basic principles of DWI emerged (25–27). All MRI techniques, including DWI, rely on the fundamental principle that water hydrogen nuclei become magnetized when placed in a strong static magnetic field. MRI contrast is formed by perturbing this magnetization using electromagnetic waves and then allowing the magnetization to “relax” back to an equilibrium state. Tissue relaxation times—namely, the spin–lattice (T₁) and spin–spin (T₂) relaxation times—form the fundamental basis of soft tissue contrast and anatomic imaging with MRI. DWI is an MRI technique that adds additional sensitivity to the intrinsic random thermal displacements of water molecules that take place during the MRI experiment as part of the natural diffusion process. A pair of pulsed magnetic field gradients are turned on and off in succession (with duration δ) to magnetically “label” and “refocus” the spin phase of hydrogen nuclei, depending on where they reside physically in the gradient field (Fig. 1). The second refocusing pulse is applied some finite time Δ after the first pulse in an effort to realign the spin phases of stationary nuclei. Thus, any residual spin phase left after the application of the refocusing pulse can be attributed to the diffusion of water along the orientation of the field gradient during the effective diffusion time of the experiment T_d, defined here as Δ − (δ/3). The net phase dispersion due to diffusion causes an attenuation of the measured signal and a decrease in the voxel intensity. Stejkal and Tanner (28) were some of the pioneers of the pulsed field gradient experiment and provided the mathematic framework to relate the diffusion coefficient to the experimental variables. To date, the Stejkal Tanner pulse sequence remains the gold standard method for measuring diffusion in the clinic and forms the basis of quantitative mapping of tumor cellularity with ADC, as discussed below.

In this section, we provide a basic review of the three principal physical modes of diffusion in tissue: free, hindered, and restricted.

Free water diffusion describes the random (Brownian) motion of water molecules due to thermal agitation, in the absence of any obstacles. The displacement distribution of free water molecules is time dependent and Gaussian and obeys a statistical law established by Einstein in 1905. Along a single direction in space, the average molecular excursion of water molecules is proportional to the square root of diffusion time T_d. Mathematically, this can be expressed as s = (2DT_d)^1/2, where s is the root-mean-squared distance and D is the diffusion coefficient (Fig. 2A). For free water at brain temperatures (37°C), the diffusion coefficient is approximately 3 μm²/ms (29), which translates to a distance of approximately 17 μm in 50 milliseconds. In brain tissue, however, water molecules are constantly bouncing off and interacting with various tissue elements such as cell membranes and macromolecules. The net result is a decrease in diffusion mobility of water and a displacement distribution that no longer follows a single Gaussian distribution. Generally speaking, tissue elements impede water diffusion both through diffusion hindrance and restriction. Hindered and restricted diffusion are two distinct processes that result from fundamentally different behavior of spins within the intra- and extracellular tissue compartments, as described next.

Hindered diffusion refers to the delay of passage of small molecules as they navigate around cellular obstacles, as in brain extracellular space (ECS; Fig. 2B). There is a long history (30) of quantifying the degree of hindrance to diffusion in the ECS compared with that of a free medium by the tortuosity λ (31), defined simply as the square root of the ratio of the free (or intrinsic) extracellular diffusion coefficient D_extra to the measured or ADC for extracellular water ADC_extra: λ = (D_extra/ADC_extra)^1/2. As such, tortuosity simply reflects the degree to which the ECS slows diffusion relative to free water. Extensive experimental studies using real-time iontophoresis (RTI) using the cation tetramethylammonium (TMA) indicate that diffusion through tortuous ECS in normal brain is generally slowed by a factor of about 2.6, corresponding to λ of about 1.6 (31). Although there may be several factors that ultimately contribute to the delay of passage of water molecules in brain ECS (such as the composition of the extracellular matrix and the transient trapping or binding of water molecules to cell surfaces; see ref. 32), perhaps the greatest contribution is ECS geometry (dictated by the extracellular volume fraction α and shape of cells). As α decreases, due to, for example, cell swelling or greater cell packing density (cellularity), water molecules must travel more circuitous paths around cellular obstructions and tortuosity increases. The maximum theoretical tortuosity due to geometry can be quantified as the ratio of the distance though the center of a spherical cell to the distance around the periphery of the cell or λ_g = π/2 = 1.57 (Fig. 2, inset). However, more complex simulations over a wide variety of packed cellular objects indicate that the maximum tortuosity due to geometry is no greater than 1.22 (33). In tumor ECS (or “tumor interstitium”), both tortuosity and volume fraction α are generally higher. For example, in anaplastic astrocytomas and glioblastomas, the average α was measured to be as high as 0.47 to 0.49, with λ around 1.67 to 1.77 (34). Increased α has been associated with necrosis and or vasogeneic edema, whereas increased λ may result from either astrogliosis, commonly observed in tumor tissue, or to changes in the extracellular matrix (32, 34). In summary, while it is often suggested that the low ADC observed in high-grade tumors results from increased packing density (cellularity) of cells due to greater hindrance imposed on extracellular water diffusion, from a physical perspective, cellular crowding can only mildly reduce the effective diffusion coefficient in tumors (λ increases only narrowly with decreasing α). Therefore, it stands to reason that a major component in decreased ADC in tumors, especially at high b-values, is restricted diffusion within the cellular compartments themselves, described next.

Restricted diffusion is a term classically used (35) to describe the trapping of water molecules within an enclosed compartment (i.e., as defined by the cell plasma membrane) such that the net distance traveled is restricted or confined by the compartment dimensions. Although the term “restricted diffusion” is often used in the clinical literature to refer to any reported decrease in ADC, the physics of restricted diffusion is fundamentally different from hindered or free water diffusion. Specifically, for restricted diffusion, the net squared displacement of water molecules is sublinear in time and therefore non-Gaussian (Fig. 2C). In addition, the time evolution of net displacements strongly depends on the size and shape of the restricting compartment as well as the intrinsic intracellular diffusivity D_intra. In contradistinction to hindered extracellular water, where ADC_extra is independent with diffusion time T_d, the effective diffusion coefficient for intracellular water (or ADC_intra) decreases with T_d over a finite range as a larger proportion of the spins “bounce off” the plasma membrane.

It is important to note that restricted (or non-Gaussian) diffusion must be discussed within the context of diffusion time, permeability, and the size of the restricting cellular compartments. For instance, in the short-time limit (T_d < 1 ms), diffusion is largely unrestricted and dictated by the intrinsic diffusivity of the medium (ADC_intra ∼ D_intra), except for a small minority of spins located in close proximity to cell membranes or other barriers. Conversely, in the long-time limit (T_d > 1 s), much greater than the average residence time for water within intra- and extracellular compartments, diffusion is dominated by exchange and can be accounted for by a single apparent tensor, reflecting the effective medium approximation of the tissue (ADC_intra ∼ ADC_tissue). In the intermediate-time regimen, relevant to typical clinical DWI acquisitions (T_d ∼ 50–100 ms), the diffusion time is short relative to exchange between intra- and extracellular compartments through the plasma membrane (36), but long enough for water molecules to repeatedly come in contact with and “bounce off” the plasma membrane. In this case, the intracellular spins approach a “fill-up” regimen, where the displacements of spins are physically restricted by the plasma membrane and, therefore, dictated by the size and shape of the cellular compartment. Note that in the intermediate-time regimen, diffusion in the ECS remains hindered and behaves according to a classic Gaussian model, where the tortuosity is a function of packing density and extracellular volume fraction (31). Therefore, the total water signal in the intermediate-time regimen reflects a superpositioning of hindered and restricted water, leading to a nonmonoexponential signal decay at high b-value. As we see below, separating the restricted water fraction from hindered and free water fraction provides a more sensitive and specific biomarker for tumor cellularity compared with traditional ADC.

Diffusion models form the fundamental basis through which quantitative information about the underlying tissue microstructure can be gleaned from DWI signals. The typical millimetric scale at which DWI measurements are made (i.e., voxel size) is large compared with the micrometric scale of the underlying physical diffusion process, and, therefore, the goal of the diffusion model is to bridge the gap such that inference can be drawn on a scale much smaller than the voxel dimensions. One of the advantages of DWI over other MRI techniques is that the physical scale probed by the measurements can be adjusted by the experimental variables, namely the diffusion time (T_d) and diffusion-weighting factor (b-value). As we see, this forms the fundamental basis through which more advanced multiscale, or non-Gaussian, diffusion models offer insight into compartmental diffusion in cancer and improved imaging biomarkers for tumors. We begin this section with brief introduction of the classic ADC and tensor model for Gaussian diffusion and follow it with a more detailed introduction to RSI, a technique developed in our laboratory for probing tissue microstructure in greater detail.

As a departure from earlier studies in which efforts were made to measure the true diffusion process in biologic systems (37), it was suggested in the mid 1980s (24) to model the complex diffusion in tissue using the free (Gaussian) diffusion equation, but replace the intrinsic diffusion coefficient D with a global statistical parameter called ADC. The ADC concept has since been used extensively in the literature as well as clinical oncology as a surrogate marker of tumor cellularity (1–3). In practice, the ADC is estimated by combining the experimental variables (i.e., the magnitude, duration, and temporal spacing of the diffusion gradients) into a single parameter called the diffusion-weighting factor, or b-value (24), and comparing the signal attenuation at one or more nonzero b-values with the baseline signal measured without diffusion weighting (i.e., with a b-value of zero). For Gaussian diffusion, the signal attenuation decays exponentially with the product of the b-value and ADC, and, therefore, the ADC reflects the slope of the best fit line to the log signal as a function of b-value. The diffusion tensor model (DTI) extends the ADC concept to three-dimensional space, allowing for different ADCs along and perpendicular to the principal axis of diffusion (38).

The main limitation of the ADC and tensor model as imaging biomarkers for tumors is their strict dependence on a single Gaussian function for the displacement distribution of water molecules within cancer tissue and, therefore, a monoexponential dependence on the b-value. Although a single Gaussian assumption may be appropriate for data collected over a fairly narrow range of b-values (up to about 1,000 s/mm²), the inferences that can be drawn are limited by lack of specificity. For example, numerous investigators have shown a negative correlation between the tumor ADC values and cellularity (1–3), which is commonly attributed to increased restricted diffusion imposed by tumor cells. However, despite increased restricted diffusion, tumor ADC values rarely fall below that of normal appearing white matter (NAWM). This is true even in highly cellular tumors that originate in white matter, such as glioblastoma multiforme (GBM) and primary CNS lymphoma (39). One explanation for higher than expected ADC values in these tumors is the presence of vasogenic edema and focal necrosis within the tumor itself, which increases the ADC through reduced hindrance imposed on the extracellular water (40, 41). Thus, increased ADC due to edema and necrosis will offset reduced ADC imposed by tumor cells, resulting in lesions that are difficult to discern from NAWM on the basis of the ADC alone. This effect becomes more pronounced as the b-value is reduced because of increasing sensitivity to the fast, hindered water fraction.

Over the past decade or so, advances in DWI acquisition and gradient hardware have made it possible to probe a whole new regimen of water diffusion in cancer tissue beyond what was previously possible on clinical MRI scanners. Specifically, the application of strong magnetic field gradient pulses has allowed a much greater range of b-values, diffusion directions, and diffusion times to be acquired during a clinical acquisition. As a result, diffusion signals can be made specific to pools of water with very low effective diffusion coefficients, well below the tortuosity limit for ECS water, likely originating from restricted water trapped within the cells themselves (42). Numerous promising methods are emerging to capture and model complex non-Gaussian diffusion in tissue, including biexponential (20, 21), stretched exponential (22), and kurtosis models (23), as well as methods to study perfusion-related effects in DWI data at low b-values, such as IVIM (24). The application and description of many of these techniques for neurooncology applications can be found in an excellent review provided by Maier and colleagues (39). Moreover, techniques are emerging to probe diffusion spectra at ultra-short diffusion times on clinical systems using oscillating gradients (43), which hold promise for probing intracelluar structures and alterations in cancer (44). For this review, we focus on one particular technique developed in our laboratory called RSI. Over the last few years, RSI has gained increased recognition as an important tool in oncology that overcomes many of the limitation of traditional DWI and ADC.

RSI is a general framework for modeling diffusion signals collected across a broad range of experimental parameters and relating these signals to underlying tissue parameters (e.g., size and shape/orientation of hindered and restricted water compartments) using a linear mixture model (45). The ultimate goal of RSI is to enable quantitative estimates of tissue microstructure based on noninvasive imaging. To achieve this, the diffusion signal is modeled as reflecting a mixture of components, where each component describes the signal dependence on specific tissue properties (e.g., cell size, density, orientation, etc.) as a function of the experimental (protocol) settings (e.g., b-value, diffusion time, echo time, etc). The total signal becomes the weighted sum of these components, and the goal is to determine the individual weights. This is achieved through the application of generalized linear estimation techniques (45, 46). As such, the RSI framework is designed to strike a balance between model complexity and interpretability by minimizing a priori assumptions on microstructure while preserving biophysical interpretability of the resulting estimates.

Our current clinical implementation of RSI acquires data with b-values of 500, 1,500, and 4,000 s/mm² and multiple diffusion directions at each b-value at a fixed intermediate diffusion time (∼90 ms), where intracellular spins would be expected to be in the fill-up regimen. The RSI design matrix includes a distribution (or “spectrum”) of effective diffusion pools spanning hindered and restricted length scales with both isotropic and anisotropic geometries. In Fig. 3, we illustrate both a schematic for the RSI spectrum model and the resultant fit of the model to data collected in a 51 year-old patient with right frontal GBM before surgical intervention (47). Note the separation of diffusion components in different tissue types, with the primary lesion exhibiting a large volume fraction of spherically restricted water, likely stemming from water trapped within cancer cells. Also note the large fraction of free and hindered water in areas of necrosis and edema. The volume fraction of spherically restricted water (either with or without additional filtering of white matter signal using a technique called “beamforming”; ref. 48) has been coined the RSI cellularity index or cellularity map (47, 49–51). In addition, the signal from the restricted anisotropic component can be used to map the density and orientation of surrounding white matter tracts (50).

In summary, advanced DWI methods such as RSI that acquire data over an extended b-value range provide the ability to quantify complex non-Gaussian diffusion in tissue (21–23, 45, 52). Although these emerging techniques offer a new class of cancer imaging biomarkers, there remain many unanswered questions and tremendous opportunity for further advancing the field. Of particular importance is to understand exactly how parameters of the diffusion model relate to specific properties of cancer tissue, such as tumor cell size, density, and nuclear volume fraction. How specific are these biomarkers to tumor cells versus healthy or inflammatory cells? Can these new imaging biomarkers characterize tumor cytotypes and address a top challenge posed by the NCI to develop in vivo imaging methods that can probe the identity, quantity, and location of different cells that make up a tumor and its microenvironment? (NIH RFA-CA-13-020). Finally, to what extent can these new imaging biomarkers help solve current clinical dilemmas in oncology such as distinguishing tumor recurrence from radiation injury? Answering these questions will undoubtedly require both advances in diffusion modeling, simulation, and MRI hardware and software [such as the latest ultra-high performance (“connectome”) MRI scanners; ref. 53] together with improved quantitative histology, multimodal image registration, and validation procedures.

Because of its ability to isolate areas of truly restricted diffusion by separating and removing the hindered diffusion signal, RSI offers a more direct measure of tumor cells than other diffusion-weighted methods. Within the past year, we have shown that by identifying areas of restricted diffusion, RSI can be applied clinically to (i) improve tumor conspicuity in patients with high-grade primary and metastatic brain tumors (47), (ii) facilitate the interpretation of tumor response on imaging following antiangiogenic treatments that significantly reduce edema (49), and (iii) improve the ability to visualize white matter pathways coursing through regions of peritumoral edema relative to standard DTI (50). In the following sections, we describe each of these clinical applications in further detail. In addition, we describe a clinical scenario in which RSI lacks specificity and a multispectral imaging approach is warranted.

As described in the Introduction, ADC is frequently used as a marker of tumor cellularity in patients with high-grade tumors (1–3). Areas of tumor are associated with decreased ADC relative to surrounding tissue (40). However, concomitant edema and tumor-related necrosis increase ADC values, thereby directly opposing the reduction in ADC associated with tumor (54, 55). This offset presents a diagnostic challenge by diminishing the conspicuity of tumor on ADC maps. Because RSI isolates areas of spherically restricted diffusion, we tested whether it could provide increased conspicuity and delineation of tumor margins relative to standard and high b-value ADC (47). To accomplish this, RSI was performed in ten presurgical patients: four with GBM, three with primary CNS lymphoma, and three with metastatic brain tumors. Tumor conspicuity, edema conspicuity, and relative sensitivity to edema were quantified for RSI cellularity maps (RSI-CM), high b-value DWI (b = 4,000), and ADC, and these values were compared in manually drawn volumes of interest. Receiver operating characteristic (ROC) curves were used to evaluate the sensitivity and specificity of each method for delineating tumor from NAWM. In addition to visible differences in conspicuity (Fig. 4), ROC curves revealed greater sensitivity and specificity for delineating tumor from NAWM with RSI-CM (AUC = 0.91) compared with both high b-value DWI (AUC = 0.77) and ADC (AUC = 0.66). In addition, the relative sensitivity to edema was greater for high b-value DWI and ADC compared with RSI, reflecting RSI's ability to suppress the fast diffusion component associated with edema. Furthermore, greater heterogeneity of the diffusion signal within the tumor was observed on the RSI-CMs compared with DWI and ADC, as evidenced by a broader histogram distribution. This may represent the intrinsic heterogeneity of tumor cellularity both within and across tumor types. These data demonstrate one promising application of RSI, i.e., improved conspicuity and delineation of high-grade tumors compared with traditional DWI models and underscore the possibility that RSI may prove helpful in delineating tumor cytotypes and infiltrating disease in peritumoral edema.

Antiangiogenic therapies, such as bevacizumab, are increasingly used in the treatment of recurrent high-grade gliomas. However, these agents decrease permeability of the blood–brain barrier and, therefore, decrease contrast enhancement and edema in patients with high-grade gliomas in a manner that may not correlate with actual tumor response—a phenomenon known as pseudoresponse (56). Given this imaging challenge, we evaluated the ability of RSI to improve conspicuity within regions of the tumor compared with ADC in patients treated with bevacizumab and to further demonstrate that RSI is minimally affected by bevacizumab-induced reductions in edema (49). RSI-CMs and DWI were available for a series of patients with recurrent gliomas at baseline and following initiation of bevacizumab. Results showed that all patients exhibited sharp decreases in contrast enhancement and edema following treatment (i.e., pseudoresponse). Bevacizumab-induced decreases in edema had a greater effect on ADC than on the RSI-CMs, with the relative sensitivity to changes in edema being more than 20 times higher on ADC than on RSI-CMs (Fig. 5). These data provide additional evidence that RSI is less influenced by changes in edema compared with ADC, which may confer an advantage of RSI for interpreting true tumor response in the setting of antiangiogenic treatment. However, this study did not include clinical follow-up. Thus, whether RSI has greater predictive validity than ADC is of key importance and requires additional investigation.

There is increasing enthusiasm for the use of DTI and tractography in neurosurgical planning (57–64), and there is some evidence that tractography-guided neuronavigation can be used to minimize neurologic morbidity (60, 65, 66). However, the ability to resolve white matter structure in peritumoral regions that include edema has remained a challenge (39, 67). We applied the same logic as in the previous examples and tested the ability of RSI to provide better visualization and quantification of white matter tracts in regions that include edema (50). In this example, isolating the slow, restricted compartment yields a better estimate of “tubularity” (i.e., models the cylindrically restricted diffusion within axons; Fig. 3B, scale 1; ref. 45). This increases the sharpness of the estimates, allowing for better delineation of fiber tract orientation. In a series of ten patients with high-grade gliomas, we were able to demonstrate that RSI yielded higher fractional anisotropy (FA) estimates in regions of edema relative to standard DTI. Furthermore, at follow-up when the edema had resolved in most patients, FA estimates increased with DTI, but remained stable with RSI, indicating that FA estimates based on DTI were artificially suppressed by the edema. Tractography performed within regions of edema revealed superior ability of RSI to track fibers through areas of significant edema relative to standard DTI (Fig. 6). These data address yet another important and growing application of diffusion imaging within the field of neurooncology (i.e., surgical planning) and speak to the advantage of using advanced DWI models for revealing the anatomic structure of peritumoral white matter. An important caveat to mention is that it is well known that gliomas grow via an infiltrative pattern and that there is often nonenhancing infiltrating tumor surrounding the enhancing portion of a high-grade glioma (68). Further investigation is required to determine whether RSI or other advanced DWI methods will prove useful for detecting these areas of tumor infiltration.

Despite the many recent advancements in MR technology and DWI methods, no single modality has emerged as the “holy grail” of tumor imaging. Thus, most diagnostic challenges in neurooncology are best addressed using a multispectral imaging approach. One particular challenge occurs following treatment with concurrent bevacizumab and radiotherapy, with a subset of patients developing regions of marked and persistent restricted diffusion that do not seem to reflect an aggressive tumor (69–71). Although the etiology of these lesions remains uncertain, pathologic confirmation in several patients has revealed atypical gelatinous necrosis. We have coined this abnormality bevacizumab-related imaging abnormality (BRIA) and have observed that the BRIA signal on RSI is quantitatively similar to that seen in a tumor (51). This illustrates a clinical scenario in which RSI lacks specificity and a multispectral imaging approach is warranted. Thus, we explored whether using RSI in combination with perfusion imaging could help to differentiate BRIA from recurrent tumor. In a series of patients, we show that these techniques are complementary in that RSI is superior to rCBV for differentiating pathology from NAWM, whereas rCBV is superior for differentiating BRIA from tumor (Fig. 7). Thus, the combination of high RSI signal and low rCBV provides a distinct imaging signature of the BRIA phenomenon. Because of the increasing use of antiangiogenic agents, imaging methods that increase our understanding of both pseudoresponse and BRIA are of high importance. These data highlight an important example in which RSI lacks specificity in its current instantiation, and information from complementary imaging modalities is essential.

Although this review is primarily focused on neurooncology applications, quantitative imaging is equally relevant to other solid organ tumors. For example, multiparametric MRI has been explored in a variety of applications, including discrimination between indolent and aggressive disease in prostate cancer. Routine T₂-weighted imaging of the prostate is the most sensitive way to evaluate anatomic detail, but is limited for disease detection, with sensitivity of around 70% and specificity of 55% (72). Functional MR techniques enhance detection, grading, and staging of prostate cancer through the use of dynamic contrast enhancement (DCE), DWI, and MR spectroscopic imaging. DCE requires intravenous administration of a T₁-shortening agent with the chief limitation being that the maximum contrast enhancement between malignant and nonaggressive disease is small (73). MR spectroscopic imaging increases specificity (74, 75), but is technically challenging and can add significantly to scan time.

Multiple studies have shown that DWI improves sensitivity and specificity in the diagnosis of prostate cancer by increasing tumor conspicuity on DWI or quantitative ADC maps. However, hemorrhage, inflammatory processes, and benign nodules in the transitional zone can all exhibit lower ADC values, leading to false positives (76). DWI can also suffer from severe spatial distortion, limiting its coregistration to anatomic images, which is necessary for tumor localization.

Increasing Gleason score correlates with loss of normal gland formation, loss of peripheral gland tubular structure, and increased cellularity (77). We hypothesize, therefore, that RSI cellularity will correlate with higher tumor grade, as measured by Gleason score, and will provide significantly greater accuracy in discriminating aggressive tumors from benign and indolent lesions when compared with current functional or anatomic imaging techniques.

Our preliminary data are encouraging. Figure 8 shows a patient with Gleason 3+4 = 7 disease. RSI cellularity map is shown color-coded and fused with the T₂ after spatial distortion correction and registration. Note the correspondence of RSI cellularity with the histopathology slice. T₂ and perfusion images are not as conspicuous, with much less contrast-to-noise ratio. This represents an example case in which conventional imaging is not as diagnostic.

One of the main limitations of diffusion imaging in general is nonlinear spatial distortion of the images due to a number of factors, including gradient nonlinearities, eddy currents, and B₀ field inhomogeneities. Although distortions due to nonlinearities of the gradient fields are commonly corrected for by software on the scanner console, B₀ field inhomogeneities are not, despite being the dominant source of spatial inaccuracy in DWI. The magnitude of B₀ distortions varies depending on a number of factors, including field strength, positioning of the subject within the scanner, and subject-specific anatomy. The typical pattern of B₀ distortion on 3T systems is illustrated in Fig. 9, along with the total whole-brain histogram of root mean square (RMS) displacements. As illustrated in the figure, on average (based scans from on 40 subjects) the typical (mode) distortion magnitude is approximately 2 mm, with a substantial proportion of voxels displaced more than 6 mm. Such distortions are of particular concern when images are used for image-guided intervention, including surgery, biopsy, or radiation dose planning. Although methods for correcting B₀ distortions have existed for some time (78–80), they are typically used only in research studies and not in clinical practice, primarily due to the additional scan time required to acquire the B₀ field maps required for standard correction methods. The method introduced by Holland and colleagues (81) overcomes this limitation by requiring only a single additional TR (2–3 seconds), using the reverse phase–encode polarity method (79, 82, 83). Moreover, this additional volume acquisition can be integrated directly in the native DWI protocol without changing the pulse sequence. Combined with efficient postprocessing methods to estimate the distortion field (81), the reverse phase–encode polarity technique provides a much needed clinical solution for accurate spatial distortion correction of DWI data. The RSI technique incorporates this procedure as part of standard preprocessing of data, and all the data presented in this review has been corrected in this manner.

The field of MR imaging is rapidly evolving, leading to new and exciting possibilities within neurooncology, urologic oncology, and beyond. In particular, the development of advanced DWI methods has allowed for improved visualization and detection of tumor cells and, thus, has great potential for better understanding of tumor biology. Numerous clinical applications of this powerful technique have already been demonstrated, including tumor characterization and grading, prognostication, early prediction of response to therapy and survival, distinguishing tumor from treatment-related confounds, detecting microinfiltration, and guiding neurosurgical and radiation planning (84). However, future advances in the field will require a fundamental understanding of the underlying DWI signal coupled with validation of diffusion contrast in unique tumor-related pathologies. In addition to improved detection of tumors, advanced diffusion methods such as RSI may also provide quantitative characterization of cellular properties such as cell size, permeability, and nuclear volume fraction, based on signal variation as a function of diffusion time and echo time (42). Validation of these measures will likely be borne out of translational efforts that include both preclinical and clinical studies in which histologic specimens are carefully coregistered to in vivo imaging. The need for targeted biopsies based on advanced DWI, precise coregistration of DWI with other imaging modalities, and careful corrections for geometric distortions will all be pivotal to providing the spatial precision needed to achieve such validation. These requirements underscore the need for a multidisciplinary approach to this validation including experts in the fields of oncology, surgery, pathology, and radiology. Through this collaboration, not only will current applications of DWI be further improved, but new possibilities will also be created that will ultimately lead to better care for patients suffering from cancer.

D. Holland has ownership interest in patent #8,160,319 (reducing distortion in magnetic resonance images). A. Bjørnerud is a board member of NordicNeuroLab AS. D. Margolis received a commercial research grant from Siemens Medical Systems. A.M. Dale received a commercial research grant from General Electric Healthcare and has ownership interest (including patents) in CorTech Labs, Inc. No potential conflicts of interest were disclosed by the other authors.

The authors thank the patients at University of California, San Diego, Neuro-Oncology Program for their generous participation.

This study was supported by grants R01AG031224, R01EB000790, and RC2 DA29475 (A.M. Dale).