Investigating intracranial tumour growth patterns with multiparametric MRI incorporating Gd‐DTPA and USPIO‐enhanced imaging

Abstract High grade and metastatic brain tumours exhibit considerable spatial variations in proliferation, angiogenesis, invasion, necrosis and oedema. Vascular heterogeneity arising from vascular co‐option in regions of invasive growth (in which the blood–brain barrier remains intact) and neoangiogenesis is a major challenge faced in the assessment of brain tumours by conventional MRI. A multiparametric MRI approach, incorporating native measurements and both Gd‐DTPA (Magnevist) and ultrasmall superparamagnetic iron oxide (P904)‐enhanced imaging, was used in combination with histogram and unsupervised cluster analysis using a k‐means algorithm to examine the spatial distribution of vascular parameters, water diffusion characteristics and invasion in intracranially propagated rat RG2 gliomas and human MDA‐MB‐231 LM2–4 breast adenocarcinomas in mice. Both tumour models presented with higher ΔR 1 (the change in transverse relaxation rate R 1 induced by Gd‐DTPA), fractional blood volume (fBV) and apparent diffusion coefficient than uninvolved regions of the brain. MDA‐MB‐231 LM2–4 tumours were less densely cellular than RG2 tumours and exhibited substantial local invasion, associated with oedema, whereas invasion in RG2 tumours was minimal. These additional features were reflected in the more heterogeneous appearance of MDA‐MB‐231 LM2–4 tumours on T 2‐weighted images and maps of functional MRI parameters. Unsupervised cluster analysis separated subregions with distinct functional properties; areas with a low fBV and relatively impermeable blood vessels (low ΔR 1) were predominantly located at the tumour margins, regions of MDA‐MB‐231 LM2–4 tumours with relatively high levels of water diffusion and low vascular permeability and/or fBV corresponded to histologically identified regions of invasion and oedema, and areas of mismatch between vascular permeability and blood volume were identified. We demonstrate that dual contrast MRI and evaluation of tissue diffusion properties, coupled with cluster analysis, allows for the assessment of heterogeneity within invasive brain tumours and the designation of functionally diverse subregions that may provide more informative predictive biomarkers.


| INTRODUCTION
The effective treatment of brain tumours is an unmet and urgent clinical need. Five year survival for patients with grade IV glioma (glioblastoma, GBM) is only 5%. 1 While patients with low grade gliomas have a better prognosis, there is, as yet, no assured cure using conventional therapies. Patients with brain metastases, which affect approximately 8-10% of cancer patients and approximately 30% of patients with breast cancer, have a similarly poor prognosis. 2,3 High grade and metastatic brain tumours exhibit considerable spatial heterogeneity in gene expression and biochemistry, resulting in regional differences in tumour cell and microvascular proliferation, angiogenesis, necrosis and invasion. Diffuse infiltration of tumour cells in the neuropil, the dense network of interwoven neuronal and glial cell processes, is a characteristic of both low and high grade brain tumours, 4 and its clinical management is one of the greatest challenges facing neuro-oncologists and radiologists treating patients with brain tumours.
Imaging biomarkers for the assessment of tumour pathophysiology and response to therapeutics are now widespread in the clinic, and diagnostic imaging is an essential tool in the treatment stratification of patients with brain tumours. MRI enables the visualization of detailed anatomical features with high resolution due to its exquisite soft tissue image contrast, 5 and is therefore the imaging methodology of choice for defining brain tumour anatomy and delineating tumours.
Advanced MRI also provides a means of defining quantitative biomarkers to inform on biologically relevant structure-function relationships in tumours, enabling an understanding of the behaviour and heterogeneous distribution of such associations. 6 There is a pressing need for refined MRI strategies and quantitative biomarkers to accurately interrogate specific growth patterns associated with infiltrative intracerebral tumours.
Conventional gadolinium-enhanced MRI, which is used extensively for diagnosis and staging, and in early stage clinical trials in solid tumours, relies upon the hyperpermeable nature of neoangiogenic tumour blood vessels. Infiltrating tumour cells in the brain obtain essential nutrients by co-opting the existing vasculature, leaving the blood-brain barrier (BBB) intact, and consequently making appropriate delineation of infiltrative brain tumours problematic. Macromolecular ultrasmall superparamagnetic iron oxide (USPIO) particles, which remain within the vasculature over the duration of the MRI timeframe, have been used extensively pre-clinically to delineate blood vessels and to provide estimates of fractional tumour blood volume and vessel calibre. [7][8][9] USPIO-enhanced MRI, alone or in combination with gadolinium-enhanced MRI, may therefore enable delineation of vessels within glioma regions with an intact BBB within co-optive, infiltrative areas.
We hypothesized that functional MRI, incorporating assessment of vascular parameters and diffusion characteristics, can report on heterogeneity within infiltrative brain tumours and inform on functionally diverse habitats. We employed a multiparametric MRI approach incorporating native measurements and both Gd-DTPA-and USPIOenhanced imaging, and assessed the data using histogram and unsupervised cluster analysis, to examine the spatial distribution of vascular permeability and volume, diffusion characteristics and invasion in RG2 and MDA-MB-231 LM2-4 intracranial tumours in vivo. Magnetic field homogeneity was optimized by shimming over the entire brain using an automated shimming routine (FASTMAP). A morphological, 20-slice, fast, multi-slice RARE spin-echo sequence was first used for localization of the tumour and measurement of tumour volume. Next, diffusion-weighted (DW) images were acquired using an echo-planar DW (EPI-DWI) sequence

| MRI data analysis
Parameter estimation was undertaken using a Bayesian maximum a posteriori algorithm, which took into account the Rician distribution of noise in magnitude MR data in order to provide unbiased parameter estimates. 14,15 Estimates of the apparent diffusion coefficient (ADC, ×10 −6 mm 2 s −1 ) were determined from the EPI-DWI data.
The dual relaxation rate sensitivity of the IR-trueFISP sequence was utilized to provide estimates of both the native longitudinal and transverse relaxation times, T 1 and T 2 (ms). The change in R 1 (=1/T 1 , ΔR 1 , ms −1 ) following delivery of Gd-DTPA was also evaluated. The transverse relaxation rate R 2 * (s −1 ) was quantified using the MGE data, and the USPIO-induced change in R 2 * (ΔR 2 *) was used to estimate fractional blood volume (fBV, %), using the following equation 7,16 : where Δχ, the change in susceptibility induced by USPIO, was taken to equal 0.408 ppm, the Larmor frequency of protons (γ) taken to equal 4.26 × 10 7 s −1 T −1 and the static magnetic field strength B 0 = 7 T. The value of Δχ is valid for the dose of USPIO contrast agent used in vivo in mice. 7 Pixels corresponding to an fBV exceeding 17% (the limit value for the linearity between ΔR 2 * and fBV) 17 were excluded. All data were fitted on a pixel-by-pixel basis using in-house software (ImageView, developed in IDL, ITT Visual Information Systems, Boulder, CO, USA), and the median value of each parameter determined from an ROI that encompassed the whole tumour. Where possible, ROIs were also drawn over an uninvolved region of the brain in the contralateral hemisphere to the injection of cells, on an imaging slice containing tumour, and the same analyses performed as for tumour ROIs.

| Cluster analysis
Unsupervised cluster analysis was performed using in-house software developed in IDL. The k-means algorithm was employed to partition the ROIs into sub-regions of similar characteristics, defined in the two-dimensional (2D) feature space formed by the two parameters ΔR 1 and fBV, ΔR 1 and ADC, or fBV and ADC, and the three-dimensional (3D) feature space formed by all three parameters. 18 The optimal number of clusters for each dataset was determined using the cluster validation method described by Sugar and James. 19    Representative anatomical T 2 -weighted images and parametric maps of native T 1 and T 2 , ADC, ΔR 1 and fBV from each tumour model are shown in Figure 1A. The quantitative MRI data for tumour and uninvolved brain ROIs is summarized in Table 1.

| Statistical analysis
RG2 tumours revealed a more homogeneous appearance on T 2 - Administration of Gd-DTPA increases the relaxation rate R 1 in areas of brain tumours where the vasculature is permeable, for FIGURE 1 A, T 2 -weighted MRI images and parametric maps of native T 1 and T 2 relaxation times, ADC, the change in relaxation rate R 1 following intravenous administration of Gd-DTPA (ΔR 1 ) and fBV, from representative RG2 (upper panels) and MDA-MB-231 LM2-4 (lower panels) tumours propagated in the brain. B, Frequency histograms displaying the distribution of ADC, ΔR 1 and fBV in RG2 and MDA-MB-231 LM2-4 tumours, and uninvolved brain tissue (data from all evaluated ROIs). Sufficient data could not be acquired from matched uninvolved brain tissue for ΔR 1 analysis; therefore, values from brain tissue in tumour bearing mice where data could be assessed were combined (uninvolved) example in areas where the BBB is disrupted. The change in the tumour R 1 (ΔR 1 ) following Gd-DTPA administration was significantly greater in MDA-MB-231 LM2-4 tumours than in RG2 tumours (Table 1) Voxel distributions partitioned with the optimal number of clusters, cluster analysis colour maps of a representative tumour slice, and associated H&E images are shown in Figure 3. Figure S1b shows 2D projections of the clustered 3D voxel distribution of all three parameters in MDA-MB-231 LM2-4, providing a 2D view of the distribution for each parameter pair. In the maps of ΔR 1 and fBV, red voxels, indicative of areas with a low blood volume (low fBV) and relatively impermeable blood vessels (low ΔR 1 ), were predominantly located at the tumour margins. Yellow voxels originate from subregions with a low blood volume but high vascular permeability. In the MDA-MB-231 LM2-4 tumours, all voxels with high fBV clustered together (blue voxels), irrespective of their relative vascular permeability, and were associated with the viable tumour mass. This cluster, driven by relatively high fBV, is also evident in cluster maps combining fBV with ADC and when the contributions from all three functional parameters are combined, although this is coupled with relatively high ΔR 1 . In RG2 tumours, however, these high fBV voxels were divided between those with low (blue) and high vascular permeability (green), and were heterogeneously distributed throughout the tumours.  ADC coincides with low fBV and is clearly identifiable on the 3D analysis incorporating all three parameters. Red voxels with low vascular permeability but relatively restricted diffusion were associated with areas of invasion without oedema ( Figure 3C, open head arrow), as were voxels ascribed to the red cluster with low fBV and low ADC. Blue voxels with higher ΔR 1 and therefore higher vascular permeability, but with a range of ADC values, corresponded to the main dense cellular mass and largely correspond to the red cluster in the three parameter maps corresponding to relatively low fBV, high ΔR 1 and low ADC and the yellow cluster in maps incorporating fBV and ΔR 1 .

| DISCUSSION
High grade and metastatic brain tumours exhibit considerable spatial variations in proliferation, cellularity, angiogenesis, invasion, necrosis and oedema. Recent publications have highlighted the potential value of spatial analysis of brain tumours, combining multiple MRI parameters to segment tumour from healthy tissue and to establish regional variations within tumours. [22][23][24][25] Amongst the greatest challenges facing neuro-oncologists and radiologists treating patients with brain tumours is diffuse tumour cell infiltration and vascular cooption. In these regions the BBB remains intact, precluding detection by conventional Gd-DTPA-enhanced MRI. We therefore used multiparametric MRI, incorporating both Gd-DTPA and USPIO particle contrast agents, coupled with histogram and unsupervised cluster analyses, to evaluate the additional information that can be obtained regarding the underlying biology of intracranial tumours when the data acquired from sequences assessing tumour vascular permeability, blood volume and water diffusion are combined. 18 A key issue at the outset of this study was the application of our multiparametric MRI strategy in intracranial tumour models displaying been used to provide a model of metastatic brain disease. 29 Interesting differences were apparent in the vascular phenotypes of the two tumour models studied. MDA-MB-231 LM2-4 tumours demonstrated higher ΔR 1 than RG2 tumours, indicative of a higher degree of extravasation of contrast agent, hence more permeable vasculature. However, their fBV was lower than that of RG2 tumours, suggesting that, in the simplest terms, high vascular permeability does not necessarily correspond to high vascular density. Cluster analysis of fBV and ΔR 1 allowed for these regions to be mapped spatially; those with low vascular density, but high vascular permeability ( Figure 3B, yellow on ΔR 1 -fBV map, red on 3D cluster map), are likely indicative of regions of vasculogenesis, where sparse, immature permeable vessels are sprouting to provide a nutritive blood supply to the surrounding tumour cells. 30 As expected, these regions were typically found within the principal tumour mass. Areas with low vascular volume and low vascular permeability ( Figure 3B, red on ΔR 1 -fBV map) include those vessels that are not perfused, but may also correspond to areas with a low density of mature, relatively impermeable vessels, likely areas where the native vessels have been co-opted in regions of invasion. 31 Indeed, these voxels, particularly in the MDA-MB-231 LM2-4 tumours, were predominantly located at the tumour margin, where tumour cell invasion was identified on H&E-stained sections.
Regions with high fBV had either more densely packed or larger calibre vessels, resulting in a relatively larger blood volume. Regions where both parameters were high, which only formed a separate cluster in the RG2 tumours (green), probably correspond to regions with either extensive neovasculature or large distended vessels. 32 MRI protocols incorporating two vascular contrast agents, either an iron oxide preparation in combination with a gadolinium-based agent, or sequential injection of low and then high molecular weight gadolinium-based agents, performed either in the same imaging session or separated by hours, have previously been used to assess tumour vasculature in intracranial tumour models in rats and mice, [33][34][35][36][37] and for assessing response to anti-angiogenic treatment in melanoma xenografts in mice. 38 The key advantage of the imaging strategy used herein is that it enabled the spatial analysis of how  where data thresholding was applied, these values are removed from the maps. Voxels without an evaluable fBV (negative ΔR 2 *) are also missing from analyses incorporating fBV. C, H&E staining of the same tumours shows spatial relationships between cluster analysis maps and the tumour physiology. The closed head arrow denotes region of invasion along blood vessels and oedema, the open head arrow denotes invasion without oedema and the dashed arrow denotes main dense tumour mass gadolinium-based contrast agents in the same imaging session in adults with primary brain cancer or brain metastases, and in children with brain tumours. 41 Studies such as these, which incorporate methods similar to those used in this study, may guide the creation of new imaging criteria/biomarkers to evaluate brain tumour progression and pseudo-progression secondary to radio-chemotherapy and antiangiogenic agents.
Susceptibility contrast MRI exploits negative contrast induced by USPIO particles, which inherently reduce signal-to-noise ratios (SNRs).
The Bayesian maximum a posteriori model used herein provides a thorough treatment of data point estimates of R 2 * involving representation of the associated uncertainties. For susceptibility contrast MRI it allows the determination of the significance of differences in R 2 * between two or more measurements, and consequently the probability that ΔR 2 * is greater than or less than zero on a pixel-by-pixel basis in vivo can be estimated. 14 This method represents a more stringent calculation of MRI parameters when SNR is modest or low. Susceptibility artefacts as a result of the air-tissue interface can emerge in MGE images, but any data affected by such artefacts were excluded from analysis in this study.
We were interested to find that, whilst susceptibility contrast MRI-derived measurements of tumour fBV have previously been validated against uptake of the Hoechst 33342 perfusion marker 42  In addition to the assessment of tumour vasculature in the brain, the ability to assess cell density in brain malignancies using DW MRI is crucial, particularly where the permeability of the BBB is low. However, care must taken when interpreting ADC measurements, as the relationship between ADC and cell density can be complex, with rela- are absent from RG2 tumours. RG2 tumours were also more cellularly dense, consistent with the higher proportion of pixels with low ADC values. 44 Interestingly, despite high cellular density and no oedema or necrosis, RG2 tumours had median ADC values and data distributions higher than the uninvolved brain. Assessment of cluster maps of MDA-MB-231 LM2-4 tumours incorporating ADC data, alongside H&E staining of the same tumours, revealed clusters of particular note driven by high ADC, denoting relatively unrestricted diffusion. These voxels displayed low ΔR 1 and fBV (low vascular permeability and vascular volume) ( Figure 3B, yellow cluster on all 2D and 3D maps including ADC), and corresponded to histologically confirmed regions of invasive tumour growth associated with oedema. A key imaging hallmark of brain tumours is an area of fluid-attenuated inversion recovery or T 2 -weighted MRI hyperintensity outside the region of contrast enhancement, consisting of a combination of infiltrating tumour cells and oedema. 45 The routine incorporation of DW imaging to gadolinium-enhanced imaging protocols in the clinic may thus assist in the more accurate resolution and characterisation of these brain tumour regions. 46 Multiparametric MRI, focussing particularly on the microvasculature, was used by Coquery et al. to establish MRI-derived clusters to characterize tumour heterogeneity, and correlate them to pathophysiological features, in rat brain tumour models. 25 MRI-based brain tumour segmentation to separate different tumour subcompartments from normal brain structures has been largely performed using supervised learning techniques, which are time consuming and expensive due to the requirement of teaching datasets and labelled images. 47 Unsupervised methods are now increasingly being developed and have been shown to perform well in comparison with supervised techniques. 24 Of particular importance is the segmentation of non-enhancing tumour from healthy tissue to monitor tumour size more accurately over time. 48 Further refinements in these techniques include 3D histogram analysis of routinely acquired images to identify radiologically defined regional habitat variations in brain tumour data to provide deeper insight into the evolutionary dynamics of brain tumours, 22  of tissue, has the potential to provide predictive models to improve brain tumour diagnosis, prognosis and monitoring.

ACKNOWLEDGMENTS
We thank Allan Thornhill and his staff for animal maintenance and