Skip to main content
NCBI home page
Journal of Cerebral Blood Flow & Metabolism logoLink to Journal of Cerebral Blood Flow & Metabolism
. 2014 May 21;34(8):1354–1362. doi: 10.1038/jcbfm.2014.90

Microvascular MRI and unsupervised clustering yields histology-resembling images in two rat models of glioma

Nicolas Coquery 1,2, Olivier Francois 2,3, Benjamin Lemasson 1,2, Clément Debacker 1,2,4, Régine Farion 1,2, Chantal Rémy 1,2, Emmanuel Luc Barbier 1,2,*
PMCID: PMC4126096  PMID: 24849664

Abstract

Imaging heterogeneous cancer lesions is a real challenge. For diagnosis, histology often remains the reference, but it is widely acknowledged that biopsies are not reliable. There is thus a strong interest in establishing a link between clinical in vivo imaging and the biologic properties of tissues. In this study, we propose to construct histology-resembling images based on tissue microvascularization, a magnetic resonance imaging (MRI) accessible source of contrast. To integrate the large amount of information collected with microvascular MRI, we combined a manual delineation of a spatial region of interest with an unsupervised, model-based cluster analysis (Mclust). This approach was applied to two rat models of glioma (C6 and F98). Six MRI parameters were mapped: apparent diffusion coefficient, vessel wall permeability, cerebral blood volume fraction, cerebral blood flow, tissular oxygen saturation, and cerebral metabolic rate of oxygen. Five clusters, defined by their MRI features, were found to correspond to specific histologic features, and revealed intratumoral spatial structures. These results suggest that the presence of a cluster within a tumor can be used to assess the presence of a tissue type. In addition, the cluster composition, i.e., a signature of the intratumoral structure, could be used to characterize tumor models as histology does.

Keywords: brain, glioma, MRI perfusion, pattern recognition, radiomics, unsupervised clustering

Introduction

To better understand a complex disease or to fully evaluate the impact of a drug, there is a strong interest in establishing a link between clinical in vivo imaging and the biologic properties of tissues. Indeed, tissue samples are not always available, and, when available, the small tissue sample may not represent the entire lesion. Major developments have been undertaken in the field of brain cancer using magnetic resonance spectroscopic images, an approach that yields several metabolic estimates for each location in the tumor.1, 2, 3 However, the spatial resolution of magnetic resonance spectroscopic images is limited and not appropriate to characterize lesions below ∼1 cm.

Another major source of contrast for in vivo imaging is tissue microvascularization, a system to which magnetic resonance imaging (MRI) is particularly sensitive. Indeed, MRI can map numerous structural and functional phenomena such as water diffusion, blood volume and blood flow,4 microvessel size and density,5 local blood oxygen saturation,6 vessel wall permeability,7 and vessel vasoreactivity.8 Major efforts have been and are made to obtain accurate and quantitative maps of each of these physiologic characteristics. Moreover, microvascular MRI may be obtained with much higher spatial resolution than magnetic resonance spectroscopic images, routinely 0.2 cm.

To derive histology-resembling images from multiparametric MRI maps,9 several techniques have been used including classification algorithms (i.e., supervised algorithms, such as linear discriminant analysis (LDA)10 or support vector machines11) and clustering algorithms (i.e., unsupervised algorithms, such as k-means or hierarchical clustering12). Supervised algorithms require a training data set, but histologic training samples are not always available (e.g., biopsy in the brain after a stroke is not feasible). Unsupervised algorithms thus appear more attractive. In medical imaging, the k-means and other clustering approaches have been evaluated.10, 13 However, there is little systematic guidance associated with k-means methods to address issues such as the number of clusters that best explain the data, which clustering method should be used, and how outliers should be handled. In addition, k-means methods can hardly assess the uncertainty on voxel assignments to each detected cluster, precluding the possibility of probabilistic prediction for new data. To overcome these limits, Fraley and Raftery14 developed an unsupervised clustering algorithm to perform cluster analysis based on probabilistic mixture models. The underlying probabilistic models can then be used to evaluate the uncertainty of voxel assignments and decide which models yield the best predictions. While there is a growing interest for such strategies in several domains, model-based clustering approaches have not been evaluated in the context of multiparametric MRI.

In this study, our aim was to evaluate whether clusters with defined histologic properties can be obtained based on an unsupervised analysis of microvascular MRI data. Microvascular MRI data were obtained from two models of glioma, C6 in Wistar rats (C6/Wistar) and F98 in Fischer rats (F98/Fischer). A combination of a manual selection of the region of interest and of an unsupervised clustering analysis is proposed to focus on tumor characterization. The identified clusters, defined by their MRI features, were found to correspond to histologic features. At a larger scale, the cluster composition of a given ROI, i.e., a signature of the intratumoral structure, could be used to characterize tumor models as histology does.

Materials and Methods

The study design was approved by the local committee for animal care and use (C2EA-04: ‘Comité d'éthique en expérimentation animale GIN'). Experiments were performed under permits (n°380820 for CR and A3851610008 for experimental and animal care facilities) from the French Ministry of Agriculture.

Animal Model

The C6 cells (105 cells, ATCC: CCL-107) and F98 cells (103 cells, ATCC: CRL-2397) were respectively implanted in the right caudate nucleus (coordinates from bregma: AP=0, ML=3.5, DV=5.5 mm) of male Wistar (n=13, 7 weeks, 175 to 200 g, Janvier, France) and male Fischer 344 rats (n=13, 7 weeks, 175 to 200 g, Charles River, France) with a stereotactic frame, as previously described.15 Briefly, 5 μL of tumor cell suspension in serum-free alpha-minimum essential medium were inoculated under anesthesia: 5% isoflurane for induction and 2.5% for maintenance in air. Bupivacaine (8 mg/kg; Centravet) was subcutaneously injected before incision to prevent postoperative pain.

In Vivo Experiment

Magnetic resonance imaging was performed at 4.7 T (Avance III console; Bruker, Germany; IRMaGe MRI Facility) between 21 and 24 days after tumor implantation. Animals were anesthetized (5% isoflurane for induction), intubated and subsequently mechanically ventilated with 2% isoflurane in 80% air-20% oxygen using a rodent ventilator (SAR-830/P; CWE, Ardmore, PA, USA). The femoral vein and artery were equipped with a catheter to collect blood samples. Partial pressures of O2 and CO2 of blood samples were determined with a blood gas analyzer (ABL 510, Radiometer, Copenhagen, Denmark). Individual values for each animal were calculated as the mean of pre and post MRI values. The tail vein was equipped with a catheter to deliver MRI contrast agents. Before the MRI session, animals received an intravenous injection of pimonidazole (60 mg/kg; Chemicon International; Temecula, CA, USA) diluted in saline (60 mg/mL). Pimonidazole is a 2-nitroimidazole reductive chemical probe that undergoes reduction below a partial pressure of O2 of ∼10 mm Hg and can then be immunohistochemically detected to assess tumor hypoxia.16 During the whole MRI session described below, the rectal temperature was monitored and maintained at 37.0°C. Two hours after pimonidazole, animals were euthanized under anesthesia with an intracardiac injection of KCl 2 mol/L. Brains were quickly removed, frozen in −40°C isopentane, and stored at −80°C until processing.

Magnetic Resonance Imaging Session

After careful shimming, the following imaging sequences were performed. Anatomic imaging was performed with a T2-weighted (T2W) spin-echo sequence (repetition time/echo time (TR/TE)=4,000/33 ms, voxel size=117 × 117 × 1,000 μm) to determine tumor volume. The apparent diffusion coefficient (ADC) was mapped using a spin-echo echoplanar imaging sequence (TR/TE=3,000/28.6 ms, three orthogonal diffusion directions, b≈0 and b=900 seconds/mm2, voxel size=117 × 117 × 1,000 μm). Cerebral blood flow (CBF) was determined using continuous arterial spin labeling (spin-echo echoplanar imaging, TR/TE=4,500/17.2 ms, labeling duration=4 seconds, post-labeling delay=0.2 seconds, 20 label/control pairs, voxel size=117 × 117 × 1,000 μm). T1 of brain tissue was determined using an inversion recovery sequence (spin-echo echoplanar imaging, TR/TE=1,000/20 ms, 25 inversion times: TI=100–3,700 ms, voxel size=117 × 117 × 1,000 μm). Tissular oxygen saturation (StO2) maps were estimated using a multiparametric qBOLD approach as previously described.17, 18, 19 A T2 map was obtained using a multiple spin-echo sequence (28 echoes, TR/TE=2,300/12–336 ms, voxel size=234 × 234 × 1,000 μm). A high resolution T2* map was obtained from a three-dimensional multiple gradient echo sequence (25 echoes, TR/TE=100/3–87 ms; voxel size=117 × 117 × 200 μm). A cerebral blood volume (CBV) map was obtained using a steady-state approach.20 A multiple gradient echo sequence (26 echoes, TR/TE=6,000/3.5–78.5 ms, voxel size=234 × 234 × 1,000 μm) was performed before and 20 seconds after an intravenous injection of ultrasmall superparamagnetic iron oxide nanoparticles (P904: 200 μmoles of iron/kg body weight; Guerbet SA, France) flushed with 250 μL of saline. The vascular wall integrity was assessed using a dynamic contrast-enhanced MRI approach as previously described.21 Briefly, 60 T1-weighted, spin-echo, images (TR/TE=800/4.2 ms, voxel size=117 × 117 × 1,000 μm; 15.6 seconds per image) were acquired. After acquisition of 10 baseline images, a bolus of Gd-DOTA (200 μmol/kg; Guerbet SA, France) was administered through the tail vein and flushed with 250 μL of saline. The image acquisition positions were identical for all MRI sequences.

Data Processing

Image processing was performed within the Matlab 7 environment (The MathWorks, Natick, MA, USA) using custom software. The CBF computation was based on the equations described in Alsop et al and Buxton et al22, 23 and assuming a transit delay much shorter than the longitudinal relaxation time of blood. The CBV was derived from the change in R2* estimated before and after injection of P904.20 StO2 was computed using the T2 map, the T2* map, and the CBV map, as previously described.17, 18, 19 The cerebral metabolic rate of oxygen (CMRO2) maps were computed using CMRO2=CBF × (1-StO2/100) and are expressed in arbitrary units. The vascular wall integrity was estimated by the area under the signal enhancement curve (AUCGd-DOTA) after the Gd-DOTA injection, after baseline removal.21 To allow voxel-wise analyses, the spatial resolution of all MRI maps was set to the lowest spatial resolution of the acquired images (234 × 234 × 1,000 μm).

Magnetic Resonance Imaging Data Analysis

For each animal, analyses were carried out on the MRI slice with the largest tumor area. Voxels were included in the analysis according to the following criteria: 0<ADC<2,500 μm2/second, AUCGd-DOTA<5 106a.u., 0<CBF<400 mL/100 g/minute, 0<CBV<20%, 0<StO2<100%. Beyond these ranges, values cannot be considered physiologically accurate. Meeting a single exclusion criterion led to an exclusion of the voxel from all maps.

Three types of analyses were performed:

  • A whole-brain clustering. The brain was manually contoured on the T2w-image. All included voxels of all animals were pooled together and analyzed with a model-based clustering approach.14 First, voxel values were standardized per parameter to minimize scale effects between parameters: standardized value=(voxel value−mean across all voxels)/s.d. across all voxels. Gaussian mixture modeling was then performed using the Expectation–Maximization algorithm implemented in the R package mclust.14 Model choice and the optimal number of clusters were determined from the values for which the Bayesian information criterion reached a plateau14 (Supplementary Figure S1).

  • A ROI selection followed by an analysis of the mean ROI values. First, three regions of interest (ROIs: tumor, contralateral striatum, and contralateral cortex) were manually delineated on the T2w-image using hematoxylin–erythrosine slices as control. Based on our experience, this approach delineates the tumor core and thereby limits partial volume effect. Average values were computed across all animals for each ROI and each parameter.

  • A ROI selection followed by a model-based clustering analysis. The included voxels of all animals and located in the three previously defined ROIs were pooled together for the clustering analysis described above with the optimal number of clusters (Supplementary Figure S2). A comparison study showed that Gaussian mixture models led to better segmentation of MRI images than hierarchical clustering or k-means methods (Supplementary Figure S3).

Brain Sectioning and Histology

For immunohistochemistry, coronal cryosections (20-μm thick) were cut along the entire striatum on a microtome (Leica, Nanterre, France). The closest sections to the MRI slice were subsequently examined. Histologic analyses were performed on three adjacent slices to allow comparison between labeling types. Hematoxylin–erythrosine staining was performed to characterize tumor cell density. Tumor hypoxia was determined using the Hypoxyprobe-1 kit (Chemicon International). Mouse anti-pimonidazole antibody was diluted 1:10; the secondary biotin-conjugated rabbit antibody was diluted 1:200 (B-8520, Sigma, St Quentin Fallavier, France). Extravidin-conjugated peroxidase was added 1:1,000 (E-2886, Sigma) and detected with the DAB Peroxidase Substrate Kit (SK-4100, Vector Laboratories, Peterborough, UK). Hematoxylin–erythrosine and pimonidazole sections were examined under a light microscope and digitized with a computerized image analyzer (Analysis; Soft Imaging System, Munster, Germany). Blood vessels were revealed with collagen IV immunofluorescence. Primary antibody was polyclonal goat anti-collagen IV diluted 1:1,000 (AO305, Southern Biotech, Birmingham, AL, USA) and secondary antibody was donkey Alexia-546 anti-goat antibody diluted 1:500 (A11056, Invitrogen, St Aubin, France). For fluorescence acquisition and large field brain section reconstruction, tissue sections were scanned with a Leica DMI6000 microscope (Leica), a × 10 dry objective and a EMCCD Quantem camera (Roper Scientific, Trenton, NJ, USA) driven by the scan slide module of Metamorph software (Universal Imaging, New York, NY, USA).

Tissue Type Prediction with Linear Discriminant Analysis

In each of the 78 ROIs (three ROIs per animal: cortex, striatum, tumor; 26 animals), the contribution of each of the five clusters was computed. Thereby, each ROI was characterized by five values, representing the number of pixels belonging to each of five clusters. In addition, six classes were determined for all ROIs: C6 tumor, Wistar cortex, Wistar striatum, F98 tumor, Fischer cortex, Fischer striatum, with each ROI assigned to a single class. To evaluate the ability of ROI classes to be predicted by the clustering outputs, a LDA was conducted using a ‘leave-one-out' cross-validation algorithm. In this algorithm, an ROI was removed from the list of all 78 ROI and LDA was performed to discriminate among the six classes based on ROI cluster composition. Using the results of LDA, a class was then predicted for each removed ROI (this step was repeated for each ROI). The fraction of correct predictions was eventually computed for each of the six classes. To further evaluate the contribution of each cluster and each MR parameter to the assignment prediction, the leave-one-out analysis was repeated but either one cluster or one MR parameter was removed for the analysis.

Statistical Analysis

Unpaired t-tests were used to evaluate differences between tumor models in MRI and blood gas parameters. Paired t-tests were used to evaluate differences between ROIs within tumor models. A P-value<0.05 was considered significant.

Results

Physiologic Controls

The tumor volumes were similar between tumor models (C6/Wistar: 85.1±46.4 mm3 and F98/Fischer: 104.9±28.0 mm3). Arterial blood gases were also similar between tumor models (C6/Wistar: PaO2=123.2±20.0 mm Hg, PaCO2=35.3±9.7 mm Hg, PvCO2=38.6±8.2 mm Hg; F98/Fischer: PaO2=126.4±22.2 mm Hg, PaCO2=30.9±6.4 mm Hg, PvCO2=34.0±8.8 mm Hg) except for the venous partial pressure of O2, which was lower in the F98/Fischer model (C6/Wistar: PvO2=54.4±6.1 mm Hg; F98/Fischer: PvO2=40.6±3.8 mm Hg). Arterial blood gases were within the normal physiologic range for isoflurane-anesthetized rats.

Microvascular Features of C6/Wistar and F98/Fischer Differ

Three ROIs (cortex, striatum, and tumor) were manually drawn on the T2-weighted image and transferred to each MRI map (Figure 1A). For each ROI, the cumulated number of voxels across all animals is as follows: tumor, n=3023 (Wistar/C6: 1500, Fischer/F98: 1523); striatum, n=579 (Wistar/C6: 315, Fischer/F98: 264); cortex, n=811 (Wistar/C6: 417, Fischer/F98: 394). After removal of voxels with non-physiologic values, the ROI-based measurement was performed on a satisfactory percentage of voxels (Figure 1B, more than 95% for the cortex and striatum ROIs and more than 85% for the tumor ROI).

Figure 1.

Figure 1

Open in a new tab

Multiparametric magnetic resonance imaging (MRI) in C6/Wistar and F98/Fischer models of glioma. (A) In addition to an anatomic T2-weighted (T2W) images, six MRI parameters were mapped: apparent diffusion coefficient (ADC), area under the signal enhancement curve (AUCGd-DOTA), cerebral blood volume (CBV), cerebral blood flow (CBF), tissular oxygen saturation (StO2) and cerebral metabolic rate of oxygen (CMRO2). The three regions of interest (ROIs) are overlaid on the maps (C, contralateral cortex; S, contralateral striatum; T, tumor). (B) Included voxels within valid physiologic values for the subsequent ROI-based analysis. Values of ADC, AUCGd-DOTA, CBV, CBF, StO2, and CMRO2 for each ROI and each animal model. Mean±s.d., paired t-test: *P<0.05 within the ROI in the same animal model, unpaired t-test: $P<0.05 between C6/Wistar (n=13) and F98/Fischer (n=13) models.

A comparison between mean ROI values allows to clearly discriminate between the tumor and the contralateral tissue. Both tumor models have increased ADC of water, vessel wall permeability (AUCGd-DOTA), CBV fraction, and decreased CBF, StO2, and CMRO2 in comparison with at least one contralateral ROI (Figure 1B). Between the C6 and the F98 gliomas, ADC, AUCGd-DOTA, CBF, and CMRO2 differ whereas CBV and StO2 are similar. Some differences for CBV, CBF, and StO2 are also found in the striatum ROI between Fischer and Wistar rats.

Both tumors exhibit a strong spatial heterogeneity in most parametric maps (Figure 1A). For example, the F98 glioma clearly exhibits a low CBF in its center and a high CBF at its periphery. On average, the mean intra-ROI variability of MRI estimates is 1.8 to 44 times higher in the tumor ROI than in the contralateral striatum ROI (tumor versus striatum expressed as fold increase; C6: ADC=3.0, AUCGd-DOTA=43.4, CBV=4.0, CBF=1.8, StO2=5.3, CMRO2=6.9; F98: ADC=5.9, AUCGd-DOTA=30.7, CBV=2.4, CBF=1.7, StO2=8.9, CMRO2=10.7). The data analysis based on mean values over ROIs masks this spatial heterogeneity. Note that the error bars in Figure 1B display the inter-animal variability and do not represent the intra-ROI variability.

The Distribution of Clusters is Spatially Structured

To enable tumor characterization, we used an unsupervised, model-based, clustering analysis, which integrates the six quantitative MRI parameters (ADC, AUCGd-DOTA, CBV, CBF, StO2, CMRO2) to identify clusters of voxels sharing the same MRI characteristics. This analysis was first performed on all voxels of the brain slices, and then on voxels of the previously defined ROIs only, the data from both tumor models being pooled. In both analyses, the optimal number of clusters was determined using a Bayesian information criterion (Supplementary Figures S1C and S2A) and was estimated to five clusters. When taking into account all voxels of the brain slices, the resulting cluster distribution shows more diversity in the non-tumoral tissue than in the tumoral tissue (Supplementary Figure S1B).

When taking into account the voxels within the three ROIs, a visual inspection of the cluster distribution (Figure 2) and a quantitative analysis of this distribution (Figure 3) show spatially coherent groups of voxels assigned to a same cluster. Cluster a, (Figure 2, purple) is specific of contralateral ROIs (Figure 3, colored pies). Cluster e (Figure 2, red) is specific of tumor ROI (Figure 3, colored pies). Cluster b (Figure 2, blue) is mainly found in C6/Wistar (Figure 3, gray pies) and is distributed in both contralateral and tumor ROIs. The two clusters c and d (Figure 2, yellow and orange) are mainly found in tumor ROI. The excluded voxels, not classified, are mainly found in the tumor ROI.

Figure 2.

Figure 2

Open in a new tab

Cluster assignment within the region of interest (ROI) overlaid on the T2-weighted (T2W) anatomic image (A) Clusters a (purple) and b (blue) are mostly found in the contralateral striatum and cortex ROIs. Cluster c (yellow), d (orange) and e (red) are mostly found in the tumor ROI. Clusters c is mostly found in C6 tumors, cluster d in both tumors, and cluster e mostly in F98 tumors. (B) Web graph representing the mean standardized value for each magnetic resonance imaging parameter and each cluster. Each map corresponds to one animal (C6/Wistar: n=13, F98/Fischer: n=13). ADC, apparent diffusion coefficient; AUCGd-DOTA, area under the signal enhancement curve; CBF, cerebral blood flow; CBV, cerebral blood volume; CMRO2, cerebral metabolic rate of oxygen; StO2, tissular oxygen saturation.

Figure 3.

Figure 3

Open in a new tab

The five clusters are labeled from ‘healthy' (voxels mainly located in the contralateral striatum and cortex regions of interest (ROIs)) to ‘pathologic' (voxels mainly located in the tumor ROI) clusters. The green and pink pies represent the distribution of voxels assigned to a given cluster between the three ROIs. The black and gray pies represent the distribution of voxels assigned to a given cluster between the two animal models. The diameter of each pie graph is proportional to the number of voxels assigned to the cluster (a pie scale is displayed on the left). The web graphs represent the mean standardized value of each magnetic resonance imaging (MRI) parameter for each cluster. The table presents, for each cluster, the number of assigned voxels and the mean value of each MRI parameter (mean±s.d.). Animal number=26, C6/Wistar: n=13, F98/Fischer: n=13. ADC, apparent diffusion coefficient; a.u., arbitrary unit; AUCGd-DOTA, area under the signal enhancement curve; CBF, cerebral blood flow; CBV, cerebral blood volume; CMRO2, cerebral metabolic rate of oxygen; StO2, tissular oxygen saturation.

Mean quantitative values of each MRI parameter in each cluster are detailed in the table of Figure 3. Cluster a, specifically found in contralateral ROIs, exhibits levels of CBF, StO2, and CBV values corresponding to a healthy tissue. Cluster b was found in both healthy and tumor ROIs in the C6/Wistar model. Magnetic resonance imaging parameters are clearly those of a healthy tissue. For pixels assigned to cluster b and located in the C6 tumor, one can thus assume that the infiltration is not high enough to significantly modify the estimated MRI parameters. Clusters d, mainly found in tumors, and e, strictly observed in tumors, show typical tumor characteristics: a high vascular permeability, a high ADC, and a reduced CBF. The characteristics of cluster c fall in between healthy and tumor tissues: a slightly increased ADC, vascular permeability, and CBV but a normal CBF. The mean quantitative values of each MRI parameter are presented as a web graph for each cluster using standardized parameter values (Figure 3, third row). These web graphs yield a graphical overview of the changes in parameter values across clusters a to e.

Regarding the distribution of each cluster in the two animal models, cluster a is mainly found in Fischer rats (65%) and cluster b in Wistar rats (86%) (Figure 3, black and gray pies). Similarly, cluster c is mainly found in the C6 tumor (59%) while cluster e is preferentially found in the F98 tumor (73%).

The cluster distribution in tumors is more heterogeneous than in contralateral tissues (Figure 2). Eight C6 tumors exhibit a bipolar pattern: the cortical part of the tumor (clusters d and e) differs from its striatal part (clusters b and c). As this striatal part is a mixture of healthy and tumoral clusters, it suggests an infiltrative behavior of the tumor. The five remaining C6 glioma shows a homogeneous cluster distribution (mainly representing cluster c). The F98 tumors are mostly composed of clusters c, d and e. Cluster c is found at the tumor periphery. Cluster d and e are found in the tumor center. One F98 tumor exhibits a bipolar pattern. As opposed to C6 glioma, very few voxels in the tumor are assigned to cluster b. Overall, a coherent spatial distribution of clusters appears in each tumor model.

Each Cluster Corresponds to a Different Histologic Signature

Immunohistological analysis was performed to detect cells and necrosis (hematoxylin–erythrosine and DAPI), hypoxia (pimonidazole), and blood vessels (collagen IV).

In contralateral brain, clusters a and b represent what we can consider as normal blood vessel density and diameter, normal cell density, and no hypoxia (Figure 4). Note that collagen IV (higher magnification) shows a higher blood vessel density in the cortex than in the striatum. This difference is in line with the CBV differences observed between these two clusters (web graph, Figures 2 and 4). Cluster b is also found in tumors (mostly C6), suggesting that this part of the tumor exhibits some similarity with ‘normal' tissue. Indeed, the histologic signature corresponds to almost normal tissue with few disseminated tumor cell foci, a slightly lower vessel density than in contralateral striatum, and no hypoxia. The histologic signature of cluster c (yellow) differs from that of cluster b. For the C6 glioma, cluster c is found in the tumor center with a high cell density. For the F98 glioma, cluster c is found at the tumor periphery with small spots of increased cell density invading normal tissue. In both cases, a vasogenic edema is observed on the hematoxylin–erythrosine sections and might account for the increased ADC. The vessel density is decreased but vessel diameter is clearly increased, which is compatible with the increased CBV observed with MRI and the absence of hypoxia assessed by both MRI and pimonidazole. Cluster d is mostly found within hypoxic areas with low vessel density. These features are in line with the low StO2 level and the low CBV observed by MRI. For both gliomas, small necrotic foci are observed. They could account for the high ADC, which characterizes this cluster. The histologic signature of cluster e is close to that of cluster c in C6 glioma. The difference between clusters e and c or d arises from a reduced CBF and an increased AUCGd-DOTA. These two features cannot be detected on the displayed histologic sections. Excluded voxels (x) are mainly found in large necrotic areas. As no vessels are present, the contrast agent does not flow through these necrotic areas, CBV, StO2, and CMRO2 cannot be determined and voxels are excluded. It, however, appears that the number of excluded voxels provides complementary information about the tumor.

Figure 4.

Figure 4

Open in a new tab

Histologic description of each cluster. To describe each cluster, a circle was drawn on the magnetic resonance imaging (MRI) map and the corresponding area is depicted as a rectangle placed on the three subsequent histologic sections (hematoxylin–erythrosine (HE), pimonidazole, and collagen IV/4′,6-diamidino-2-phenylindole (DAPI)). The last histologic section corresponds to a zoom in the previously mentioned rectangles. For sake of clarity, the mean standardized value for each MRI parameter is displayed below the cluster label. Cluster a and cluster b in contralateral tissue. Tissue is normoxic (no pimonidazole detection) with a high density of thin vessels (collagen IV). Cluster b in tumor tissue. Low tumor cell density (HE and DAPI), normoxia (no pimonidazole detection), blood vessel density is slightly lower than that of contralateral tissue. Cluster b in tumor reflects infiltration of tumor cells in normal brain tissue. Cluster c. Compared with cluster b, the tumor cell density and vessel diameter are larger. Cluster d. The tissue is hypoxic (pimonidazole detection), exhibits rare blood vessels, and a lower tumor cell density than cluster c. Clusters d are found near excluded voxels (marked as x), which correspond to necrotic or hemorrhagic areas. Cluster e. The number of blood vessels and the density of tumor cells are comparable with that of cluster c. From a histologic point of view, clusters e and c appear comparable. In MRI, they appear different owing to the difference in blood flow and permeability. Scale bar, 1 mm (HE, pimonidazole, DAPI/collagen IV). Scale bar, 200 μm (magnified DAPI/collagen IV). ADC, apparent diffusion coefficient; AUCGd-DOTA, area under the signal enhancement curve; CBF, cerebral blood flow; CBV, cerebral blood volume; CMRO2, cerebral metabolic rate of oxygen; StO2, tissular oxygen saturation.

Composition of Clusters within a Region of Interest Allows Tissue Type Prediction

Each ROI is characterized by its cluster composition and is classified into six classes (Wistar cortex, Wistar striatum, C6 tumor, Fischer cortex, Fischer striatum, F98 tumor). To evaluate misclassification rates based on cluster composition, a cross-validation procedure was performed on all seventy-eight ROIs after applying LDA (Table 1). Among the 26 tumor ROIs, 11 out of 13 ROIs (84.6%) are correctly predicted for each tumor model. Tumor ROIs not correctly predicted (15.4%) are attributed to the other tumor model. Among the 52 healthy ROIs, 20 out of 26 ROIs (76.9%) are correctly predicted for the Wistar rat strain and 25 out of 26 ROIs (96.2%) for the Fischer rat strain. The prediction is less accurate for cortex ROIs (14 out of 26, i.e., 53.8%, are correctly predicted) than for the striatum ROIs (19 out of 26 ROI, i.e., 73.1%, are correctly predicted). Performing the clustering on the voxels from the two tumor ROIs did not change the level of prediction (data not shown).

Table 1. Tissue type prediction based on the ‘leave-one-out' procedure.

ROIs Class prediction
  C6 Wistar
 F98 Fischer
  S C T S C T
C6 Wistar
 S (13) 10 3        
 C (13) 4 3     6  
 T (13)     11     2
             
F98 Fischer
 S (13) 1     9 3  
 C (13)       2 11  
 T (13)     2     11
Open in a new tab

C, cortex; ROIs, regions of interest; S, striatum; T, tumor.

Each of the 78 ROIs (26 animals, 3 ROIs: tumor (T), Cortex (C), Striatum (S)) was assigned to one among six classes: C6 tumor, Wistar cortex, Wistar striatum, F98 tumor, Fischer cortex, Fischer striatum.

We further explored the discriminatory power of each cluster and each MRI parameter for tissue prediction. Table 2 shows the classification accuracy when using all clusters and when removing a cluster from the analysis. Whatever the cluster removed, the overall classification is not improved. Similarly, Table 3 shows the classification accuracy when using all MRI parameters and when removing a MRI parameter from the analysis. Whatever the MRI parameter removed, the overall classification is not improved.

Table 2. Classification power of each cluster on the classification.

ROIs All clusters are used One cluster is removed
    a b c d e
C6 Wistar
 S (13) 76.9 61.5 76.9 69.2 76.9 76.9
 C (13) 23.1 30.8 23.1 23.1 23.1 23.1
 T (13) 84.6 92.3 76.9 69.2 76.9 69.2
             
F98 Fischer
 S (13) 69.2 93.3 38.5 69.2 69.2 69.2
 C (13) 84.6 61.5 84.6 84.6 84.6 84.6
 T (13) 84.6 92.3 46.2 92.3 84.6 84.6
Open in a new tab

C, cortex; ROIs, regions of interest; S, striatum; T, tumor.

Value: % of correct prediction.

Each of the 78 ROIs (26 animals, 3 ROIs: tumor (T), cortex (C), striatum (S)) was assigned with a ‘leave-one-out' procedure to one among six classes: C6 tumor, Wistar cortex, Wistar striatum, F98 tumor, Fischer cortex, Fischer striatum. To assess the contribution of each cluster to the classification, the percentage of correct prediction for each tissue type is presented when all clusters are used and when one cluster is excluded.

Table 3. Contribution of each MRI parameter on the classification.

ROIs All MRI parameters are used One MRI parameter is excluded
    ADC CBV CBF AUCGd-DOTA SO2 CMRO2
C6 Wistar
 S (13) 76.9 76.9 76.9 53.8 61.5 15.4 15.4
 C (13) 23.1 76.9 69.2 69.2 92.3 69.2 69.2
 T (13) 84.6 15.4 23.1 7.7 53.8 15.4 15.4
               
F98 Fischer
 S (13) 69.2 76.9 92.3 76.9 84.6 38.5 23.1
 C (13) 84.6 76.9 92.3 69.2 84.6 84.6 84.6
 T (13) 84.6 84.6 84.6 92.3 76.9 84.6 84.6
Open in a new tab

ADC, apparent diffusion coefficient; AUCGd-DOTA, area under the signal enhancement curve; C, cortex; CBF, cerebral blood flow; CBV, cerebral blood volume; CMRO2, cerebral metabolic rate of oxygen; MRI, magnetic resonance imaging; ROIs, regions of interest; S, striatum; StO2, tissular oxygen saturation; T, tumor.

Value: % of correct prediction.

Each of the 78 ROIs (26 animals, 3 ROIs: tumor (T), cortex (C), striatum (S)) was assigned with a ‘leave-one-out' procedure to one among six classes: C6 tumor, Wistar cortex, Wistar striatum, F98 tumor, Fischer cortex, Fischer striatum. To assess the contribution of each MRI parameter to the classification, the percentage of correct prediction for each tissue type is presented when all MRI parameters are used and when one MRI parameter is excluded.

Discussion

This study presents a combination of spatial selection and of unsupervised, model-based cluster analysis to integrate the large amount of information obtained with microvascular MRI in the context of brain tumors. The identified clusters, defined by their microvascular-based MRI features, correspond to different histologic features, and reveal an important intratumoral spatial structure. This suggests that the presence of a cluster within a tumor can be used to assess the presence of a tissue type within this tumor. At a larger scale, the cluster composition of a given ROI, i.e., a signature of the intratumoral structure, could be used to characterize tumor models as well as histology does.

Imaging complex cancer lesions, which are heterogeneous in space and time, is a real challenge. For tumor diagnosis, histology is the reference. However, several studies report that histologic diagnosis of glioma established after stereotactic biopsy, from a single point in the tumor, is associated with a substantial risk of inaccuracy.24 Moreover, a biopsy cannot always be obtained: some tumors are not accessible to the surgeon and no repeated biopsies of a brain tumor are performed during the therapeutic follow-up.

To characterize the entire tumor in a non-invasive way, we propose the use of quantitative microvascular imaging. Indeed, it is well known that a developing tumor impacts vessel wall permeability, blood volume and flow, oxygenation and others physiologic parameters.25, 26, 27, 28, 29 Moreover, the physiologic links between CBF, CBV, StO2, CMRO2 are altered in tumors, as can be seen from the independent variations of perfusion estimates between clusters (Figure 3). Microvasculature is also a target for several anti-tumoral strategies. Beyond tumors, several brain diseases impact the microvasculature, such as stroke,30, 31 Alzheimer,32 or epilepsy.33 Outside of the brain, several diseases, such as hypertension, diabetes impact the microvasculature as well. Quantitative microvascular imaging, which assesses both structural and functional aspects of perfusion, appears thus as a promising multiparametric MRI approach to obtain microvascular parameter-based in vivo images closely resembling histologic features.

To obtain microvascular MR-resembling histology,9 several approaches have used classification algorithms.10, 11 They are however limited by the absence of an underlying statistical model, which prevents the use of probabilistic tools. In this study, we used an unsupervised, model-based, classification approach implemented in the software mclust. A comparison study provided evidence that mclust outperformed other clustering methods (k-means, hierarchical clustering, Supplementary Figure S3).

The use of model-based clustering on the entire brain led to a larger number of clusters in healthy tissue than in tumor. To focus on the tumor, we coupled spatial knowledge with the use of the clustering approach, as previously proposed with machine learning techniques.13, 34, 35, 36 The conservative strategy used for tumor delineation in this study is not optimal. Instead of using a two-step approach (selection, clustering), further clustering strategies should directly combine spatial and physiologic informations. Such a combination should allow studying the tumor periphery, which cannot be properly estimated with manual delineation.

The use of contralateral tissue was the key to highlight tumor regions with mixed tumoral and healthy properties (cluster b in tumor) that were related to the infiltrative component of the C6 tumor, in agreement with the infiltrative behavior of the C6 model.37 The combination of a spatial selection and an unsupervised clustering approach yields statistically characterized data on the tumor, including its infiltrative part (Supplementary Figure S2).

The clusters defined with an unsupervised approach and using quantitative microvascular MRI characteristics were in line with the histologic features assessed in this study. This result relies on previously validated MRI methods. Indeed, the estimate of CBV was validated against histology,38 the CBF against PET39 and autoradiography,39 the permeability against autoradiography,40 the oxygenation against gas challenges17 and pimonidazole.19 As observed in the MRI maps, each parameter exhibits a different spatial distribution. This difference between spatial patterns is exploited by the unsupervised cluster analysis and yields the intratumoral structures reported in this study. Removing an MRI parameter from the analysis did not improve the overall tissue classification. Furthermore, as therapies can modulate all the measured MRI parameters,18, 41, 42 it therefore appears important to include them all.

In case of stereotactic biopsy performed for diagnostic tumor grading, a small piece of the tumor is characterized by a pathologist. The limits of this approach are well known: the small tumor sample may not arise from the most aggressive part of the tumor. Moreover, it may not represent the tumor heterogeneity. The cluster analysis proposed in this study could provide an in vivo histology-resembling image of the entire tumor. While this approach remains to be evaluated in humans, several applications can already be foreseen:

  • Design of homogeneous groups for preclinical pharmacological studies by removing tumors whose cluster composition is too far from the mean of the group or from a reference cluster data set. Previous reports have shown that two subtypes of the same tumor can indeed yield two different glioblastoma phenotypes in vivo;43

  • Classification of tumors in patients, using a reference database;

  • Evaluation of tumor evolution or tumor response after therapy, using the cluster composition of the tumor as the tumor signature.

The combination of spatial selection and cluster analysis can be used to structure the wealth of information gathered with multiparametric microvascular MRI. This approach, which highlights several types of pathophysiological features across two glioma models, is well suited for the characterization of tumor heterogeneity. Its translation to a clinical setting has a potential to ease tumor diagnosis, prognosis, and treatment follow-up.

Acknowledgments

We thank Fondation ARC and ANR Imoxy for their financial support and the animal care facility of GIN, the IRMaGe facility, Yasmina Saoudi and Carole Carcenac, for their technical support. This work was partly funded by the French program ‘Investissement d'Avenir' run by the ‘Agence Nationale pour la Recherche' grant ‘Infrastructure d'avenir en Biologie Santé' - ANR-11-INBS-0006.

The authors declare no conflict of interest.

Footnotes

Supplementary Information accompanies the paper on the Journal of Cerebral Blood Flow & Metabolism website (http://www.nature.com/jcbfm)

Supplementary Material

Supplementary Figure 1
Click here for additional data file. (9MB, tif)
Supplementary Figure 2
Click here for additional data file. (8.4MB, tif)
Supplementary Figure 3
Click here for additional data file. (2.2MB, tif)
Supplementary Figure Legends
Click here for additional data file. (63KB, doc)

References

  1. Preul MC, Caramanos Z, Leblanc R, Villemure JG, Arnold DL. Using pattern analysis of in vivo proton MRSI data to improve the diagnosis and surgical management of patients with brain tumors. NMR Biomed. 1998;11:192–200. doi: 10.1002/(sici)1099-1492(199806/08)11:4/5<192::aid-nbm535>3.0.co;2-3. [DOI] [PubMed] [Google Scholar]
  2. Lin G-C, Wang W-J, Wang C-M, Sun S-Y. Automated classification of multi-spectral MR images using linear discriminant analysis. Comput Med Imaging Graph. 2010;34:251–268. doi: 10.1016/j.compmedimag.2009.11.001. [DOI] [PubMed] [Google Scholar]
  3. Li Y, Sima DM, Van Cauter S, Himmelreich U, Sava ARC, Pi Y, et al. Unsupervised nosologic imaging for glioma diagnosis. IEEE Trans Biomed Eng. 2013;60:1760–1763. doi: 10.1109/TBME.2012.2228651. [DOI] [PubMed] [Google Scholar]
  4. Barbier EL, Lamalle L, Décorps M. Methodology of brain perfusion imaging. J Magn Reson Imaging. 2001;13:496–520. doi: 10.1002/jmri.1073. [DOI] [PubMed] [Google Scholar]
  5. Troprès I, Grimault S, Vaeth A, Grillon E, Julien C, Payen J-F, et al. Vessel size imaging. Magn Reson Med. 2001;45:397–408. doi: 10.1002/1522-2594(200103)45:3<397::aid-mrm1052>3.0.co;2-3. [DOI] [PubMed] [Google Scholar]
  6. He X, Yablonskiy DA. Quantitative BOLD: mapping of human cerebral deoxygenated blood volume and oxygen extraction fraction: default state. Magn Reson Med. 2007;57:115–126. doi: 10.1002/mrm.21108. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Tofts PS, Brix G, Buckley DL, Evelhoch JL, Henderson E, Knopp MV, et al. Estimating kinetic parameters from dynamic contrast-enhanced T1-weighted MRI of a diffusable tracer: standardized quantities and symbols. J Magn Reson Imaging. 1999;10:223–232. doi: 10.1002/(sici)1522-2586(199909)10:3<223::aid-jmri2>3.0.co;2-s. [DOI] [PubMed] [Google Scholar]
  8. Krainik A, Hund-Georgiadis M, Zysset S, von Cramon DY. Regional impairment of cerebrovascular reactivity and BOLD signal in adults after stroke. Stroke. 2005;36:1146–1152. doi: 10.1161/01.STR.0000166178.40973.a7. [DOI] [PubMed] [Google Scholar]
  9. Dominietto M, Rudin M. Could magnetic resonance provide in vivo histology. Front Syst Biol. 2014;4:298. doi: 10.3389/fgene.2013.00298. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. De Edelenyi FS, Rubin C, Estève F, Grand S, Décorps M, Lefournier V, et al. A new approach for analyzing proton magnetic resonance spectroscopic images of brain tumors: nosologic images. Nat Med. 2000;6:1287–1289. doi: 10.1038/81401. [DOI] [PubMed] [Google Scholar]
  11. Devos A, Simonetti AW, van der Graaf M, Lukas L, Suykens JAK, Vanhamme L, et al. The use of multivariate MR imaging intensities versus metabolic data from MR spectroscopic imaging for brain tumour classification. J Magn Reson. 2005;173:218–228. doi: 10.1016/j.jmr.2004.12.007. [DOI] [PubMed] [Google Scholar]
  12. MacQueen JB.Some Methods for Classification and Analysis of MultiVariate ObservationsIn: Cam LML, Neyman J, (eds).. Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability University of California Press; 1967281–297. [Google Scholar]
  13. Mouthuy N, Cosnard G, Abarca-Quinones J, Michoux N. Multiparametric magnetic resonance imaging to differentiate high-grade gliomas and brain metastases. J Neuroradiol. 2012;39:301–307. doi: 10.1016/j.neurad.2011.11.002. [DOI] [PubMed] [Google Scholar]
  14. Fraley C, Raftery AE. Model-based clustering, discriminant analysis, and density estimation. J Am Stat Assoc. 2002;97:611–631. [Google Scholar]
  15. Coquery N, Pannetier N, Farion R, Herbette A, Azurmendi L, Clarencon D, et al. Distribution and radiosensitizing effect of cholesterol-coupled dbait molecule in rat model of glioblastoma. PLoS One. 2012;7:e40567. doi: 10.1371/journal.pone.0040567. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Lyng H, Sundfør K, Rofstad EK. Oxygen tension in human tumours measured with polarographic needle electrodes and its relationship to vascular density, necrosis and hypoxia. Radiother Oncol. 1997;44:163–169. doi: 10.1016/s0167-8140(97)01920-8. [DOI] [PubMed] [Google Scholar]
  17. Christen T, Lemasson B, Pannetier N, Farion R, Segebarth C, Rémy C, et al. Evaluation of a quantitative blood oxygenation level-dependent (qBOLD) approach to map local blood oxygen saturation. NMR Biomed. 2011;24:393–403. doi: 10.1002/nbm.1603. [DOI] [PubMed] [Google Scholar]
  18. Lemasson B, Christen T, Tizon X, Farion R, Fondraz N, Provent P, et al. Assessment of multiparametric MRI in a human glioma model to monitor cytotoxic and anti-angiogenic drug effects. NMR Biomed. 2011;24:473–482. doi: 10.1002/nbm.1611. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Lemasson B, Christen T, Serduc R, Maisin C, Bouchet A, Le Duc G, et al. Evaluation of the relationship between MR estimates of blood oxygen saturation and hypoxia: effect of an antiangiogenic treatment on a gliosarcoma model. Radiology. 2012;265:743–752. doi: 10.1148/radiol.12112621. [DOI] [PubMed] [Google Scholar]
  20. Troprès I, Lamalle L, Péoc'h M, Farion R, Usson Y, Décorps M, et al. In vivo assessment of tumoral angiogenesis. Magn Reson Med. 2004;51:533–541. doi: 10.1002/mrm.20017. [DOI] [PubMed] [Google Scholar]
  21. Lemasson B, Serduc R, Maisin C, Bouchet A, Coquery N, Robert P, et al. Monitoring blood-brain barrier status in a rat model of glioma receiving therapy: dual injection of low-molecular-weight and macromolecular MR contrast media. Radiology. 2010;257:342–352. doi: 10.1148/radiol.10092343. [DOI] [PubMed] [Google Scholar]
  22. Alsop DC, Detre JA. Reduced transit-time sensitivity in noninvasive magnetic resonance imaging of human cerebral blood flow. J Cereb Blood Flow Metab. 1996;16:1236–1249. doi: 10.1097/00004647-199611000-00019. [DOI] [PubMed] [Google Scholar]
  23. Buxton RB, Frank LR, Wong EC, Siewert B, Warach S, Edelman RR. A general kinetic model for quantitative perfusion imaging with arterial spin labeling. Magn Reson Med. 1998;40:383–396. doi: 10.1002/mrm.1910400308. [DOI] [PubMed] [Google Scholar]
  24. Trembath D, Miller CR, Perry A. Gray zones in brain tumor classification: evolving concepts. Adv Anat Pathol. 2008;15:287–297. doi: 10.1097/PAP.0b013e3181836a03. [DOI] [PubMed] [Google Scholar]
  25. Cha S. Neuroimaging in neuro-oncology. Neurotherapeutics. 2009;6:465–477. doi: 10.1016/j.nurt.2009.05.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Galbán CJ, Chenevert TL, Meyer CR, Tsien C, Lawrence TS, Hamstra DA, et al. The parametric response map is an imaging biomarker for early cancer treatment outcome. Nat Med. 2009;15:572–576. doi: 10.1038/nm.1919. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Hamstra DA, Chenevert TL, Moffat BA, Johnson TD, Meyer CR, Mukherji SK, et al. Evaluation of the functional diffusion map as an early biomarker of time-to-progression and overall survival in high-grade glioma. Proc Natl Acad Sci USA. 2005;102:16759–16764. doi: 10.1073/pnas.0508347102. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Aronen HJ, Gazit IE, Louis DN, Buchbinder BR, Pardo FS, Weisskoff RM, et al. Cerebral blood volume maps of gliomas: comparison with tumor grade and histologic findings. Radiology. 1994;191:41–51. doi: 10.1148/radiology.191.1.8134596. [DOI] [PubMed] [Google Scholar]
  29. Emblem KE, Nedregaard B, Nome T, Due-Tonnessen P, Hald JK, Scheie D, et al. Glioma grading by using histogram analysis of blood volume heterogeneity from MR-derived cerebral blood volume maps. Radiology. 2008;247:808–817. doi: 10.1148/radiol.2473070571. [DOI] [PubMed] [Google Scholar]
  30. Moisan A, Pannetier N, Grillon E, Richard M-J, de Fraipont F, Rémy C, et al. Intracerebral injection of human mesenchymal stem cells impacts cerebral microvasculature after experimental stroke: MRI study. NMR Biomed. 2012;25:1340–1348. doi: 10.1002/nbm.2806. [DOI] [PubMed] [Google Scholar]
  31. Boehm-Sturm P, Farr TD, Adamczak J, Jikeli JF, Mengler L, Wiedermann D, et al. Vascular changes after stroke in the rat: a longitudinal study using optimized magnetic resonance imaging. Contrast Media Mol Imaging. 2013;8:383–392. doi: 10.1002/cmmi.1534. [DOI] [PubMed] [Google Scholar]
  32. Biron KE, Dickstein DL, Gopaul R, Jefferies WA. Amyloid triggers extensive cerebral angiogenesis causing blood brain barrier permeability and hypervascularity in Alzheimer's disease. PLoS One. 2011;6:e23789. doi: 10.1371/journal.pone.0023789. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Marchi N, Lerner-Natoli M. Cerebrovascular remodeling and epilepsy. Neuroscientist. 2012;19:304–312. doi: 10.1177/1073858412462747. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Zacharaki EI, Wang S, Chawla S, Soo Yoo D, Wolf R, Melhem ER, et al. Classification of brain tumor type and grade using MRI texture and shape in a machine learning scheme. Magn Reson Med. 2009;62:1609–1618. doi: 10.1002/mrm.22147. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Zacharaki EI, Kanas VG, Davatzikos C. Investigating machine learning techniques for MRI-based classification of brain neoplasms. Int J Comput Assist Radiol Surg. 2011;6:821–828. doi: 10.1007/s11548-011-0559-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Zöllner FG, Emblem KE, Schad LR. SVM-based glioma grading: optimization by feature reduction analysis. Z Für Med Phys. 2012;22:205–214. doi: 10.1016/j.zemedi.2012.03.007. [DOI] [PubMed] [Google Scholar]
  37. Jacobs VL, Valdes PA, Hickey WF, De Leo JA. Current review of in vivo GBM rodent models: emphasis on the CNS-1 tumour model. ASN Neuro. 2011;3:171–181. doi: 10.1042/AN20110014. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Valable S, Lemasson B, Farion R, Beaumont M, Segebarth C, Remy C, et al. Assessment of blood volume, vessel size, and the expression of angiogenic factors in two rat glioma models: a longitudinal in vivo and ex vivo study. NMR Biomed. 2008;21:1043–1056. doi: 10.1002/nbm.1278. [DOI] [PubMed] [Google Scholar]
  39. Boś A, Bergmann R, Strobel K, Hofheinz F, Steinbach J, van den Hoff J. Cerebral blood flow quantification in the rat: a direct comparison of arterial spin labeling MRI with radioactive microsphere PET. EJNMMI Res. 2012;2:47. doi: 10.1186/2191-219X-2-47. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Paudyal R, Ewing JR, Nagaraja TN, Bagher-Ebadian H, Knight RA, Panda S, et al. The concordance of MRI and quantitative autoradiography estimates of the transvascular transfer rate constant of albumin in a rat brain tumor model. Magn Reson Med. 2011;66:1422–1431. doi: 10.1002/mrm.22914. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Chung C, Jalali S, Foltz W, Burrell K, Wildgoose P, Lindsay P, et al. Imaging biomarker dynamics in an intracranial murine glioma study of radiation and antiangiogenic therapy. Int J Radiat Oncol. 2013;85:805–812. doi: 10.1016/j.ijrobp.2012.07.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Hamstra DA, Rehemtulla A, Ross BD. Diffusion magnetic resonance imaging: a biomarker for treatment response in Oncology. J Clin Oncol. 2007;25:4104–4109. doi: 10.1200/JCO.2007.11.9610. [DOI] [PubMed] [Google Scholar]
  43. Thorsen F, Jirak D, Wang J, Sykova E, Bjerkvig R, Enger PØ, et al. Two distinct tumor phenotypes isolated from glioblastomas show different MRS characteristics. NMR Biomed. 2008;21:830–838. doi: 10.1002/nbm.1263. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Figure 1
Click here for additional data file. (9MB, tif)
Supplementary Figure 2
Click here for additional data file. (8.4MB, tif)
Supplementary Figure 3
Click here for additional data file. (2.2MB, tif)
Supplementary Figure Legends
Click here for additional data file. (63KB, doc)

Articles from Journal of Cerebral Blood Flow & Metabolism are provided here courtesy of SAGE Publications

RESOURCES