2.1.

Cell Lines

The two cell lines investigated in this study were the rat 9L gliosarcoma cell line transfected with green fluorescent protein (GFP) and the human glioma cell line U251. Cells were grown in Dulbecco’s Modified Eagle’s Medium (DMEM; Mediatech, Inc., Manassas, Virginia, Cat. # 10-013-CV) supplemented with penicillin $(100\text{units}∕\mathrm{ml})$ -streptomycin, $100\mu \mathrm{g}∕\mathrm{ml}$ (HyClone, Logan, Utah, Cat. # SV30010). In vitro studies of the tumor cells have shown that U251 cells have a 20-fold higher expression of EGFR than 9L cells.²⁷ In this study, U251 tumors that overexpress EGFR are referred to as EGFR(+) tumors, while the negative control 9L line tumors are termed EGFR(-) tumors. Cells were grown to 80% confluency in culture, trypsinized, and brought into solution in phosphate buffered saline (PBS) at $5\times {10}^7\text{cells}∕\mathrm{ml}$ in preparation for injection.

2.2.

Animal Preparation

All procedures using animals were conducted under protocols approved by the Institutional Animal Care and Use Committee (IACUC) at Dartmouth. Animal subject preparations for the nude mice used in this study have been described previously⁷ and are summarized here. Mice were $6\text{to}7\text{week}$ -old male athymic NCr-nu/nu nude mice (strain 01B74) purchased through the NIH Animal Procurement Program. All mice underwent intracranial surgery prior to imaging. During the surgical procedure, the skin on the skull was prepared using betadine. Five $\mu \mathrm{l}$ of the cell suspension were injected slowly via a $1\text{-}\mathrm{mm}$ burr hole in the skull with a $25\text{-}\mathrm{ga}$ needle at $2\mathrm{mm}$ anterior and $3\mathrm{mm}$ to the left of the bregma. Control mice underwent sham-surgery and were injected with PBS only. Tumors grew for $14\text{to}23\text{days}$ , and each tumor-bearing animal was imaged with gadolinium-enhanced MRI (Gd-MRI) at least two days prior to the MRI-FMT scans. Only animals with visible Gd-enhanced features in the brain were included in the study. The final distribution of animals used in the study was six mice with U251 tumors, five with 9L tumors, and four in the control group. Forty-eight hours prior to MRI-FMT scanning, each animal was administered 1 nmole of Licor IRDye 800CW EGF (Licor Biosciences, Lincoln, Nebraska) reconstituted per manufacturer’s directions into the tail vein. This imaging agent consists of an NIR fluorescence dye conjugated to the EGF ligand, which has high affinity for the associated EGFR.

2.3.

MRI-FMT Imaging System

The hybrid MRI-optical system has been detailed in previous publications,^{7, 17, 18} and a brief overview is included here. A diagram of the system is provided in Fig. 1 , which illustrates the spectroscopic FMT system integrated into a Philips $3\mathrm{T}$ clinical MRI. This integrated platform allows simultaneous optical and MR scanning that facilitates straightforward integration of data between the modalities. The optical detection component consists of eight spectrometers, each with cooled imaging CCD sensors, which are coupled to the tissue surface through long optical fiber bundles. A specialized rodent MRI coil, shown in Fig. 1, positions the optical fibers in a circular ring around the animal’s head. The bifurcated branch of the fibers was used to illuminate each channel sequentially with a $690\text{-}\mathrm{nm}$ laser diode, while the remaining seven channels act as light pickups to the spectrometers. Both fluorescence emission and excitation spectra were measured for each source–detector pair. In this configuration, a total of 56 emission and 56 excitation measurements (eight sources by seven detectors) were acquired for each animal.

Fig. 1

Diagram of the MRI-FMT scanner illustrates the contact-mode configuration of eight optical fibers around the head (a). The fiber array installed in the RF coil is shown in (b). A close-up 3-D rendering of the head (blue) and brain (yellow) of one of the animals is shown in (c). The red spheres mark the location of the eight spectroscopy fibers encircling the head in a single plane. A T1-weighted coronal MR image corresponding to this fiber plane is shown in (d), overlain with blue and yellow regions to emphasize the location of the brain. Once this plane is selected for a given animal, the image is segmented into tissue regions broadly defined as brain, abnormal structures in the brain, and the region outside the brain. Examples of segmented regions for a mouse with a U251 tumor, a 9L tumor, and no tumor are shown in (e), (f), and (g), respectively. (Color online only.)

During MRI-FMT scanning, an anesthetized (1.5% isoflurane, $1\mathrm{L}∕\min$ oxygen) animal was positioned in the specialized rodent coil and the fiber terminals moved inward to contact the tissue. Scout MRI scans were used to locate the plane of the fibers with respect to the brain and/or tumor (if visible pre-Gd contrast). This procedure was repeated until the animal alignment was satisfactory. Once the animal was positioned, both optical and MRI scans were initiated. The MRI acquisition protocol included a T1-weighted turbo-spin-echo (TSE) pre-Gd contrast followed by T2-weighted TSE, intraperitoneal gadolinium-DTPA (Magnevist, Bayer Healthcare Pharmaceuticals, Wayne, New Jersey) administration $\left(0.03\mathrm{ml}\right)$ , and finally another T1-weighted TSE post-Gd contrast. Besides scout scans, all MRI image slices were in the coronal direction. The optical measurements were completed within the $23\min$ required for the MRI sequences, and the entire procedure including anesthetization could be completed in about $30\min$ .

Figure 1 shows a portion of a mouse head rendered from an MR image stack, with the brain shown as a rendered surface in yellow. The red spheres surrounding the head lie in a coronal plane and mark the location of the optical fibers. The 2-D image slice corresponding to this plane, shown in Fig. 1, represents the MRI-FMT image plane. With the exception of control mice, only animals with Gd contrast-enhancement in this plane were included in the analysis. For MRI-FMT imaging, the image was segmented into regions of broadly defined tissue types using MIMICS image processing software (Materialise Medical Software, Leuven, Belgium). Tissues were delineated as brain, the area outside the brain, any gadolinium-enhanced features in the brain, and any other abnormal-looking features not associated with Gd-enhancement. Since no abnormal features were evident in the brains of control animals, only the brain and area surrounding the brain were included in the segmentation. Examples of segmented MR images from each group (U251, 9L, and control) are shown in Figs. 1, 1, 1, respectively, where the color overlays delineate different regions. Similar segmented masks were produced for each animal and used to generate finite element method (FEM) meshes compatible with the imaging algorithms.

2.4.

Optical Image Reconstruction

FMT image recovery involves fitting calculated data from a light propagation model to the measured data. The diffusion approximation for photon propagation in tissue is the most commonly used model for FMT imaging and was used in this study. These algorithms recover images of fluorescence yield, which is the product of the fluorophore quantum yield, $\eta$ , and the absorption coefficient of the fluorescence compound at the excitation wavelength, ${\mu }_{af}$ . While early FMT systems were designed as stand-alone imaging systems, the latest generation of research FMT instruments are hybrid systems integrated with established imaging modalities. Paralleling the advances in hardware design, several methods to synthesize the dual data sets have been explored. Two of these methods are considered herein.

In this study, four diffusion-based recovery techniques were investigated, each representing a different degree to which the segmented MRI masks guide the imaging algorithm. The general formulation used to match the measured and calculated fluence rates, ${\Phi }_{fl}^{\mathit{Meas}}$ and ${\Phi }_{fl}^C$ , respectively, and thus recover fluorescence yield is

Internal tissue structure was incorporated into the algorithm in two ways. The hard-priors technique assumes that all segmented regions have homogeneous values of fluorescence yield and thus puts complete trust in the quality of the segmented MR images. The formulation is straightforward and here is accomplished by reducing the Jacobian matrix based on tissue regions;

Another spatially guided approach introduces the segmented information through a spatially varying regularization parameter and is termed the “soft-priors” or “L-matrix” approach.^{17, 30} The updated formulation is written as

The fourth diffusion model–based technique investigated in this study is formulated on the assumption that the fluorescence yield value in the entire head is homogeneous. Like the unguided technique, the geometry of the imaging domain is assumed known, and internal tissue information is not included in the algorithmic reconstructions; however, all nodal values of fluorescence yield are locked to the same value. This “homogeneous fitting” technique is analogous to a single-region hard-priors reconstruction.

All four diffusion-based reconstruction techniques used in the study are summarized in the first four rows of Table 1 . A critical consideration when implementing these techniques is the estimation of the optical properties (absorption and reduced scattering coefficients, ${\mu }_a$ and ${\mu }_s^{∕}$ ) that form the foundation of the light propagation model. Typically, these values are not recovered explicitly and thus must be estimated. For the hard- and soft-priors techniques, the segmented MR images were used to assign heterogeneously distributed optical properties based on two broad tissue types, the brain and all tissue outside the brain. The optical properties in the brain region were estimated from values for rat brain published in the literature³¹: ${\mu }_{ax}=0.03{\mathrm{mm}}^{-1}$ , ${\mu }_{sx}^{∕}=2.25{\mathrm{mm}}^{-1}$ at the excitation wavelength and ${\mu }_{am}=0.03{\mathrm{mm}}^{-1}$ , ${\mu }_{sm}^{∕}=2.75{\mathrm{mm}}^{-1}$ at the emission wavelength. Suspected tumor regions in the brain were assigned these values as well. The tissue surrounding the brain is a mixture of bone, muscle, adipose tissue, and skin, all of which were lumped into a single homogeneous region with estimates of ${\mu }_{ax,m}=0.01{\mathrm{mm}}^{-1}$ and ${\mu }_{sx,m}^{∕}=1{\mathrm{mm}}^{-1}$ at the exciting and emitting wavelengths. Since the internal distribution of tissue types in the unguided and homogeneous fitting algorithms were assumed unknown, the optical properties for these techniques were assumed to be homogeneous throughout the head ( ${\mu }_{ax}=0.02{\mathrm{mm}}^{-1}$ , ${\mu }_{sx}^{∕}=1.88{\mathrm{mm}}^{-1}$ , ${\mu }_{am}=0.02{\mathrm{mm}}^{-1}$ , and ${\mu }_{sm}^{∕}=1.63{\mathrm{mm}}^{-1}$ ). These values are the averages of the properties in the two tissue regions used in the hard- and soft-priors implementations.

Table 1

Descriptions of the eight data processing techniques considered in this study.

Spectra measured for each source–detector combination were calibrated and then preprocessed with a simple spectral unmixing algorithm described in previous publications.^{17, 18} This procedure fits premeasured fluorescence spectra of the imaging agent and tissue autofluorescence to the measured data and thus decouples these signals from one another. Once decoupled, the unmixed dye fluorescence spectra were integrated and input into the imaging algorithms.

2.5.

Bulk Fluorescence Spectroscopy

To compare the diagnostic capacity of MRI-FMT to simpler optical spectroscopy methods, the same optical data used in the image recovery algorithms were also analyzed independent of the imaging problem. These spectroscopy techniques consider the measured spectra alone and therefore do not require geometry-specific FEM modeling or diffusion approximation computations. The four different approaches considered in this study involved calculating the ratios of the unmixed dye fluorescence intensity with either the excitation intensity or the unmixed autofluorescence signal. Descriptions of each method are included in the last four rows of Table 1. The diagnostic performance of the four diffusion model–based techniques and the four bulk spectroscopy techniques were compared using one-tailed student’s $t$ -tests and receiver operating characteristic (ROC) curves.

3.1.

MRI-FMT Imaging

Examples of MRI-FMT images for three mice are shown in Fig. 2 . The first column shows images of a mouse with a U251 tumor, the second column shows a mouse with a 9L tumor, and the third column shows images from a control mouse. Each row corresponds to images recovered using a different reconstruction technique. Images in row I were recovered using hard-priors reconstructions; row II, soft-priors reconstructions; and row III, unguided reconstructions. For illustrative purposes, the color scales are consistent for each example animal (columns) but vary between subjects.

Fig. 2

Representative images of fluorescence yield overlying the corresponding MR images. Each column provides results for a single animal from one of the three mouse populations (U251 tumors, 9L tumors, and controls, respectively). The rows show images recovered using the hard-prior technique (I), the soft-prior technique (II), and an unguided reconstruction technique that operates under the assumption that only the outer boundary was known (III). Given the wide range of recovered values between subjects, the color scales were kept consistent only within the image set for each animal. Transparency levels were varied between images to ensure that the MR images are easily discerned through the optical image overlay. (Color online only.)

Examining the images collectively reveals dramatic difference between the U251 animal and the other two animals for all three imaging techniques. Considering the images in row I, the maximum fluorescence yield value in the U251 animal is approximately tenfold higher than in either the 9L or control animals. The elevated fluorescence activity was well localized in the Gd-enhanced regions of the brain with the U251 tumor, indicating that the fluorescence activity is primarily in the tumor region. This is consistent for the hard- and soft-priors reconstructions for this mouse, which produced similar qualitative results. The unguided reconstruction image for the U251 animal shown in row III produced elevated levels of fluorescence in the same general region, although the spatial resolution and contrasts were substantially lower than produced by the hard- and soft-priors algorithms. Levels of fluorescence yield in the 9L tumor were lower than in the head, indicating poor tumor-to-normal tissue contrast, even when MRI data was used in the imaging algorithms. The unguided solution indicated no tendency toward elevated fluorescence in the Gd-enhanced region, unlike the results in the U251 animal. Images of the control animal also show relatively low overall levels of fluorescence activity and low contrasts throughout the entire head.

The examples in Fig. 2 are representative of the entire image set, demonstrated by Fig. 3 , which presents images of fluorescence yield recovered using the soft-priors reconstruction for all animals included in the study. The images are plotted on a consistent color scale, which reveals the large differences between the animal groups. According to the images, U251 tumors produce much higher levels of fluorescence activity than found in 9L or control animals, while there is no noticeable difference between the 9L and control animals themselves. Similar image sets were generated for the hard-priors and unguided reconstruction techniques but are not included here for the sake of brevity.

Fig. 3

Images of fluorescence yield for all animals in the study recovered using the soft-priors image reconstruction technique. The columns correspond to the U251, 9L, or control mouse populations. All images are plotted on the same color scale. (Color online only.)

The fluorescence activity was highest in the Gd-enhanced region of the head in all but one of the animals in the U251 group. In this animal (the last image in the U251 column of Fig. 3), the elevated region of fluorescence activity corresponded to a dark, abnormal feature in the brain that was segmented during the MR image preprocessing. The Gd-enhanced region, circled in blue, produced lower fluorescence activity than the maximum value, but higher activity than the brain background. (Color online only.) This value was also higher than all recovered values in the 9L and control animals.

To quantify the imaging results, the mean fluorescence yield value in the region defined by Gd-enhancement was calculated for all tumor-bearing animals. This corresponds to the light blue region in the example shown in Fig. 1 and the light pink region in the example in Fig. 1. (Color online only.) Since no Gd-enhancement was observed in the brains of control mice, the mean value in the entire brain region was used. These values are compiled in the box and whisker plots presented in Fig. 4 for the hard-priors, soft-priors, and unguided reconstructions [Figs. 4, 4, 5 , respectively]. The values plotted in Fig. 4 were recovered using the homogeneous fitting approach, which essentially averages the fluorescence activity over the entire domain. These plots show reasonable separation between U251 and the other two groups, although the difference between 9L tumors and control animals is less obvious. $P$ -values between the animal groups calculated using a one-tailed student’s $t$ -test are tabulated in the first four rows of Table 2 . The differences between U251 tumors and the other two groups were statistically significant for all four diffusion-based recovery algorithms. However, only the hard-priors imaging technique revealed a statistically significant difference between the 9L tumor-bearing and control groups.

Fig. 4

Box and whisker plots of fluorescence yield calculated using diffusion-based reconstruction techniques. The values in (a) through (c) were determined by calculating the mean value of fluorescence yield in the region corresponding to gadolinium enhancement for animals with tumors and the mean value in the entire brain region for control animals that showed no gadolinium enhancement. The panels show the results for hard-prior reconstructions (a), soft-prior reconstructions (b), and the unguided reconstruction (c). Values in (d) were recovered using homogeneous fitting, which assumes that the fluorescence activity distribution is homogeneous through the entire head of a mouse subject. Note that the $y$ -axis values are plotted on a different scale for each panel.

Fig. 5

Raw optical spectra measured with the MRI-FMT are preprocessed using a simple spectral fitting algorithm. Examples of measured fluorescence spectra are shown in (d) through (f) (solid line) for U251, 9L, and control mice, respectively. These spectra correspond to a source-detector pair that transmits light from the top of the head to the bottom, as shown in (a) through (c). For illustrative purposes, the sensitivity values between the source and detector are plotted on the corresponding MR images, and an outline of the tumor region as defined by the gadolinium contrast enhancement is included to demonstrate the extent to which the tumor lies in the sensitivity field. Prerecorded spectra of the fluorescent probe and tissue autofluorescence from mice are used to decouple the signals originating from tissue autofluorescence and the optical probe itself. The results of this process are shown in (g) through (i), and the fitted spectra are compared to the measured spectra in (d) through (f).

Table 2

p -values from one-tailed student’s t -test.

3.2.

Bulk Fluorescence Spectroscopy

Figure 5 shows the results of the spectral unmixing algorithm in three example animals, each from a different group. The source–detector measurement pair used in these examples was between the top of the head and the opposing detector below the jaw (dorsal to ventral source–detector pair), illustrated in panels (a) through (c) for the three mice. For illustrative purposes, the sensitivity fields for the source–detector pair overlie the MR images. Also included are outlines of the tumor region as defined by Gd-contrast. The solid lines in the graphs in panels (d) through (f) show the measured spectrum for each animal. The sharp jump in intensity at $720\mathrm{nm}$ corresponds to the cut-on wavelength of the long-pass filter positioned at the entrance of the spectrometers. The results of the spectral unmixing for each animal are plotted in (g) through (i) and illustrate the relative contributions of the imaging agent fluorescence and the tissue autofluorescence. In all cases, the tissue autofluorescence is the dominant signal, revealing the importance of the spectral unmixing process. The sums of the unmixed basis spectra are plotted as fitted spectra in (d) through (f) (dashed lines), demonstrating an excellent match between measured and fitted data.

Visual inspection of the results in Fig. 5, 5, 5 suggests that the spectra alone contain diagnostic information, without the need for light modeling and image reconstruction. While the absolute intensities of the measured spectra are not correlated with the presence of tumors in these examples, the relative intensity of the dye fluorescence to the autofluorescence, or perhaps the excitation intensity (not shown), may well correlate with the presence of a particular tumor type.

The box and whisker plots and corresponding $p$ -value calculations in Fig. 6 and Table 2, respectively, confirm these observations. In all cases, significantly higher ratio values were calculated for the U251 group relative to the 9L and control groups. Also, two ratio-based techniques produced statistically significant differences between 9L and control animals.

Fig. 6

Box and whisker plots of different measurement ratios. The plots in the top panels were determined by calculating either the sum (a) or average (b) of all fluorescence-to-excitation ratios for a given animal. The plot in (c) contains data determined by calculating the sum of the fluorescence-to-autofluorescence ratios for each animal. Using this ratio eliminates the need to measure the excitation intensity separately and ensures that the numerator and quotient are perfectly calibrated to one another. Last, the fluorescence-to-autofluorescence ratio of single measurement point (shown in Fig. 2) for each animal was used to produce the results plotted in (d). Note that the $y$ -axis values are plotted on a different scale for each panel.

3.3.

ROC Analysis

The results compiled in Tables 2, 3 indicate that nearly all methods identify the presence of U251 tumors versus 9L tumors or controls. Area-under-the-curve (AUC) values determined from ROC analysis for U251 versus control animals show perfect diagnostic performance for all imaging methods and bulk spectroscopy methods. Similar performance was observed between U251 and 9L tumors, with one exception. The AUC determined from the sum of the fluorescence-to-autofluorescence ratios slipped to 0.93, suggesting that this simple data processing method is a less powerful diagnostic technique.

Table 3

Area-under-curve values from ROC analysis.

Differentiating subjects with 9L tumors from healthy subjects proved to be more difficult for all methods. Three techniques produced statistically significant differences between these groups; the hard-priors imaging technique and the maximum and mean of the fluorescence-to-excitation spectroscopy ratios. These statistical differences translated into reasonably reliable diagnostic performance with AUC values of 0.95, 0.9, and 0.85 for the imaging technique, and maximum and mean fluorescence-to-excitation ratio spectroscopy methods, respectively. The soft-priors imaging technique reported the poorest diagnostic performance between these groups, as the AUC value was just over 0.55, indicating that the groups were nearly indistinguishable. The unguided and homogeneous fitting reconstructions also demonstrated poor diagnostic characteristics with an $\mathrm{AUC}=0.65$ for both methods. All spectroscopic methods produced AUC values over 0.8 for the 9L versus control tests.