HiSStology: High Spectral and Spatial Resolution Magnetic Resonance Imaging Detection of Vasculature Validated by Histology and Micro–Computed Tomography

Abstract

High spectral and spatial resolution (HiSS) data, acquired with echo-planar spectroscopic imaging (EPSI), can be used to acquire water spectra from each small image voxel. These images are sensitive to changes in local susceptibility caused by superparamagnetic iron oxide particles (SPIO); therefore, we hypothesized that images derived from HiSS data are very sensitive to tumor neovasculature following injection of SPIO. Accurate image registration was used to validate HiSS detection of neovasculature with histology and micro–computed tomographic (microCT) angiography. Athymic nude mice and Copenhagen rats were inoculated with Dunning AT6.1 prostate tumor cells in the right hind limb. The tumor region was imaged pre– and post–intravenous injection of SPIO. Three-dimensional assemblies of the CD31-stained histologic slices of the mouse legs and the microCT images of the rat vascular casts were registered with EPSI. The average distance between HiSS-predicted regions of high vascular density on magnetic resonance imaging and CD31-stained regions on histology was 200 µm. Similarly, vessels identified by HiSS in the rat images coincided with vasculature in the registered microCT image. The data demonstrate a strong correlation between tumor vasculature identified using HiSS and two gold standards: histology and microCT angiography.

Angiogenesis has been linked to tumor aggressiveness^1–4 and, by extension, tumor hypoxia.⁵ As early as 1965, thalidomide was used to treat cancer patients.⁶ Since then, many antivascular or antiangiogenic agents have been used in both clinical and preclinical studies.⁷ Therefore, the ability to monitor angiogenesis, especially in vivo, is of great interest for cancer therapy. Clinically, computed tomography (CT) and magnetic resonance imaging (MRI) are often used with contrast agents to assess response to antivascular therapy.

Two common MRI-based methods for assessing vasculature both use gadolinium (Gd), namely magnetic resonance angiography (MRA) and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). In addition to concerns with nephrotoxicity, DCE-MRI suffers from poor spatial resolution and therefore is useful only as a functional imaging tool. DCE-MRI might not be sensitive to early or subtle vascular changes. MRA without contrast agents, for example, time-of-flight MRA, suffers from lower signal to noise ratio and flow artifacts.⁸ Susceptibility or T₂*-weighted imaging is another MRI technique used for examining vasculature.^9–11 CT-based angiography may limit longitudinal studies owing to fear of radiation exposure, especially with older scanners. In general, CT does not image soft tissue as well as MRI, so soft tissue with low contrast medium uptake may not be well defined, for example, a necrotic core of a tumor.

Despite the limitations of low-molecular-weight Gd-based contrast agents, they are easily detectable as they leak into the extracellular-extravascular space readily. However, low-molecular-weight Gd-based contrast agents can underestimate the effects of antiangiogenic treatments.^12,13 Blood-pool agents have been shown to be more effective in detecting changes in vasculature after antiangiogenic treatment.^12,14,15 However, blood-pool agents produce smaller changes in signal intensity than low-molecular-weight agents, and increased sensitivity is desirable when these agents are used.

We propose that magnetic resonance spectroscopic imaging of the water resonance could increase sensitivity to the effects of blood-pool agents. Echo-planar spectroscopic imaging (EPSI) can be used to acquire water and fat spectra from each small image voxel.^16,17 High spectral and spatial resolution (HiSS) data can be constructed from various features of the water lines.^18–20 HiSS MRI provides an extra dimension of information (not available with conventional MRI) that can be used to identify blood vessels and contrast agent effects. Because HiSS images are derived from the complete free induction decay (FID), artifacts common to standard T₂*-weighted imaging are avoided or filtered out. For example, chemical shift artifacts from fat can often degrade conventional images.^21,22 The morphology of breast images has been shown with more detail using HiSS MRI.^23,24 (For a detailed background on EPSI, see Mulkern and Panych and Tal and Frydman.^25,26) Note that EPSI is typically performed at low spatial resolution and the acronym HiSS is preferred (to describe the protocol used here) as it emphasizes the use of both high spectral and high spatial resolutions.

HiSS MRI is extremely sensitive to changes in local susceptibility, for example, changes caused by deoxyhemoglobin or superparamagnetic iron oxide particles (SPIO).²⁷ Therefore, we proposed that images derived from HiSS data are very sensitive to tumor vasculature following injection of a blood-pool contrast agent. To test this hypothesis, we developed accurate image registration methods that allow precise correlation of MRIs, CD31-stained tissue slices, and microCT images of tumor vasculature. Note that vessel structural information can be obtained at a scale below the actual image resolution, that is, like some functional images.

We also demonstrate a new approach to registration of histology and tomographic images (MRI in this case). Histology is often used as a gold standard to assess the effect of an antiangiogenic agent in mouse models. It is desirable to have a system that registers multiple histologic slices with in vivo functional and anatomic images. Furthermore, the histologic slices require some degree of warping to account for tearing, shrinking, and so on. To date, many of the published image histology registration methods are missing one of these requirements or have other limitations. For example, Humm and colleagues have reported an elegant fiducial-based histology and MRI registration method.²⁸ However, they inserted polyethylene tubes into the sample as fiducial markers. One limitation would be the possibility of inserting the fiducial into an area of interest. It would also seem that exchanging the fiducials accurately for multiple modalities could be difficult (eg, exchanging water [MRI] fiducials for positron emission tomographic [PET] fiducials). Similarly, Breen and colleagues inked tissue samples prior to histologic preparation and then inserted pins into the ink marks to visualize them better.^29,30 They demonstrated a robust warping and registration system, but, again, for a small sample, the ink or pins could disturb an interesting region of the sample. Furthermore, the method does not assist in registration with multimodality in vivo images.

In many of the published methods, the serial capability is not apparent; that is, one cannot likely adjust a single histology slice and have that change propagate through all of the transformations to the functional image. In addition, the non–fiducial-based systems would have difficulty registering tissue sections with a low-resolution functional image. Also, some of the image registration/immobilization systems do not lend themselves to slicing for histology owing to geometry (whole body) or the materials used.^28,31 Therefore, we developed a cast immobilization/registration system that facilitates multimodality image registration. A rat prostate xenograft model was used in athymic nude mice to validate HiSS neovascular identification with immunohistochemistry. Using a syngeneic rat prostate model, vascular casts imaged with microCT were compared with HiSS images.

Materials and Methods

Animal Protocol

Athymic nude mice (n = 4) and Copenhagen rats (n = 4) (Charles River, Wilmington, MA, and Harlan Laboratories, Inc., Indianapolis, IN, respectively) were inoculated with Dunning AT6.1 prostate tumor cells³² in the right hindlimb. Mice were used for histologic comparison as their size allowed the use of the whole leg. Conversely, rats were used for the microCT vascular cast comparison because their larger femoral arteries and veins were easier to work with. All animals were anesthetized during all procedures with 1.5 to 3% isoflurane mixed with medical air to maintain a surgical plane of anesthesia. Heart and respiratory rates, along with temperature, were continuously monitored using a computerized system (SA Instruments, Inc., Stony Brook, NY). A standard 24-gauge intravenous (IV) catheter was inserted into the tail vein for the injection of contrast agent (300 µL SPIO, 1 µm diameter, tagged with fluorescent Dragon Green, in 1.4 mL saline, Bangs Laboratories, Fishers, IN). The temperature of the animals was maintained at 37°C during imaging procedures via computer-controlled warm air.

All animal procedures were performed under Institutional Animal Care and Use Committee–approved protocols.

Magnetic Resonance Imaging

MRIs were acquired at 4.7 T using a 30 cm diameter bore Bruker MRI Scanner (Bruker BioSpin, Billerica, MA) equipped with self-shielded imaging gradients. Custom eight-leg low-pass, volume birdcage coils designed to couple with the tumor-bearing leg (16 mm diameter for mice and 32 mm diameter for rats) were used. For image registration and anatomic guidance, the multislice, rapid acquisition with relaxation enhancement (RARE) spin-echo sequence was used (repetition time [TR] = 4,100 ms, effective echo time [TE] = 56 ms, field of view [FOV] = 3.5 cm [mice] and 5.0 cm [rats], matrix size = 256 × 256, slice thickness = 0.6 mm, number of excitations [NEX] = 2, RARE factor = 8). Voxel sizes for the mice and rat images were 0.14 × 0.14 × 0.6 mm and 0.20 × 0.20 × 0.6 mm, respectively. Mouse images were acquired in planes roughly orthogonal to the long axis of the leg, approximating a transaxial plane to correspond with the orientation used for histology. Rat images were acquired in sagittal planes (parallel to the long axis of the leg). To improve shimming to higher orders, Fastmap³³ was used with a cubic voxel centered within the leg and encompassing the whole tumor volume. For HiSS, three to four slices were selected from the anatomic images using the same image geometry. To minimize motion artifacts, respiratory gating was used. Therefore, the approximate TR was 900 ms (between 800 and 1,200 ms, determined by the respiratory rate) (FOV = 3.5 cm [mice] and 5.0 cm [rats], slice thickness = 1 mm, matrix size = 256 × 256 × 64; readout steps, phase encoding steps, time points for each point in kspace, 5.12 ms echo spacing, flip angle = 30°, NEX = 2, spectral resolution = 3.1 Hz, 50 kHz receiver bandwidth, scan time ≈9 min). EPSI collects the decay in the magnetic resonance signal at each point in k-space as a function of time by use of an oscillating (bipolar) readout gradient train (see Figure 1, EPSI pulse sequence, in Foxley and colleagues³⁴). This is equivalent to collecting the FID for each voxel in real space. The Fourier transform of the FID produces a spectrum for each small image voxel. HiSS images were reconstructed using code written in-house using Interactive Data Language (IDL, ITT Visual Information Solutions, Boulder, CO), reported previously.¹⁹ Only water peak-height (PH) images were used in this study, although the full proton spectrum was collected. Water PH images were constructed from the HiSS data in which the signal intensity at each pixel in the two-dimensional matrix was proportional to the maximum signal intensity in the corresponding voxel’s water resonance. Postcontrast images were subtracted from precontrast images, and these subtraction images were correlated with histology and microCT. A mask was made from the subtraction images with a threshold, 4 to 30%, with spikes removed by means of the histogram. This apparently large range of thresholds was likely due to the geometry of vessels as the blooming artifact of SPIO is influenced by the geometry of the vessels. This mask was taken as the in vivo MRI estimate of blood vessel locations. The results reported are the mean over four experiments for each species (four mice and four rats).

Figure 1.

Example of 3 mm cast slices (above) with corresponding MRI slices (below) of the postmortem leg.

Histology

Mice were sacrificed and tumor-bearing legs were removed and fixed in formalin to correlate in vivo MRI and histology. After fixation, legs were embedded in vinyl polysiloxane (VPS) (Regular Type, Examix NDS, GC America, Alsip, IL) dental impression material. MRI-detectable fiducial markers were embedded in the dental cast (Figure 1 and Figure 2). The VPS cast was made using a custom jig that was machined to precisely match both the 16 mm birdcage coil used for MRI of the mice and the slicing device described below. The legs embedded in dental casts and the associated fiducials were reimaged with MRI (for details about the immobilizing cast, see Haney and colleagues³⁵). The postmortem images served as a link between in vivo images and histologic slices. A custom slicing device was fabricated so that precise 3 mm slices of the immobilized leg could be made. The whole leg was cut into five or six 3 mm thick slices, and both faces of each slice were scanned optically using a standard desktop scanner. The digital images of the cut pieces were cropped and aligned manually using Adobe Photoshop v7.0.1 (see Figure 1 and Figure 2). The aligned slices were assembled into a three-dimensional volume in MATLAB (The MathWorks, Natick, MA) (see Figure 2). This volume allowed translation between histologic image coordinates and postmortem MRI coordinates. The fiducial markers leave holes in the cast, and these holes were used to align each 3 mm slice and register the fiducials visible on postmortem MRI (see Figure 1 and Figure 2). Each 3 mm slice was processed for hematoxylin-eosin staining and staining for platelet/endothelial cell adhesion molecule 1 (PECAM-1, also referred to as CD31) receptors on alternating 5 µm thick slices. The location of each 5 µm slice within the 3 mm slabs, described above, was known. The optical scans facilitated assembly of the three-dimensional histologic image from the 5 µm slices. Histology slices in the three-dimensional volume were aligned using Stack Reg, an ImageJ plug-in based on reference 36.³⁶ Also, fluorescence microscopy images were made in vitro on the 3 mm thick slice of tissue prior to embedding in paraffin for microtome slicing because the histologic processing removed the fluorescence.

Figure 2.

Left: Three-millimeter cast slices assembled into a three-dimensional volume (green) registered with a postmortem spin-echo MRI of the mouse leg (red). Right: Five-micrometer thick histologic slides assembled into a three-dimensional volume (green) and registered with the spin-echo image (red). The histologic slices are located accurately in the three-dimensional volume of the tumor and leg. The gap in the slices is due to sparse sampling of histology slides where the tumor was absent.

The full-resolution images of CD31 slides (necessary for identifying blood vessels) were typically 100 MB at 25 µm resolution (scanned with a Clarient Automated Cellular Imaging System, now part of Carl Zeiss, Inc). About 10 to 20 slides were used to make a three-dimensional histologic image. The full-resolution images were resized (typically down to 300 × 300 pixels) to be manageable for three-dimensional image registration. Once the transformation from in vivo MRI to histology was determined, a slice of interest could be displayed at full resolution (Figure 3).

Figure 3.

CD31-stained axial section of a mouse leg with an AT6.1 tumor. The blue contour denotes the perimeter of a registered spin-echo image. The red contour denotes changes in signal (above the threshold) in a registered EPSI image after injection of Bangs particles. The yellow contour denotes the tumor boundary defined from the spin-echo image. All contours have been warped to account for distortions in the histology. Inset: An EPSI image (postcontrast image subtracted from precontrast image) of approximately the same slice. The green arrow denotes the large blood vessel in Figure 7A. The yellow arrow denotes dense vasculature in Figure 7B.

A blood vessel mask was made for each CD31-stained image by converting the image to CIE L*a*b* color space and automatically matching pixels stained for CD31, resulting in a binarized image. A euclidean distance transform was used to measure the distance from each pixel in the histology vessel mask to its nearest neighbor in the HiSS MRI vessel mask using the Image Processing Toolbox in MATLAB. The mean distance, sensitivity, and specificity were calculated with the criterion that a distance less than twice the standard deviation would be considered a true positive. Data are reported as the mean ± standard deviation.

MicroCT

Rats were prepared for microCT as follows. The right femoral artery and vein were cannulated for infusion of 10% barium sulfate (BaSO₄) (Maxibar, E-Z-EM Canada, Westbury, NY) and withdrawal of blood, respectively. BaSO₄ was injected until the vasculature at the skin surface was visibly almost white (typically 1–1.5 mL). Each rat was sacrificed after BaSO₄ injection and was kept in the same orientation as it was during MRI in the refrigerator for 24 hours. The leg was then removed from the rat and imaged with microCT. It was important to ensure that rigor mortis occurred while the leg was at the same articulation as during the in vivo MRI acquisition to obtain good image registration.

A Faxitron MX 20 Specimen Radiography System (Faxitron X-ray Corp., Wheeling, IL) was retrofitted with a sample rotation stage for use as a microCT with a resolution of 100 to 150 µm. Exposure time was typically 4.7 seconds per projection at 30 mA and 27 kVp. However, owing to image communication and storage overhead, the total image acquisition time was ≈2 hours for 360 projections. Images were reconstructed on a 512 × 512 × 300 matrix using a Feldkamp Davis Kress approximate cone beam algorithm, written in IDL. Unlike the histology images, the microCT images could be used at nearly full resolution, without compromising detection of the blood vessels by downsizing. This made voxel-based analysis possible. For each EPSI voxel, the distance to the nearest microCT blood vessel was measured using a distance volume transformation on the registered images. The mean distance, sensitivity, and specificity were calculated with the criterion that a distance less than twice the standard deviation would be considered a true positive. Data are reported as the mean ± standard deviation.

Image Registration

Image registration was performed using in-house software written with MATLAB. Matching features were used to manually translate and rotate one image relative to the other. For histologic comparisons, the order of transformations was as follows: in vivo MRI to postmortem MRI, then to the 3 mm optical scans of sliced cast/leg, and then finally to 5 µm thick histology slices (see the diagram in Figure 4). During the assembly of the 5 µm histology three-dimensional volume, the face of each 3 mm thick slice was used. Therefore, the three-dimensional histology volume could use the same transformation as that of the 3-mm three-dimensional volume registered to postmortem MRI. The microCT comparison was more straightforward; in vivo MRI was manually registered (rigid body) directly to postmortem microCT, with the bone being a major landmark. Although there were slight differences in articulation between the rat leg in MRI and postmortem microCT, warping was not used.

Figure 4.

Diagram outlining the procedure to register histology with in vivo MRI. The rightmost column contains three three-dimensional images resulting from each stage of the process. The registration of ex vivo microCT to in vivo MRI was done in one step, described in the Methods. H&E = hematoxylin and eosin.

Because of tearing and nonuniform shrinking of tissue, a method of warping the histologic images had to be developed. Again, MATLAB was used to create a program that allows the user to move points on registered regions of interest from postmortem MRI to make them match the corresponding distorted histologic slice. B-splines were used to warp the images, based on the user’s placement of the “control points.”

To evaluate how well the images matched, the Jaccard index was calculated. This is a standard metric obtained by dividing the intersection of the histology and in vivo MRI images by the union of the binarized images. The histologic processing degrades the fidelity of the sample through the fixing and cutting procedures. However, the integrity of the microCT images was not compromised. Therefore, the bone was a very reliable landmark for MRI/microCT registration. The Jaccard index is less appropriate for MRI/microCT registration because the bone was used as the main guide for registration and its small size would negatively bias the comparison of how well the binary images match. Because the Jaccard index is essentially a percentage, bone-to-bone comparisons could show a large percentage mismatch when the absolute deviation is less than half a millimeter. Therefore, the mean distance between the bone in microCT and in vivo MRI was measured.

Results

Image registration methods were developed and used to show that regions containing high vascular density on in vivo MRI were reliably identified based on changes in HiSS images following contrast medium injection. There was strong agreement between gold standards (histology and microCT) with HiSS images regarding the location of dense vasculature. Examples of HiSS images before and after injection of SPIO are shown in Figure 5. The red arrows locate regions of negative contrast in the subtraction images (postcontrast image minus precontrast image), which correspond to dense vasculature.

Figure 5.

Example of pre– (A and D) and post– (B and E) intravenous injection and subtraction images (C and F). Top row: Mouse leg, axial slice; the tumor is denoted by the yellow region of interest (ROI). The tibia is in the upper-left quadrant. Bottom row: Rat leg, sagittal slice; the tumor is denoted by the yellow ROI. The whole rim of the tumor shows a decrease in contrast, especially at the tumor/muscle boundary. The red arrows indicate areas of dense vasculature. The color bars are percent difference.

Image Registration

For the middle three slices of the tumor-bearing portion of the leg, where the tissue damage was minimal and where the MRIs provided the best coverage, the Jaccard index was 79% with a standard deviation of 10%. The mean distance between the center of tibia in the CT images and in vivo MRI images was 0.74 ± 0.22 mm. However, for the slices of interest (ie, slices containing most of the tumor), the mean distance was 0.30 ± 0.11 mm, which translates to about 1.3 pixels. Differences in articulation and changes postmortem were the main issues influencing registration accuracy between microCT and in vivo MRI.

Correlation of HiSS MRI with Histology

CD31 staining provides very high resolution and high sensitivity for blood vessels. An example is shown in Figure 3. The blood vessels are brown/black with CD31 staining, which can be seen in Figure 6. Note a blood vessel or vessels along the edge of a lobe of tumor (yellow arrow). However, there is the potential for discrepancy between the histology and in vivo MRI as CD31 also stains the endothelium of non- or poorly perfused tumor vessels. This problem was minimized by avoiding regions that were necrotic based on hematoxylin-eosin slides. Vasculature foci at least 50 microns in diameter were identified on CD31-stained tissue slices and used as a gold standard. The mean distance between all pixels identified as vessels in the CD31 images and any pixels identified as vessels on HiSS water PH images was 0.74 ± 1.1 mm with a sensitivity of 74.8 ± 5.4% and a specificity of 75.4 ± 12.3%. The positive predictive value was 79.4 ± 9.3%, and the negative predictive value was 63.7 ± 11.3%. However, the average distance (in two dimensions) between regions of high vascular density on in vivo MRI and CD31-stained regions on the 5 µm thick histologic slides was 200 µm, indicating excellent spatial correlation.

Figure 6.

Four times magnification of the slide in Figure 3 (denoted by the black box region of interest). A and B, Hematoxylin-eosin- and CD31-stained slides, respectively. Inset: CD31 with a fluorescent Bangs particle image (see Figure 7) overlaid, taken approximately from the region marked by the yellow arrows in this figure and Figure 3.

Correlation of HiSS MRI with Fluorescence Image

Unfortunately, image registration and the signal to noise ratio of the fluorescence images were not adequate to provide a strong test of MRI sensitivity and specificity. This was due to the histologic processing, which either cleaved or quenched the fluorescent tag. In spite of this difficulty, Figure 3, Figure 6, and Figure 7 show that the Bangs particles are detected at the tumor rim (high vascular region) in both HiSS and fluorescence images. Following contrast medium injection, HiSS shows a maximum change in signal intensity where the vasculature is dense. The green arrow in Figure 3 (CD31-stained histologic slice with registered in vivo MRI contours) points to a large blood vessel with many Bangs particles seen in the fluorescence image (see Figure 7).

Figure 7.

Fluorescent Bangs particles at 10× original magnification seen in a 3 mm thick tissue slice prior to histologic processing. A, Large blood vessel seen in Figure 3. The black arrow marks muscle tissue, and the red arrows denote the dense vascular patch that is shown by MRI in Figure 3. B, Region within the tumor shown in Figure 3 shows fluorescent Bangs particles at the location where MRI shows high vascular density. C and D, CD31 of the same region as A and B, respectively.

MicroCT

Regions of high vascular density were selected using a 26 to 31% threshold on CT, and this was used as the gold standard. Blood vessels were identified in HiSS data sets based on the change in water PH following contrast medium injection. The mean distance between voxels containing blood vessels in CT and the nearest EPSI voxel with blood vessels was 1.34 ± 0.45 mm (n = 4). The sensitivity and specificity for detection of blood vessels by in vivo MRI were 77.1 ± 10.1% and 77.3 ± 6.4%, respectively. The positive predictive value was 51.3 ± 10.8%, and the negative predictive value was 90.6 ± 6.5%.

An example of a microCT of a rat leg vascular cast is shown in Figure 8; ROI from the registered in vivo MRI are shown on the right. Note that some large vessels are at the edge of the in vivo MRI (cyan contour) owing to the FOV used. Even though these vessels were outside the optimally shimmed area, a change in susceptibility following contrast medium injection was detected with excellent signal to noise, as evidenced by the relatively small (red) contour.

Figure 8.

Left: Example of a rat leg microCT (vascular cast). Right: Same image as on the left, with contours from registered MRI. The red contour denotes changes in signal (above the threshold) in the registered EPSI image after injection of SPIO. The cyan contour denotes the tumor boundary defined from the spin-echo image. All contours did not require warping.

Discussion and Conclusions

Using image registration, the gold standards histopathology and microCT angiography validated the use of HiSS MRI to detect dense foci of microvessels. The use of an SPIO with HiSS as a blood-pool agent, detected by HiSS MRI, showed good agreement with the gold standards.

We recently demonstrated that off-resonance components by themselves can be used to identify tumor vasculature, especially at the rim of the tumor.¹⁹ Because of the complexity of the Fourier component images (FCIs) used to identify vasculature, we took a more straight-forward approach and used a blood-pool agent. Note that the same precontrast images used in the rat images of this study were the same as used in the FCI study.¹⁹ Therefore, the FCI results could be validated with the registered microCT data shown here. As concern grows regarding the use of contrast agents, identifying blood vessels with FCI (without the use of contrast agents) will become more valuable.

There were two major limitations for this study that future studies will address. First, there were difficulties in processing the samples for histology, with the need for good registration in mind. The sample typically shrinks inhomogeneously when using formalin fixing, and the tissue can be torn during sectioning. For example, material lost from cutting created some ambiguity as to the starting position of each histologic slice in its three-dimensional stack.

The second major limitation was the lack of a “turnkey” microCT for small-animal imaging. Images for the present study were acquired using home-built equipment, and this complicated data acquisition and processing. In addition, it was not possible to image the whole tumor or to image a live animal. Future CT imaging will be performed using a more convenient commercial microCT unit, and this will improve image quality, simplify the measurements, and allow us to image larger volumes of tissue. The possibility of repeated imaging, especially after an intervention (eg, an antivascular agent), will enhance future studies.

Improvements in the registration software to make it more automated are in progress, along with more advanced vessel identification algorithms. Although other work from this laboratory used an immobilizing cast in vivo to improve registration,³⁷ the cast was not used here in vivo because of concerns that paramagnetic impurities in the cast would degrade spectral integrity. In future work, we plan to use a cast that has no significant effect on water spectra.

Recall that the SPIO had a fluorescent tag. Fluorescence measurements were made in vitro on the 3 mm thick slice of tissue prior to embedding in paraffin for microtome slicing. Therefore, the depth of the Bangs particles is difficult to assess. The thick slice prevents focusing all of the Bangs particles; therefore, the signal to noise ratio and spatial resolution were reduced. Unfortunately, image registration and the signal to noise ratio of the measurements of the fluorescence images were not adequate to provide a strong test of MRI sensitivity and specificity. This was due to the histologic processing, which either cleaved or quenched the fluorescent tag.

In spite of these limitations, the data demonstrate strong correlation between tumor vasculature on in vivo MRI and the gold standards: histology and microCT. This suggests that HiSS data are highly sensitive to tumor vasculature. Therefore, clinical and preclinical HiSS MRI could be used serially to track changes in vasculature, for example, before and after treatment with an antiangiogenic agent. The image registration methods (especially the mechanical aspects) described here have broad applications. They could be used, for example, to correlate PET images with immunohistochemical staining for a variety of important biomarkers, such as vascular endothelial growth factor, Hoechst, and pimonidazole.