|Home | About | Journals | Submit | Contact Us | Français|
Angiogenesis in breast cancer helps fulfill the metabolic demands of the progressing tumor and plays a critical role in tumor metastasis. Therefore, various imaging modalities have been used to characterize tumor angiogenesis. While micro-CT (μCT) is a powerful tool for analyzing the tumor microvascular architecture at micron-scale resolution, magnetic resonance imaging (MRI) with its sub-millimeter resolution is useful for obtaining in vivo vascular data (e.g. tumor blood volume and vessel size index). However, integration of these microscopic and macroscopic angiogenesis data across spatial resolutions remains challenging. Here we demonstrate the feasibility of ‘multiscale’ angiogenesis imaging in a human breast cancer model, wherein we bridge the resolution gap between ex vivo μCT and in vivo MRI using intermediate resolution ex vivo MR microscopy (μMRI). To achieve this integration, we developed suitable vessel segmentation techniques for the ex vivo imaging data and co-registered the vascular data from all three imaging modalities. We showcase two applications of this multiscale, multi-modality imaging approach: (1) creation of co-registered maps of vascular volume from three independent imaging modalities, and (2) visualization of differences in tumor vasculature between viable and necrotic tumor regions by integrating μCT vascular data with tumor cellularity data obtained using diffusion-weighted MRI. Collectively, these results demonstrate the utility of ‘mesoscopic’ resolution μMRI for integrating macroscopic in vivo MRI data and microscopic μCT data. Although focused on the breast tumor xenograft vasculature, our imaging platform could be extended to include additional data types for a detailed characterization of the tumor microenvironment and computational systems biology applications.
Angiogenesis, the process of new blood vessel formation, plays a central role in breast cancer development, invasion and metastasis . Therefore, new approaches for elucidating the angiogenesis pathway and the effects of anti-angiogenic therapies in breast cancer are required to confer survival benefits for patients . Much of our understanding of angiogenesis has been the result of preclinical research  that includes a wide array of imaging studies . Microscopic imaging with optical methods has revealed different aspects of tumor angiogenesis . For example, intravital microscopy of window chamber tumor models has provided invaluable insights into vascular function, gene expression, the tumor microenvironment and drug delivery [6–8]. Compared to histology, advanced optical techniques do not require sectioning of the tumor and permit relatively larger spatial coverage. However, since most optical methods cannot image tissue deeper than ~1 mm, they are unsuitable for imaging the complete 3D vasculature of tumor xenografts . Furthermore, since angiogenic gene expression has been shown to differ between orthotopically and ectopically (e.g. in window chambers) grown tumors , one requires a technique capable of imaging the whole tumor.
In contrast, one can characterize angiogenesis in vivo on the macroscopic scale using computed tomography (CT) , or magnetic resonance imaging (MRI) [11, 12]. MRI is a versatile imaging modality for assessing tumor angiogenesis because it permits the repeated measurement of numerous angiogenesis-related parameters such as blood volume, vessel size index and vessel permeability . In vivo CT offers analogous parameters for quantifying in vivo angiogenesis [10, 14], but involves the use of X-ray radiation.
In spite of advances in imaging angiogenesis, integration of imaging data acquired at microscopic and macroscopic spatial scales is challenging for visualization  and computational applications . Therefore, the goal of this study was to develop and validate an imaging platform for characterizing breast cancer angiogenesis at multiple spatial scales. We imaged orthotopic MDA-MB-231 human breast cancer xenografts using three imaging modalities: (1) in vivo susceptibility contrast-enhanced MRI at the macroscopic scale (0.1 × 0.1 × 1 mm3 resolution), (2) ex vivo magnetic resonance microscopy (μMRI) at the mesoscopic scale (40 μm isotropic resolution) and (3) micro-CT (μCT) at the microscopic scale (8 μm isotropic resolution) (Fig. 1). Susceptibility contrast-enhanced MRI was used to measure in vivo tumor blood volume, while diffusion-weighted imaging was used to assess tumor cellularity by measuring the apparent diffusion coefficient (ADC). μCT provided high-resolution 3D images of the whole tumor microvasculature without the need for tissue maceration. Co-registration of these complementary tumor angiogenesis data was facilitated by mesoscopic scale μMRI, which exhibits both excellent soft tissue contrast (for co-registration with in vivo MRI) and visualization of blood vessel architecture (for co-registration with μCT) as demonstrated by us earlier .
We demonstrate the utility of our multiscale imaging platform by: generating co-registered blood volume maps from macroscopic, mesoscopic and microscopic angiogenesis data acquired using different contrast mechanisms; and integrating tumor vascular data with complementary information on tumor viability to show how the distribution of tumor blood vessels varies between viable and necrotic tissue. The availability of a preclinical multiscale imaging platform can provide a wealth of clinically relevant information about tumor angiogenesis. This includes identification of novel drug targets, development of novel antiangiogenic therapies, tests of therapeutic efficacy, development of novel clinical biomarkers, correlation of genotype with phenotype and ‘bioimage’ based computational models for cancer systems biology.
Human MDA-MB-231 breast cancer cells were cultured in RPMI-1640 medium (Sigma-Aldrich, St. Louis, MO), supplemented with 10 % fetal bovine serum and penicillin–streptomycin (Sigma-Aldrich). Three million cells were orthotopically inoculated in 50 μl of Hanks balanced salt solution (Sigma-Aldrich) in the left thoracic mammary fat pad of ten female athymic NCr nu/nu mice. All animal studies were performed in accordance with the institutional Animal Care and Use Committee guidelines.
Tumor xenografts were imaged with in vivo MRI, ex vivo μMRI and μCT. Five tumors were imaged at post inoculation week three (PIW3) and another five at PIW5.
For in vivo MRI, mice were anesthetized with a cocktail of ketamine and acepromazine, and the tail veins were cannulated for administration of the superparamagnetic iron oxide contrast agent Feridex® (Bayer Healthcare Pharmaceuticals Inc., Wayne, NJ). The mice were placed into a Bruker 9.4 Tesla horizontal bore MRI scanner with an 18 mm solenoid radio frequency (RF) coil and kept under isoflurane anesthesia while the body temperature was maintained at 37 °C using a heating blanket. The xenografts were imaged using the following pulse sequences: (1) Diffusion weighted (DW) images were acquired using a DW spin-echo sequence: echo time (TE) = 26.6 ms, repetition time (TR) = 1,000 ms, number of averages (NA) = 2. One non-DW image and three DW-images were acquired with a b-value of ~300 s/mm2 and diffusion sensitizing gradient orientations of [1,0,0], [0,1,0] and [0,0,1]. (2) T2*weighted (T2*w) images were acquired using a two-dimensional (2D) multi-echo gradient-echo (MGE) sequence with the following parameters: TE = 4.2/7.2/10.2/13.2/16.2/19.2 ms, TR = 800 ms, flip angle (θ) = 90°, NA = 4, in-plane resolution = 100 × 100 μm2, slice thickness = 1 mm. Next, Feridex® (3.73 mg/ml Fe in saline) was injected into the catheterized tail vein at a dose of ~25 mg/kg of Fe. After allowing 5 min for the contrast agent to equilibrate, T2*w scans were acquired to obtain post-contrast images.
After in vivo MRI, each tumor was perfusion fixed followed by perfusion with the radio-opaque silicon-rubber compound Microfil® (FlowTech Inc., Carver, MA), according to a method described by us in . After letting the Microfil® cure for ~60 min, the tumor and surrounding connective tissue were excised and fixed in zinc formalin. To preserve the anterior-posterior orientation of the tumor following excision, a cotton string soaked in Gd-DTPA (Magnevist®, Bayer Healthcare Pharmaceuticals) enriched water was attached with superglue along the anterior-posterior axis of the tumor. Two days later, excess connective tissue was trimmed from the tumor while leaving the cotton string attached.
Twenty-four hours before μMRI, each tumor was immersed in Phosphate Buffered Saline (PBS) doped with 1 mM Magnevist® for soft tissue contrast enhancement via shortening of the tissue T1. Right before the μMRI scans, each tumor was placed in an NMR tube filled with Fomblin® (Solvay Solexis, Milano, Italy), which provides background free MRI images because it does not supply any proton signal. Fomblin also prevents sample dehydration. Any trapped air bubbles were removed using a vacuum pump.
The tumors were imaged on a vertical bore Bruker 9.4 Tesla spectrometer using a 10 mm volume RF coil (Bruker BioSpin Corp, Billerica, MA) and a T2*w MGE sequence with the following parameters: TE = 4.9/9.7/14.5/19.3/24.0/28.8 ms, TR = 150 ms, θ = 30°, NA = 14, resolution = 40 × 40 × 40 μm3.
Following μMRI, samples were imaged by Numira Biosciences (Salt Lake City, UT) on a high-resolution, volumetric μCT scanner (μCT40, ScanCo Medical, Brüttisellen, CH). All images were acquired with the following parameters: 8 μm isotropic resolution at 55 kVp, 300 ms exposure time, 2,000 views and 5 frames per view.
Since we were interested in characterizing the tumor vasculature using multiple imaging modalities, it was necessary to develop vessel segmentation pipelines unique to each contrast mechanism. A brief description ensues.
3D MGE μMRI images corresponding to the first echo were imported into ImageJ (Rasband, W.S., National Institutes of Health, Bethesda, Maryland, USA, http://rsb.info.nih.gov/ij/). Blood vessels appear dark in the μMRI images due to the lack of mobile water protons in the polymerized Microfil®. The tumor vasculature was segmented from the background tissue in multiple steps as summarized in Fig. 2a.
First, a tubeness filter was used to determine how ‘tube-like’ each 3D structure in the image was. This filter is based on the computation of a 3D Hessian matrix which is described in detail in . It first convolves the 3D image data with a spherical Gaussian kernel with a standard deviation ‘σ’ that sensitizes the filter to tubes or vessels of different radii. We determined optimal values of σ = 0.8, 1.0, 1.2, 1.4 by visual inspection and based on recommendations for morphological segmentation described in . The outputs of the tubeness filter were combined via a voxel-wise ‘maximum’ operation (i.e. output intensity of voxel i image3i = max(image1i, image2i)) and a conservative threshold was applied to the result, which identified the most ‘tube-like’ structures and resulted in an initial binary vascular tree. Second, for regions that exhibited low contrast between the vasculature and tumor tissue, we enhanced the vessel extraction using a C-means clustering approach. The software MIPAV (Medical Image Processing, Analysis, and Visualization; National institute of Health, Bethesda, MD) was used to automatically classify three to five regions of different image intensities in the raw μMRI data. A lower threshold was applied to the tubeness output, which identified vessels in low contrast regions as well as ‘non-vascular’ structures in the tumor tissue. The latter were removed by masking out the regions of intermediate image intensity determined by the C-means classes.
The tubeness filter was not sensitive to vessels with diameters greater than approximately four μMRI voxels, which led to holes in these vessels, and using larger σ would have resulted in an unwanted dilation of smaller vessels. Therefore, in the third step, these vessels were segmented separately using C-means clustering. The class of voxels exhibiting the lowest signal intensity accounted for vessels greater than ~80 μm in radius that were not completely segmented in the previous steps. Next, the outputs of all three vessel extraction steps were combined, and a 3D morphological closing operation was performed in ImageJ with a spherical structuring element of one voxel radius to fill remaining holes. Finally, to eliminate small, disconnected objects in the extracted vasculature, 3D regions smaller than six connected pixels determined using a 26-pixel-connected neighborhood were removed.
Due to the excellent soft tissue contrast of μMRI, it was possible to manually mask non-tumor tissue using the segmentation editor in the Amira® visualization platform (VSG Inc., Burlington MA). This tumor mask was multiplied with the resulting binary vasculature to exclude non-tumor vasculature from the surrounding tissue.
Due to the radio-opacity of Microfil®, all tumor blood vessels appeared bright in the μCT images and the vasculature was extracted using a similar procedure as for the μMRI data (Fig. 3a). First, the 3D μCT data were imported into ImageJ and the tubeness filter applied with σ = 1.0 and 1.5. The results were combined via a ‘maximum’ operation and then subjected to a 3D mean filter to fill holes and create a well-connected vascular structure, which was then thresholded to produce a binarized vasculature. This threshold was chosen by visual comparison with the raw data and a compromise was made between segmenting small or low intensity blood vessels, dilating vessels and including noise. The selected threshold resulted in some noise (which was removed in the final processing step) while resultant vessel diameters were visually consistent with the raw data. To fill holes in large caliber blood vessels, a conservative intensity threshold was applied to the raw μCT data to segment vessels larger than approximately four μCT voxels in diameter, which were then combined with the initial binarized vessels. The final binarized vasculature underwent a 3D closing operation and in the final step, regions smaller than 27 connected voxels were removed.
The unprocessed in vivo MRI and μMRI data were co-registered using Amira®. The string fiducial attached to the tumor to indicate the anterior-posterior axis for ex vivo imaging helped to orient the μMRI dataset relative to the in vivo MRI data. A thorough comparison of the tumor shapes between both datasets was then used to define the final rigid-body transformation for the alignment of the μMRI tumor volume to the in vivo MRI volume. For the computational execution of the 3D co-registration, a Lanczos interpolation was applied to register the volumes while preserving voxel size. The same transformation was used to co-register the μMRI vascular data and the tumor mask to the in vivo MRI data. The tumor mask was downsampled to the in vivo MRI spatial resolution.
The co-registration of the unprocessed 3D μCT data to the transformed μMRI data was performed in two steps using Amira®. (1) Since the string fiducial was visible in both μMRI and μCT datasets, it was used for an initial manual alignment of the two image volumes. The transformation was executed using Lanczos interpolation while preserving the voxel size. (2) To improve the manual alignment, two fiducial sets were created by manually placing landmarks on corresponding vessels in the raw μMRI and μCT datasets. Following slice-wise inspection of the 3D μMRI and μCT data, characteristic vessel bifurcations or conspicuous vessels were identified in both datasets. Between 5 and 10 landmarks were set per tumor and used for the rigid-body landmark registration in which the μCT data was automatically aligned with the μMRI data by minimizing the spatial differences between the manually set landmarks in both datasets. The extracted μCT vasculature was transformed with the same parameters as those obtained from the registration of the raw μCT data with the μMRI data.
The quality of the vessel extraction was assessed by visual comparison with optical microscopy of whole mount tumor sections. Due to positive contrast of the Microfil in μCT images, it was possible to segment the superficial tumor blood vessels. The same vessels were then imaged by bright-field microscopy and visually compared to the μCT data. Superficial vessels were not visible in the μMRI data since both the microfilled vessels and background appear dark, so intratumoral vessels were compared with optical microscopy images as follows. A 1 mm thick tumor section was optically cleared by sequential immersion in a 50 % (v/v) water-glycerin mixture followed by increasing concentrations every 24 h to 75, 85 and 100 %. A bright-field image of the tumor section was then acquired at 2× magnification.
The co-registration of extracted μMRI and μCT vasculature was visually evaluated by overlaying the extracted tumor vasculature obtained from both modalities. Moreover, we compared the co-registered blood volume maps from the two modalities. Fractional blood volume (FBV) maps for μMRI and μCT data were computed by superimposing a spatial grid of 1 mm isotropic voxels on each co-registered dataset and computing the fractional occupancy of vessels within each grid voxel. Linear regression between the μMRI and μCT FBV was used to assess the quality of co-registration and vessel segmentation.
It has been shown that the change in the gradient-echo relaxation rate or ΔR2* (i.e. 1/T2post* – 1/T2pre*) is a measure of the total blood volume . Thus, for each voxel of the in vivo MGE images, the signal intensity versus echo-time curve was fit with a mono-exponential function, which yielded pre- and post-contrast R2* maps. An F-test was used to select only those voxels for which the full exponential model fit the data significantly better (p < 0.05) than the reduced constant model. Next we performed a voxel-wise computation of ΔR2* to create in vivo ΔR2* maps. Voxels with negative ΔR2* were excluded from all analyses.
The ΔR2* maps were converted to FBV maps using an equation derived by Tropres et al. :
where γ is the gyromagnetic ratio of the 1H nucleus, B0 = 9.4 T the external magnetic field strength and Δχ = 0.112 ppm (in cgs units) the difference in magnetic susceptibility of blood before and after contrast agent injection as measured by us earlier .
The measurement of the fractional blood volume of the tumors was compared for the three imaging modalities. The in vivo MRI, μMRI and μCT FBV maps were computed using the same spatial grid to make comparison of FBV maps possible. For the purposes of this study, the high spatial resolution (8 μm isotropic) μCT data were treated as the ‘gold standard’ to which the in vivo MRI and μMRI data were compared. For each imaging modality the FBV values were pooled for PIW3 and PIW5 groups, and quantile–quantile (Q–Q) plots of in vivo MRI and μMRI versus μCT FBV measurements were used to examine the similarity between their distributions. Nineteen quantiles (i.e. the 5th, 10th, 15th … 95th) were computed for the Q–Q plots.
For a representative tumor sample, we characterized differences in vasculature between viable and necrotic tumor regions of interest (ROI). Since necrotic ROI are known to exhibit a higher apparent diffusion coefficient (ADC) than viable tumor tissue , we used an in vivo ADC map to define the necrotic ROI. The ADC map was computed from the in vivo DW MRI data using DTI-Studio (H. Jiang and S. Mori, Radiology Department, Johns Hopkins University, Baltimore, MD, USA). The MR signal intensity depends on the ADC and the b-value according to:
where S and S0 are the MR signal intensities with and without diffusion weighting, respectively. The ADC value along each gradient direction ADCx, ADCy and ADCz was obtained by solving Eq. 2 for ADC for the three DW images acquired with orthogonal diffusion-sensitizing gradients. The final ADC map is a map of the mean diffusivity:
This ADC map was smoothed, and for one 1 mm slice we used C-means clustering to create four classes of ADC values. The class of highest ADC values of (0.43–0.93) × 10−3 mm2/s was used to define the necrotic ROI. Voxels with high ADC values at the tumor edges were excluded.
A nearest-vessel distance map was computed for the 1 mm thick section of the μCT image corresponding to the ADC slice. This was done in Amira® by calculating the Euclidean distance between every μCT voxel within the 1 mm section and the nearest tumor blood vessel. The nearest-vessel distances were then averaged over the z-direction (i.e. the 1 mm thickness) to obtain a mean nearest-vessel distance map, which was overlaid with the necrotic ROI.
Multi-modality vascular data were acquired from ten breast cancer xenografts at complementary spatial scales. However, motion artifact precluded computation of in vivo blood volume maps for one tumor, and two other tumors could not be imaged with μMRI because of sample dehydration and incomplete tumor exsanguination, respectively. Consequently, these tumors were excluded from further analyses.
Figure 2b–j illustrates the outputs of the steps necessary to reliably extract the breast tumor vasculature from the 3D μMRI data. Figure 2b shows the raw μMRI data in which the microfilled blood vessels appear dark. Figure 2c shows the result of the “tubeness” filter on the raw μMRI data wherein the most tube-like structures are brightest. The result of the C-means clustering for the raw data is presented in Fig. 2d. Figure 2e demonstrates that a “tubeness” filter alone was not sufficient for extracting all blood vessels because large caliber tumor blood vessels and those in regions of low contrast-to-noise ratio were not segmented by this method. Figure 2f illustrates how the C-means clustering successfully filled holes in tumor blood vessels larger than ~80 μm in radius from the raw μMRI data. Figure 2g shows how the combination of C-means clustering and “tubeness” filtering permitted extraction of blood vessels with low contrast relative to surrounding tumor tissue. In Fig. 2h, the results of these three approaches are overlaid, indicating that each step detects a different subset of vessels. In Fig. 2i one can see how the morphological 3D “closing” operation is able to fill small holes remaining in the extracted tumor vasculature. Figure 2j illustrates the final extracted 3D tumor vasculature.
The results of the μCT vessel extraction are demonstrated in Fig. 3b–f. The raw μCT data is displayed in Fig. 3b in which the microfilled blood vessels appear bright. Figure 3c shows the result of the tubeness filter on the raw μCT data, which was then subjected to a 3D mean filter and thresholded to obtain the vasculature shown in red in Fig. 3d. A few remaining holes in large caliber tumor blood vessels remained as illustrated in green. The final extracted tumor blood vessels after 3D morphological closing and small region removal are displayed in Fig. 3e. To better visualize the result of this 3D vessel segmentation pipeline, a volume rendering of the vasculature for a 0.5 mm tumor slice is shown in Fig. 3f.
We visually evaluated the extraction of tumor blood vessels from μCT data by directly comparing the resulting vasculature with an optical microscopy image. Figure 4a, b shows that a large number of blood vessels were visible in both the μCT and bright-field image, while vessels smaller than ~8 μm were only visible in the latter. A similar comparison was carried out between the μMRI-derived tumor vasculature and bright-field microscopy. Figure 4d, e shows the concordance between the vasculature extracted from μMRI and a bright-field image of a closely matching 1 mm tumor section.
The co-registered and extracted μMRI and μCT vessels were overlaid as shown in Fig. 4c to permit visual assessment of the co-registration. This overlay illustrates the large degree of overlap between the extracted tumor vasculature from the two datasets and confirms the inter-modality co-registration. In addition to the spatial overlap, we observed dilation of the μMRI-derived blood vessels compared to both optical microscopy and μCT-derived vessels (Fig. 4c–e). Small vessels (<20 μm radius) were only visible in the μCT data and not in the μMRI data (Fig. 4c).
Additionally, we used linear regression for a voxel-wise comparison of the μMRI and μCT-derived FBV (Fig. 4f). The R2 values for the linear regression ranged between 0.60 and 0.87 (p < 0.001) with a mean R2 of 0.72 (n = 8 tumors). The μMRI-derived FBV overestimated the μCT-derived FBV for all tumors.
We used the multi-modality FBV maps to compare the blood volume distribution between the different image contrast mechanisms. Figure 5a–f illustrates the spatial distribution of FBV in a tumor slice from all three imaging modalities. There was good spatial agreement of the μMRI and μCT FBV maps, but absolute FBV values were higher for the μMRI data. The mean μMRI derived FBV was approximately twofold higher for the PIW3 tumors and threefold higher for the PIW5 tumors, compared to the μCT FBV (Fig. 5j). Although in vivo MRI-derived FBV maps showed less spatial agreement with those computed from μMRI and μCT, all three modalities were capable of visualizing the elevated FBV of the tumor rim compared to the center. The mean in vivo MRI FBV values overestimated those derived from μCT by a factor of two (Fig. 5j). Q–Q plots of in vivo MRI and μMRI FBV versus μCT FBV were strongly linear for the PIW3 tumors. This suggests that the shapes of the in vivo MRI and μMRI FBV distributions were similar to that of the μCT FBV distribution (Fig. 5g). Interestingly, the in vivo MRI resembled the μCT FBV distribution more closely than the μMRI did. The slope of the in vivo MRI and μMRI FBV versus μCT FBV Q–Q plot was >1, indicating that the former overestimated the latter by a scaling factor. An identical trend was observed for the Q–Q plots of the PIW5 tumors (Fig. 5h). Finally, Fig. 5i illustrates that the FBV assessed with all three modalities decreased with tumor volume.
An application of our multi-modality imaging approach was the characterization of differences in vascularity between viable and necrotic tumor regions. Figure 6 demonstrates how the necrotic region of a tumor, as identified from in vivo ADC measurements, co-localizes with an area exhibiting large distance from the nearest blood vessels, calculated from the μCT data (Fig. 6c). In addition, an earlier stage (PIW3) tumor with no identifiable necrotic regions exhibited higher ADC values in areas that were at larger distances from the nearest blood vessels (Fig. 6g).
Quantifying angiogenesis-induced changes in the tumor vasculature from medical images is often challenging due to limited contrast-to-noise ratio, image artifacts and spatial resolution limitations. Additionally, each imaging modality has its own unique biophysical contrast mechanism, strengths and weaknesses when it comes to imaging the tumor vasculature as summarized in Table 1. In this work, we exploit the advantages of in vivo MRI and ex vivo μMRI and μCT to develop an integrated platform for characterizing angiogenesis at multiple spatial scales in a human breast cancer model.
Due to its high spatial resolution, μCT is able to resolve blood vessels down to the capillary level . Although the 8 μm isotropic resolution μCT employed in this study cannot resolve the smallest capillaries (~2–3 μm diameter) and angiogenic sprouts, we were able to confirm with optical microscopy that μCT could recapitulate the 3D tumor vasculature down to ~8 μm diameter vessels. Moreover, our μCT data is 3D, covers the whole tumor and tumor vessel radii over several orders of magnitude. Therefore, we regarded μCT as the ‘gold standard’ for 3D blood vessel visualization in whole tissue specimens. Others have also demonstrated the utility of μCT for quantifying tumor angiogenesis down to the capillary level , as well as for validating tumor vascular response to antiangiogenic therapies [24–26]. Moreover, we recently demonstrated μCT’s usefulness as a tool for validating the accuracy of in vivo susceptibility-contrast based MRI biomarkers of breast cancer angiogenesis such as the FBV, vessel size index and vessel density .
In this study, we employed ‘mesoscopic’ scale μMRI to facilitate co-registration between macroscopic in vivo MRI and microscopic μCT angiogenesis data. The intermediate (~40 μm) spatial resolution of μMRI in conjunction with its excellent soft tissue contrast facilitates co-registration and distinguishing tumor tissue from normal tissue for ROI analyses. The spatial resolution of μMRI is limited by the MRI scanner hardware such as the magnetic field and gradient strengths, and by the acquisition time that increases with spatial resolution. In the current study, the intermediate spatial resolution of μMRI did not permit imaging of vessels smaller than 20 μm in radius while the median vessel radius assessed using μCT was ~10 μm . Consequently, partial volume effects led to overestimation of absolute blood vessel size by μMRI, and the μMRI-derived FBV values were approximately twofold to threefold higher than those derived from μCT. However, in spite of this resolution limitation, μMRI-visible vessels greatly enhance the co-registration of in vivo MRI and μCT data to enable multiscale imaging of the tumor vasculature. The strength of μMRI therefore does not lie in resolving small capillaries, but in imaging larger blood vessels while simultaneously providing excellent soft tissue contrast. By exploiting the range of available contrast mechanisms, μMRI can provide detailed information about the tumor tissue that can complement high-resolution vascular images from co-registered μCT data.
In vivo MRI is the only imaging modality used in this study that is translatable to patients for repeated and noninvasive tracking of disease progression and anti-angiogenic treatment response [13, 21]. Subject motion and low spatial resolution make the computation of in vivo angiogenic biomarkers and direct spatial correlation with the μCT data challenging. This was apparent from the co-registered FBV maps of all three modalities wherein the in vivo FBV maps did not exhibit voxel-wise agreement with the ex vivo FBV maps. However, the in vivo FBV maps did show spatial similarities with the ex vivo FBV maps, such as elevated blood volume in the tumor rim compared to the center. In addition, the Q–Q plots validated the usefulness of in vivo FBV as a clinical biomarker for angiogenesis since it resembled the μCT FBV distribution more closely than the μMRI FBV distribution, and was sensitive to the shift in FBV values exhibited by tumors at different stages. The shift towards lower FBV values in PIW5 tumors could be due to larger relative areas of necrosis or lower FBV values in viable tumor regions. Compared to μCT, in vivo MRI overestimated the FBV. A detailed description of the possible biophysical factors (e.g. contrast agent dose, magnetic field strength etc.) contributing to this overestimation can be found in . An extensive comparison of FBV, vessel size and vessel density for μCT and in vivo MRI in that study showed a similar overestimation of blood vessel size. Finally, that study conclusively demonstrated that relative MRI derived parameters (i.e. those based only on ΔR2* and ΔR2 measurements) may be better biomarkers of angiogenesis than absolute MRI derived measures (e.g. FBV, which requires knowledge of the intravascular susceptibility difference and yields an absolute blood volume with the appropriate unit). In addition, others and we have previously shown that there is a significant correlation between several in vivo MRI angiogenesis biomarkers and their μCT analogs [21, 27].
Our multi-modality visualization of differences in tumor vascularity between viable and necrotic tumor tissue takes advantage of the high spatial resolution provided by μCT and the sensitivity of diffusion-weighted MRI to the underlying tissue cytoarchitecture. The integration of these complementary physiologic data helps us understand the relationship between tumor vascularization and development of necrosis. In our breast cancer model, necrotic tumor areas co-localized with poorly vascularized areas. Moreover, the distance of voxels in the necrotic areas to the nearest blood vessel often exceeded ~200 μm, which is the typical diffusion distance of oxygen in tissue . In addition, smaller tumor regions with higher ADC values tended to exhibit larger distances to the nearest vessel, suggesting that necrotic areas were probably developing because of diffusion-limited hypoxia.
Due to the range of available MRI contrast mechanisms, one can foresee several potential applications of this novel imaging platform. For instance, one could relate in vivo measures of vascular function to the tumor vessel architecture by overlaying maps of Ktrans (i.e. the volume transfer constant between blood plasma and extracellular, extravascular space) obtained from dynamic contrast enhanced (DCE)-MRI with the well-resolved μCT vasculature. This would facilitate interpretation of Ktrans maps that depend on both blood flow and vessel permeability . We recently demonstrated that the high resolution μCT-derived vasculature can be used in more realistic computational models of tumor blood flow  and in ‘bioimage informatics’ approaches for simulating tumor hemodynamics . In conjunction with histology, these developments help bridge the genotype to phenotype divide and are a major step towards creating tumor vasculature-based ‘digital atlases’ for cancer systems biology applications.
More recently, investigators have begun to employ realistic vascular geometries to elucidate the biophysics underlying the different MRI contrast mechanisms  as well as improve the accuracy of clinical biomarkers of tumor angiogenesis . For example, Pannetier et al. developed a computational tool for simulating DCE-MRI experiments and validated it by testing it on a 2D mouse brain cortical vascular network acquired using two-photon laser scanning microscopy [34, 35]. We recently extended this approach by incorporating the ‘whole-tumor’ 3D vasculature in biophysical models of dynamic  and steady-state susceptibility contrast (DSC or SSC) MRI .
The multi-modality nature of the present study necessitated the development of suitable vessel segmentation techniques for the 3D μCT and μMRI data, since each imaging method has its own unique contrast mechanism . For optimal results, our technique incorporated vessel enhancement , intensity thresholding and binary morphological operations. Although in some cases, simple intensity thresholding could be applied to isolate vascular structures from the background tumor tissue, we demonstrated that the quality of vessel segmentation (i.e. vessel completeness, elimination of ‘holes’ and isolated voxels) was greatly improved by additional pre- and post-processing steps. One limitation of our vessel segmentation procedure was that it involved the manual determination of segmentation thresholds. Although this made our procedure operator dependent, it resulted in higher-fidelity tumor vasculature than automatic thresholding alone. Both correlation analysis and visual inspection demonstrated that high-quality co-registration was achievable between the μMRI and μCT-derived tumor vasculature. While it would have been ideal to conduct a vessel-to-vessel co-registration for the whole tumor, even the slightest misalignment between the two vascular structures would affect the computed correlation coefficient. Since sample embedding and fixation for μMRI and μCT imaging can introduce tissue deformation as discussed in , tissue distortions on the order of the μMRI spatial resolution of ~40 μm would adversely affect the correlation at the original spatial resolution. Collectively, these data demonstrate that although not without limitations, our multiscale multi-modality imaging platform is a valuable tool for studying tumor biology at multiple spatial scales. In addition, although it necessitates the availability of different imaging modalities, the integration of multiscale, multidimensional data is a powerful platform for visualizing and integrating angiogenesis-related data in pre-clinical cancer models.
Supported by a Susan G. Komen for the Cure Career Catalyst Grant (KG090640) and a Bayer Fellowship to J.C. We are also grateful to Dr. Zaver M. Bhujwalla, Director of the Division of Cancer Imaging Research, JHU ICMIC program and Cancer Functional Imaging Core for useful discussions and support (P50CA103175 and P30CA006973).
Jana Cebulla, Department of Circulation and Medical Imaging, Norwegian University of Science and Technology, Trondheim, Norway.
Eugene Kim, Department of Biomedical Engineering, The Johns Hopkins University School of Medicine, Baltimore, MD, USA.
Kevin Rhie, Division of Cancer Imaging Research, Russell H. Morgan Department of Radiology and Radiological Science, The Johns Hopkins University School of Medicine, 720 Rutland Ave, 217 Traylor Bldg., Baltimore, MD 21205, USA.
Jiangyang Zhang, Division of Cancer Imaging Research, Russell H. Morgan Department of Radiology and Radiological Science, The Johns Hopkins University School of Medicine, 720 Rutland Ave, 217 Traylor Bldg., Baltimore, MD 21205, USA.
Arvind P. Pathak, Division of Cancer Imaging Research, Russell H. Morgan Department of Radiology and Radiological Science, The Johns Hopkins University School of Medicine, 720 Rutland Ave, 217 Traylor Bldg., Baltimore, MD 21205, USA. Sidney Kimmel Comprehensive Cancer Center, The Johns Hopkins University School of Medicine, Baltimore, MD, USA.