Imaging Tumor Vascular Heterogeneity and Angiogenesis using Dynamic Contrast-Enhanced Magnetic Resonance Imaging

Development of quantitative biomarkers has led to increased use of imaging in the evaluation of tumor angiogenesis (1, 2). Magnetic resonance imaging (MRI) has been the most commonly used modality due to its wide availability and lack of ionizing radiation—two features that provide significant benefits over positron emission tomography. Also, although MRI does not offer the high level of biological specificity available with positron emission tomography ligands, it does provide a wide range of physiologically relevant image contrasts that make it a flexible imaging modality.

In dynamic contrast-enhanced MRI (DCE-MRI), the distribution of i.v. administered contrast agent is repeatedly imaged. Subsequent analyses produce biomarkers of the tumor vasculature. The technique is of interest to clinical oncologists for two main reasons. First, DCE-MRI biomarkers of drug efficacy and clinical outcome are used in an increasing number of early-phase clinical trials of antiangiogenic and vascular disrupting agents. Second, the technique has been applied in the nontrial setting to differentiate benign pathology from malignant disease, to grade tumors noninvasively, to detect early relapse, and to predict clinical outcome after both cytotoxic chemotherapy and radiation therapy (3).

DCE-MRI data analysis is complex. The variables obtained are heavily dependent on the choice of acquisition strategy (e.g., single slice, multislice, or three-dimensional imaging) and image analysis method [e.g., definition of the region of interest (ROI), selection of an arterial input function (AIF), and choice of pharmacokinetic model]. Crucially, all these factors can influence not only the value but also the precise meaning of derived variables. Furthermore, acquisition and analysis methods vary considerably between studies (4–6). These technical issues are considered in Appendix 1. In the remainder of this article, we begin by reviewing the application of DCE-MRI in clinical studies and in trials of angiogenesis inhibitors. We then discuss the application of DCE-MRI to investigate tumor heterogeneity because this illustrates an important potential use of functional imaging techniques in the evaluation of anticancer therapy.

Clinical adoption of DCE-MRI has been slow. Many studies have used simplified measurements of change in signal intensity for their analyses. Such metrics characterize the shape of the signal-time course curve or describe the rate of increase in signal intensity (7, 8), but show poor reproducibility and do not distinguish the effects of blood flow, blood volume, and contrast agent leakage (9). The application of T₁-weighted DCE-MRI to early-phase clinical trials of angiogenesis inhibitors has driven developments in quantitative image analysis, particularly the modeling of variables that describe tumor microvascular structure and function (Fig. 1). The details of these alternative analysis techniques are discussed in Appendix 1.

Analysis of enhancement kinetics (commonly from very low temporal resolution acquisitions) has become standard for the diagnostic classification of lesions in the breast and the liver. In the breast, it is common to calculate enhancement rates from regions of interest placed within suspicious lesions (10) with simple signal intensity metrics such as the proportion of maximal enhancement that occurs over a given time (e.g., 80%/min; ref. 8). Although malignant lesions tend to have higher enhancement rates, no fixed threshold provides a reliable distinction. Evidence of early washout of contrast agent has also been strongly associated with malignancy (11). Furthermore, enhancement patterns help identify histologic subtypes and predict response to radiotherapy and overall prognosis (12–14). Characterization of liver lesions is improved by examination of the enhancement rate and pattern (i.e., the spatial heterogeneity) and by examination of changes in enhancement on late scans (15).

In the brain, T₂*-weighted imaging of glioblastoma multiforme has repeatedly shown heterogeneity in cerebral blood volume (CBV) parametric maps (Fig. 2), where areas of high CBV correlate to regions of high-grade malignancy (16–18). Consequently, the practice of targeting biopsy to an area of high CBV has been adopted in many specialist centers (19). In addition, measurement of volume transfer coefficient of contrast agent across the capillary wall (K^trans) has been shown to independently predict prognosis using T₁-weighted sequences, further supporting use of DCE-MRI in the management of brain tumors (18, 20).

There is also increasing evidence that DCE-MRI may have further oncological applications. Characterization of prostate carcinoma lesions and their discrimination from benign disease can be improved by using dynamic contrast-enhanced analysis (21). Other potential applications include monitoring the response of osteosarcoma to chemotherapy (22), evaluation of the effect of radiotherapy early in treatment in cervical carcinoma (23), and using DCE-MRI to predict event-free survival in multiple myeloma (24). In general, as simple analysis packages, often using signal intensity metrics, become commercially available, indications for using dynamic contrast-enhanced computed tomography and MRI in follow-up and treatment evaluation are increasing (1, 25, 26).

Approximately 30 DCE-MRI–based studies of antiangiogenic (27–50) and vascular targeting agents (51–56) have been reported to date, predominately in phase I trials, although a few phase II studies have incorporated the technique. Image acquisition, analysis, and principal findings have been extensively reviewed elsewhere (57). Nearly all of these studies have utilized T₁-weighted DCE-MRI. Some studies have also incorporated positron emission tomography, T₂*-weighted DEC-MRI, and/or other functional imaging techniques.

Modeled DCE-MRI variables are subject to random error, biological variation and measurement inaccuracies. All of these factors cause day-to-day variation in measured values; thus, it is necessary to calculate confidence intervals for each combination of DCE-MRI model, choice of AIF, and variable. Some consensus has been reached on what degree of change is likely to represent true change—for example, most studies accept a change in K^trans of >40% as likely to represent a true difference due to drug effect—but in practice the precise confidence interval varies between studies depending on image acquisition and analysis variables (58, 59). For this reason, many centers have routinely done two baseline examinations to calculate variable reproducibility for an individual phase I trial (32, 44, 53).

Some trials have reported model-free analyses based on the initial area under the gadolinium contrast agent concentration-time curve (IAUGC), but the majority have used some form of pharmacokinetic analysis either alone or in combination with IAUGC (57). When modeling has been applied, most have adopted a simple model with only a minority attempting to correct for intravascular contrast effects. Despite this, evidence for drug efficacy has been found in several compounds, including the anti–vascular endothelial growth factor antibody bevacizumab (Avastin; Genentech; ref. 49); the tyrosine kinase inhibitors vatalanib (PTK787/ZK222584; Novartis; refs. 29, 36, 37, 47, 60), axitinib (AG-013736; Pfizer; ref. 34), BIBF1120 (Boehringer Ingelheim; ref. 39), and cediranib (AZD2171; AstraZeneca; ref. 30); and the antitubulin agent combretastatin A4-phosphate (CA-4-P; OxiGene; ref. 53). In each of these trials, significant reduction in IAUGC, K^trans, or K_i (the unidirectional transfer constant, a similar variable to K^trans) was observed after treatment. Finally, in some cases, the effective or biologically active dose was, in part, defined by the DCE-MRI biomarker (30, 46, 53).

Despite these encouraging results, identification of change in the tumor microvasculature does not necessarily predict success at phase III evaluation (61). Few trials of angiogenesis inhibitors have described a clear correlation in DCE-MRI biomarker with progression-free survival, the notable exception being sorafenib (BAY43-9006; Bayer; ref. 42). Details of image acquisition, analysis, and findings in some of the key trials are presented in Table 1, which highlights the considerable variation in choice of image acquisition and analysis.

No DCE-MRI variable has yet been accepted as a validated surrogate end point for use in phase III trials or routine clinical practice. Several points must be considered here. First, variables represent composites of physiologic processes rather than absolute measurements of flow or capillary permeability. Hence, a variable such as K^trans will not accurately measure flow or permeability—indeed, its precise meaning will vary between examinations. Second, there is evidence that the pharmacokinetic models used in DCE-MRI systematically overestimate K^trans and underestimate the blood plasma volume fraction (v_p; ref. 62). Third, considerable variations in measurement reproducibility reduce the confidence limits for individual measurements in a clinical setting.

Fourth, validation of DCE-MRI variables against histologic and serum biomarkers of angiogenesis is incomplete. Broad correlations between various MRI variables have been shown with tumor microvascular density (63); however, the marked variation in image acquisition and analysis between studies make comparison difficult. For example, several studies have correlated signal intensity metrics with microvascular density in patients with breast cancer. Amplitude and time to enhancement have been shown to correlate with microvascular density (64, 65); however, this has not been widely replicated (66). Furthermore, a linear relationship between the MRI efflux rate constant (k₂₁), which approximates to a measure of vascular permeability and surface area, has been reported with tissue vascular endothelial growth factor staining, although no relationship was found with microvascular density (10). This, in part, reflects the problem of comparing MRI changes measured at the millimeter level with either histology markers of vessel structure or distribution of receptors at the micrometer scale, or with circulating levels of angiogenic growth factors. Furthermore, microvascular density is a measurement of vessel density, whereas MRI variables such as v_p reflect the average density and are sensitive only to perfused vessels.

Finally, only a small number of DCE-MRI clinical studies have shown evidence for predicting clinical outcome. Examples include signal intensity metrics predicting local recurrence (67) and survival (68) in patients with cervical carcinoma and K^trans as an independent prognostic indicator of survival in patients with high-grade glioma (20). Lack of comprehensive validation against clinical outcome data currently prevents the use of DCE-MRI biomarkers as an accepted surrogate end point in a phase III trials or in clinical studies.

Human solid tumors are biologically heterogeneous (69), exhibiting considerable variation in microvascular structure and function (70, 71). In particular, growth factor expression is spatially heterogeneous within lesions of different size and may change during tumor growth (72). Furthermore, treatment of tumors with angiogenesis inhibitors has been shown to promote alternative angiogenic growth factor pathways, further promoting tumor heterogeneity (73). Imaging-based biomarkers have the potential to evaluate intratumoral heterogeneity and its relationship to tumor growth and response to therapy (2).

Technical considerations and practical approaches. The values of DCE-MRI variables are strongly affected by the definition of the tumor ROI. In each study, a decision must be made as to whether nonenhancing pixels are included in the ROI. Inclusion of zero values artificially lowers the mean and/or median variable values. Exclusion can obscure important heterogeneity information and can increase motion errors by effectively increasing the surface area of the normal tissue boundary around the sampled tumor. Three-dimensional ROI cover the entire tumor and are considered to be more accurate and reproducible than two-dimensional ROI, which are error prone despite attempts to correct for in-plane motion (36).

DCE-MRI data analysis assumes that signal intensity measurements taken over time originate from exactly the same volume of tissue in each sample. This is untrue outside of the brain due to physiologic (respiratory and cardiac) motion and patient movement. Use of a single ROI covering an entire tumor can reduce misregistration errors. Here, a single average enhancement curve can be extracted (34, 36, 37, 47, 51) to produce averaged kinetic variable estimates, for example of K^trans, for the entire tumor. However, this approach clearly ignores heterogeneity of tumor enhancement.

Alternatively, variables may be produced for each pixel (74), allowing the production of DCE-MRI variable maps, from which summary statistics, such as mean and/or median K^trans, may be calculated (28, 32, 33, 52, 53). This is routine practice in most cerebral studies because physiologic movement is negligible, allowing techniques such as tumor grading and targeted biopsy (19). However, the presence of motion reduces measurement reproducibility, especially in tumors with spatially heterogeneous enhancement around the edges of the selected ROI, resulting in a lower power to studies (75). Physiologic motion through the imaging plane further compounds motion error, particularly in studies that utilize a single imaging slice. Although these problems occur in whole-tumor ROI analysis, they are even greater in pixel mapping.

Hence, pixel mapping is technically challenging in areas of significant physiologic motion, such as the thorax and abdomen. Data may be acquired during breath holding (36, 56); however, such methods inevitably require concessions in the range and accuracy of biomarkers that can extracted (76). An alternative approach is to correct physiologic motion through postprocessing techniques (75). Spatial coregistration of MRI is routinely applied in dynamic brain imaging, but is more complex in the abdomen, where tissues are subject to greater motion and are not only displaced but also distorted during movement.

Specific DCE-MRI heterogeneity analysis techniques. Simple summary statistics (e.g., median K^trans) do not accurately describe heterogeneity within tumors (77). However, the distribution of variables can add further information. For example, the SD of pixel values for K^trans and similar variables within a ROI can improve diagnostic accuracy for distinguishing breast cancer from benign lesions, compared with summary variables alone (78). In a study of high-grade gliomas, median K^trans did not relate to tumor grade or behavior, whereas measurement of the upper part of the distribution of K^trans (95th percentile) showed close correlation with overall survival (20). In another study, mean values of relative contrast agent recirculation—a variable that reflects areas of under perfusion—had no significant relationship with tumor behavior, although the distribution of pixel relative recirculation values followed a normal distribution in low-grade tumors but was significantly skewed toward higher values in glioblastoma multiforme (Fig. 3; ref. 79).

These alternative analysis techniques suggest that description of tumor heterogeneity is possible and that such analyses can significantly strengthen the use of DCE-MRI in studies of tumor pathophysiology or drug action, for example in detecting differential regional changes in the tumor core and enhancing rim (53). More specialized statistical analyses of heterogeneity based on DCE-MRI data are in their infancy but have already shown important clinical findings.

Data from several studies suggest that histogram analysis is sensitive at detecting change in tumor heterogeneity after treatment. In patients receiving cytotoxic chemotherapy for locally advanced breast cancer, SD and kurtosis of enhancement amplitude before therapy predicted subsequent radiological response. Furthermore, tumors that responded had significant reduction in SD of enhancement amplitude and showed change from a heterogeneous distribution into a more homogeneous one (80). Change in histogram heterogeneity has also been seen in a second clinical study of breast cancer (81) and in patients with rectal cancer after radiotherapy (82). Relative signal intensity (a model-free measure of tumor enhancement) of the 10th percentile has been shown to better predict tumor recurrence than mean or median relative signal intensity in patients with cervical carcinoma after radiotherapy (83). Histograms of enhancement variables (including K^trans) have discriminated tumors treated with the thioredoxin-1 inhibitor PX-12 (Prolx Pharmaceuticals) versus untreated tumors in an animal model of colon cancer (84) and in a FSaII mouse fibrosarcoma models receiving NS-398 (85), where treated animals exhibited narrow homogeneously distributed histograms.

Interestingly, subtle or regional differences in antiangiogenic therapy efficacy can be shown using histogram analysis (Fig. 4), which may improve the overall sensitivity of preclinical studies and early-phase clinical trials. In one study, of the PC-3 prostate adenocarcinoma mouse tumor model, significant reductions in K^trans and the extracellular extravascular volume (v_e) were shown throughout entire tumors, after administration of the tyrosine kinase inhibitor ZD6474 (AstraZeneca). However, change in K^trans and v_e were significantly greater in the pixels corresponding to the lower 80% of absolute baseline K^trans values, which represented the tumor core (86). Importantly, application of a principal components analysis to the data showed disappearance of a second peak from a bimodal distribution after treatment—evidence of change in both the median value of K^trans and the heterogeneity within the tumor K^trans distribution (87). Tumor heterogeneity of response has also been shown in a mouse tumor model of human breast cancer after inhibition with AG-013736 (88). Histogram analysis of distributions of K^trans showed highly significant differences between progressive and stable disease in a phase I trial of the α_ν-integrin receptor antibody (CNTO 95; Centocor; Fig. 5; ref. 89), although no significant changes were found in median K^trans in response to therapy (40).

Histogram analysis may be useful in grading tumors. In a limited number of studies, principally in gliomas, clear differences have been shown in both the median values of CBV and the distributions of values of normalized CBV for high-grade gliomas and cerebral metastases versus low-grade tumors (90). Furthermore, CBV pixel histogram analysis can accurately grade gliomas (91) and has been shown to detect and quantify subtle changes in tissue enhancement that predict disease progression (92). It is important to note that although histogram analysis may provide useful information regarding tumor heterogeneity, it does so at the expense of including spatial information.

Useful heterogeneity information may be produced from other features of the dynamic data series. One simple but powerful method is to measure the fraction of pixels that enhance after contrast agent administration. A phase I trial of CNTO 95 has shown that enhancing fraction can discriminate between stable and progressive disease in the absence of a clear treatment effect (40). More sophisticated approaches to assess spatial heterogeneity include fractal analysis, which may provide useful markers of response to treatment (93). In another application of fractal analysis, v_e variable maps (a variable previously assumed to be of little physiologic interest) captured the heterogeneous nature of glioblastoma multiforme and predicted grade (94). Finally, factor analysis enables the characterization of the signal intensity profiles within a tumor and can identify intratumoral heterogeneity, as shown by facilitating prediction of response to chemotherapy in patients with osteosarcoma (22).

DCE-MRI has been shown to provide noninvasive imaging biomarkers of tumor biological response to antiangiogenic and vascular targeting agents in a number of clinical trials. It also has an increasing number of clinical applications. Although further animal and clinical validation of analytic techniques are required, DCE-MRI appears to be a useful tool in the investigation of tumor heterogeneity and has already shown its potential to improve sensitivity to subtle drug effects and provide additional understanding of tumor biology and response to therapy.

Image acquisition strategies. Principal contrast mechanisms in MRI emerge from the relaxation properties of tissue water (or fat) hydrogen nuclei in the scanner magnetic field after undergoing radiofrequency excitation. These properties can be manipulated to reflect features of the molecular and physiologic environment of a tissue. T₁-weighted images emphasize the rate (different for different tissue types) at which energy is redistributed between these nuclei and the tissue as a whole. T₂*-weighted images emphasize the different rates at which the same nuclei lose phase coherence with one another and the way that this process is affected by local variations in field strength due to differences in magnetic susceptibility between tissues. Conventional gadolinium chelate-based contrast media affect T₁-weighted images causing a signal increase, and T₂*-weighted images causing a signal decrease.

Nearly all clinical DCE-MRI is done at 1.5 T. The data have a typical spatial resolution of a few millimeters, which reflects the technical limits of the scanner specifications and the need for an acceptable signal-to-noise ratio. This, in turn, limits temporal resolution, in-plane resolution, and the number of slices that can be imaged. There is a direct trade-off between these factors so that rapid sampling at high spatial resolution can be achieved only by sacrificing the volume of tumor sampled (6, 95).

Dynamic susceptibility contrast MRI. T₂* dynamic imaging acquisitions have been widely used in the brain (96), where rapid loss of magnetic resonance signal is used to calculate the change in concentration of contrast agent for each individual voxel. Studies typically use fast imaging techniques such as single or multishot echo planar imaging to produce a temporal resolution of around 2 s, during which it is possible to acquire 5 to 15 slices at a matrix resolution of 128 × 128 or greater, depending on scanner specifications. The acquisition variables vary hugely in T₂ and T₂* dynamic imaging but all have a long time to repetition (e.g., 200-2,000 ms) and a long echo time (e.g., 50 ms; ref. 95).

Analysis of T₂*-weighted DCE-MRI is mainly relevant to a small number of glioma studies and the reader should refer to references (97, 98) for a more detailed description. In brief, the high initial concentration of contrast agent during the first pass makes T₂- and T₂*-weighted imaging sensitive to both flow and blood volume; hence, measures of the CBV and cerebral blood flow are commonly done (95). It is important to note that the contrast agent is assumed to remain in the vessels as a blood pool marker throughout the examination (99, 100). This is untrue in malignancy, where leakage of contrast agent occurs through a disrupted blood-brain barrier. Also, T₂* techniques generally suffer from susceptibility artifacts at interfaces between air and fat/soft tissue, further limiting their use outside the brain.

T₁-weighted DCE-MRI: data acquisition and analysis. Most DCE-MRI studies use T₁-weighted images, which are generally less sensitive to susceptibility artifact and allow estimation of capillary permeability and surface area. Investigators usually adopt a single/multislice or three-dimensional gradient echo sequence. Typical scan variables include a time to repetition of 3 to 10 ms, echo time under 5 ms, a variety of choice of flip angle, and in-plane imaging matrix between 128 × 128 and 256 × 256 at a spatial resolution of a few millimeters (32, 36, 44). Volumetric acquisition methods allow up to around 25 slices to be acquired over a period in excess of 5 min to provide adequate data. Pharmacokinetic analysis of contrast agent distribution requires transformation of measurements of signal change over time to measurements of contrast agent concentration (5). For T₁-weighted protocols, the relationship between signal change and contrast concentration is nonlinear; thus, a separate measurement of baseline T₁ values is done before the dynamic series (101, 102).

Pharmacokinetic analyses are complex and the choice of technique has major effect on the meaning of the derived variables. The IAUGC (103) is the simplest pharmacokinetic variable and appears reproducible. This is simply a measure of the amount of contrast agent delivered to and retained in the tumor tissue over a given time period and offers no biological specificity (2, 5). All other pharmacokinetic analyses use curve fitting to characterize the tissue contrast agent concentration curve, together with an estimate of the AIF, which needs to be measured in each patient (104) to portray the “true” input function for each tumor (5). However, accurate measurement of the AIF is technically demanding, due to blood flow artifacts in the image (101, 105), and requires high temporal resolution of less than a second (106). Such rapid data acquisition considerably limits the spatial resolution and extent of tissue coverage that can be achieved. An alternative approach adopts an idealized AIF—generally from population-based studies of contrast agent clearance from the blood pool—which relaxes requirements on temporal resolution, slice positioning, and sequence choice (107, 108).

Modeled variables are in theory independent of the scanning acquisition protocol, thus avoiding problems associated with signal intensity metrics, supporting their use in multicenter studies (26, 33). Simple analyses based on the Kety model provide estimates of K^trans (or a related variable, K_i) and v_e (109, 110). Here, changes in K^trans reflect changes in blood flow, blood volume, endothelial permeability and endothelial surface area and the individual effects of each component cannot be distinguished (5). These models are widely used but assume that signal change results entirely from contrast agent in the extravascular space. This produces significant errors in voxels that contain large vascular spaces—a relatively common occurrence in tumors (74, 111).

Attempts have been made to provide more accurate estimates of individual microvascular variables, particularly the signal contribution produced by contrast medium within blood plasma volume (112), and to model perfusion (F) and the product of capillary endothelial permeability and capillary surface area (113). These more complex models appear desirable because they allow specific analysis of biological effects; however, increasing the number of variables involved in the curve fitting process produces concomitant reductions in accuracy and reproducibility (9, 62). Therefore, many investigators calculate the IAUGC, which partially reflects K^trans, v_e, and v_p (114), and choose one or more of the modeled variables; choice of pharmacokinetic model depends on the degree of specificity required with regard to the mechanism of drug action, the acceptable level of reproducibility in the estimation of biomarkers, and the quality of acquired MRI data. Commonly used variables in DCE-MRI studies (4) are summarized in Table 2.

We thank Caleb Roberts for the original creation of Fig. 1.