2.1.

Imaging System

Zeiss Axioskop II (Duke University) and AxioImager (University of Florida) microscopes (Carl Zeiss, Inc., Thornwood, New York) served as the imaging platform. A $100\text{-}\mathrm{W}$ tungsten halogen lamp was used for transillumination of the window chamber. Images were acquired at $1388\times 1024\text{pixels}$ and $12\text{-}\text{bit}$ dynamic range with a DVC 1412 charge-coupled device (CCD) camera (DVC Company, Austin, Texas) that was thermoelectrically cooled to $-20°\mathrm{C}$ . The objectives used were all air immersion with long working distances (WD) (Carl Zeiss, Inc.): $2.5\times$ [numerical aperture $\left(\mathrm{NA}\right)=0.12$ , $\mathrm{WD}=6.3\mathrm{mm}$ ] and $5\times$ ( $\mathrm{NA}=0.25$ , $\mathrm{WD}=12.5\mathrm{mm}$ ) Fluars, and a $10\times$ ( $\mathrm{NA}=0.3$ , $\mathrm{WD}=5.5\mathrm{mm}$ ) Plan-NeoFluar. Bandlimited optical filtering for hyperspectral imaging was accomplished with a C-mounted liquid crystal tunable filter (CRI, Inc., Woburn, Massachusetts) with a $400\text{-}\text{to}720\text{-}\mathrm{nm}$ transmission range and $10\text{-}\mathrm{nm}$ nominal bandwidth placed in front of the camera. Images were saved as $16\text{-}\text{bit}$ tagged image files. The field-of-view captured by the CCD camera was different for the two microscope systems due to slight demagnification ( $\sim 0.75\times$ , Duke microscope) or magnification ( $\sim 1.5\times$ , University of Florida microscope) in relay lenses used to create an infinity space for the liquid crystal tunable filter. Thus, the fields-of-view imaged by the CCD camera were the following: $2.5\times =3.6\times 4.9\mathrm{mm}$ (Duke) or $2.3\times 3.1\mathrm{mm}$ (Florida), $5\times =1.8\times 2.5\mathrm{mm}$ (Duke) or $1.2\times 1.6\mathrm{mm}$ (Florida), $10\times =0.9\times 1.2\mathrm{mm}$ (Duke) or $0.6\times 0.8\mathrm{mm}$ (Florida). Scale bars in the figures give an indication of the field-of-view and object dimensions in the images.

2.2.

Image Acquisition

Custom designed software created with LABVIEW (National Instruments Corp., Austin, Texas) was used to control the tuning of the filter and operation of the camera. The software allows automated image acquisition with user-specified camera exposure time and gain for each filter wavelength. For each HbSat time point, 16 images were acquired from $500\text{to}575\mathrm{nm}$ in $5\text{-}\mathrm{nm}$ intervals. The transmission of the liquid crystal tunable filter is lower at shorter wavelengths and higher at longer wavelengths; therefore, the image acquisition time was adjusted a priori for each wavelength such that the full dynamic range of the camera was used. The minimum exposure time used was $300\mathrm{ms}$ to average out fluctuations in the signal due to random red blood cell motion.²³ A typical data set was acquired in about $13\mathrm{s}$ , which included image acquisition, filter tuning, image transfer, and saving of the images to the computer hard drive.

2.3.

In vivo Surgery and Imaging

All in vivo experiments were conducted under protocols approved by the Duke University Institutional Animal Care and Use Committee and the University of Florida Institutional Animal Care and Use Committee. A titanium window chamber was surgically implanted under anesthesia [ketamine $100\mathrm{mg}∕\mathrm{kg}$ intraperitoneal (IP) and xylazine $10\mathrm{mg}∕\mathrm{kg}$ IP] on the back of athymic (nu/nu) nude mice. The tumors used in these experiments were the 4T1 metastatic mouse mammary adenocarcinoma and the 4TO7 nonmetastatic subclone of the 4T1 tumor. A window chamber tumor was established during chamber implantation by injecting $10\mu \mathrm{L}$ of a single cell suspension of $5\times {10}^3$ cells into the dorsal skin flap prior to placing a $12\text{-}\mathrm{mm}$ diameter #2 round glass coverslip (Erie Scientific, Portsmouth, New Hampshire) over the exposed skin. In this window chamber model, one piece of skin was removed completely and replaced with a coverslip, but the other piece of skin in the window chamber remained intact and without any coverslip. The tumor had unrestricted growth on the side of the chamber without a coverslip and thus bulged out during growth. This limited transillumination imaging to a maximal tumor size of about $2\text{to}3\mathrm{mm}$ because the light that penetrated the tumor was insufficient for the hyperspectral imaging technique due to poor signal to noise ratio and experiments were terminated at this point. Animals were housed in an environmental chamber with free access to food and water and standard $12\text{-}\mathrm{h}$ light and dark cycles. Animals were placed on a heating pad attached to the microscope stage during the imaging session. A custom-built window chamber holder secured the window chamber under the microscope objective. The holder was fastened to a stage mount that allowed the window chamber to be positioned under the microscope objective using the standard manual controls on the microscope stage. Anesthesia for immobilization during imaging was provided by isoflurane (1 to 1.5%) in medical air. Image data sets were acquired as follows. (1) For investigations of time-averaged values of HbSat in microvessels, image data sets were acquired at intervals of 1 or $3\min$ for approximately $1\mathrm{h}$ . (2) For HbSat fluctuation analysis, image data sets were acquired at $20\text{-}\mathrm{s}$ intervals for $40\min$ . Data set acquisition at $20\text{-}\mathrm{s}$ intervals represents the maximum achievable time resolution with the current imaging system parameters, with image exposure time being the limiting factor. In one set of experiments, animals were imaged while the breathing gases were changed from room air to 100% oxygen. In this case, anesthesia for immobilization during imaging was provided by ketamine $100\mathrm{mg}∕\mathrm{kg}$ IP and xylazine $10\mathrm{mg}∕\mathrm{kg}$ IP.

2.4.

Image and Data Processing

Image processing was performed using MATLAB software (The Mathworks, Inc., Natick, Massachusetts). All images were converted into double-precision arrays for mathematical processing. The raw pixel values were converted to absorbance values by manually selecting avascular regions in the images to use as an estimate of unattenuated light. Pure reference spectra of oxy- and deoxyhemoglobin obtained with the imaging system were used in a linear mixing model to solve for HbSat by a linear least-squares regression of the data according to the method of Shonat²⁵ as described previously.^{23, 24} Regions-of-interest (ROIs) on individual microvessels were manually selected for data points of HbSat. The calculated values of HbSat for each pixel were accepted or rejected based on the $R^2$ value for the fit of the model to the data based on the calculated unknowns. $R^2$ was calculated according to the standard definition for the coefficient of determination:²⁶ $R^2=1-S{S}_e∕S_t$ , where $S{S}_e=\text{error}$ sum of squares (variance considering regressors), and $S_t=\text{total}$ variance (no regressors considered). Pixels with $R^2$ values $<0.90$ were rejected and not considered in calculations or included in saturation maps. The HbSat value of the ROI is reported as the mean HbSat value of the pixels within the ROI. The 95% confidence intervals of the mean HbSat for the ROIs at a single time point were calculated as $\overline{x}\pm z_c(\sigma ∕\sqrt{n})$ , where $\overline{x}=\mathrm{ROI}$ mean, $z_c=\text{critical}$ $\text{value}=1.96$ for 95% confidence level, $\sigma =\mathrm{ROI}$ standard deviation, and $n=\text{number}$ of pixels in the ROI. For statistical analysis, analysis of variance (ANOVA) was performed followed by Newman-Keuls comparison when statistical differences were indicated. A value of $p<0.05$ was used as the threshold for statistical significance.

For HbSat fluctuation analysis, median ROI HbSat values from $40\text{-}\min$ time series were analyzed by examining the power spectra of each time series. Each series of data contained 123 data points sampled at $20\text{-}\mathrm{s}$ intervals. In some cases, excessive animal movement or other artifact corrupted the data such that a time point was lost. In this case, linear interpolation was used to fill in the missing data point. No more than two data points in any data set were lost and had to be treated in this manner. To obtain the power spectrum, the data were mean-subtracted to remove the dc component and padded with zeros at the end of the time series to 128 data points (a power of 2) so the fast Fourier transform algorithm could be used for the Fourier transform. Fourier transform of the time series and power spectra were calculated with MATLAB software.

3.1.

Oxygen Transport Longitudinal Gradients

Figure 2 shows a region of 4T1 tumor microvasculature nine days after tumor cell implantation with various ROIs indicated. The ROIs were chosen such that they followed the course of several microvessels to observe the presence of gradients in HbSat along the vessels. Figure 3 is a graph of the fluctuations recorded at $3\text{-}\min$ intervals in HbSat over time for one group of ROIs associated with a particular vessel path. Table 1 is a list of the time-averaged HbSat for each of the ROIs in Fig. 3, with the ROIs grouped by association with a particular vessel path (see Fig. 2). ANOVA and Newman-Keuls comparisons $(p<0.05)$ of the time-averaged HbSat were performed on the separate groups of ROIs (i.e., ROIs 1 to 6, 7 to 10, and 11 to 15). It is evident from Fig. 3 and Table 1 that, for the selected vessels, there is little indication of time-averaged gradients in HbSat. Although for one group of ROIs the difference in the average HbSat of ROI 1 $(82\%\pm 3\%)$ from ROIs 2 to 6 rises to the level of statistical significance $(p<0.05)$ , biologically this difference is probably not meaningful (maximum average of 85% for ROIs 2 to 6). In ROIs 11 to 15, ROIs 13 and 11, respectively, are statistically the highest $(83\%\pm 3\%)$ and lowest $(73\%\pm 5\%)$ time-averaged values with the other ROIs somewhere in between and not statistically different from each other. The largest measured fluctuation range in the ROIs for this tumor occurred with ROI 11 (minimum $\mathrm{HbSat}=55\%$ , maximum $\mathrm{HbSat}=81\%$ , $\text{range}=26\%$ ). The average fluctuation range of all the ROIs is $18\%\pm 5\%$ .

Fig. 2

Tumor nine days after implantation of $5\times {10}^3$ cells. The numbered ROIs correspond to the ROIs in Fig. 3 and Table 1. The scale bar is $200\mu \mathrm{m}$ ( $10\times$ objective, $0.9\times 1.2\mathrm{mm}$ field-of-view).

Fig. 3

HbSat fluctuations in ROIs 1 to 6 as indicated in Fig. 2. The data points plotted are the median $\mathrm{HbSat}\pm \text{interquartile}$ range in the ROI. The lines connecting the data points were added to aid in visualization and do not imply interpolation between time points. Data points were obtained about every $3\min$ .

Table 1

Time-averaged values of mean HbSat±standard deviation for the ROIs in Fig. 2. The ROIs were grouped for statistical analysis (see Sec. 2.4 for details) according to apparent common microvascular flow pathways with groups indicated by horizontal borders in table. Statistically significant differences (ANOVA and Newman-Keuls, p<0.05 ) are indicated in the table (Y).

Figure 4 shows a region of 4T1 tumor microvasculature seven days after tumor cell implantation in a tumor different from the one in Fig. 2 with various ROIs chosen similarly as before. A graph of the fluctuations in HbSat over time recorded at $3\text{-}\min$ intervals for a group of ROIs is shown in Fig. 5 , and Table 2 is a list of the time-averaged HbSat for each of the ROIs in Fig. 4. Statistical comparisons were performed as before. The graph of HbSat temporal fluctuations in Fig. 5 for the selected group of ROIs clearly clearly shows the presence of gradients in HbSat that persist over time. The time-averaged HbSat values in Table 2 indicate the existence of time-averaged gradients in HbSat in the various groups of ROIs, particularly for ROIs 5 to 9. In this tumor, the fluctuations in HbSat tended to be larger than those for the previously described tumor. The largest measured fluctuation range in the ROIs for the tumor in Fig. 4 occurred with ROI 11 (minimum $\mathrm{HbSat}=35\%$ , maximum $\mathrm{HbSat}=83\%$ , $\text{range}=48\%$ ). The average fluctuation range of all the ROIs is $26\%\pm 12\%$ . An interesting observation is that within each group of ROIs in this tumor, the ROIs with the highest time-averaged HbSat (ROIs 4, 8, 9, and 13) are all located in close proximity (see Fig. 4), and the average HbSat tends to decrease along each of the respective microvessel branches toward the junction where they meet (ROI 1), at which point the average HbSat increases. This may imply a common underlying mechanism that affects the general region of the tumor where the observed vessels are located.

Fig. 4

Tumor seven days after implantation of $5\times {10}^3$ cells. The numbered ROIs correspond to the ROIs in Fig. 5 and Table 2. The scale bar is $200\mu \mathrm{m}$ ( $10\times$ objective, $0.9\times 1.2\mathrm{mm}$ field-of-view).

Fig. 5

HbSat fluctuations in ROIs 5 to 9 as indicated in Fig. 4. The data points plotted are the median $\mathrm{HbSat}\pm \text{interquartile}$ range in the ROI. The lines connecting the data points were added to aid in visualization and do not imply interpolation between time points. Data points were obtained about every $3\min$ .

Table 2

Time-averaged values of mean HbSat±standard deviation for the ROIs in Fig. 4. The ROIs were grouped for statistical analysis (see Sec. 2.4 for details) according to apparent common microvascular flow pathways with groups indicated by horizontal borders in table. Statistically significant differences (ANOVA and Newman-Keuls, p<0.05 ) are indicated in the table (Y). ROIs 8 and 9 were not statistically different from each other (indicated by “*”) but were different from the other ROIs in the analysis group.

In contrast to the tumor microvascular networks, gradients in HbSat such as those found in the tumor were not evident in normal microvessels. Figure 6 shows vessels in a nontumor bearing window chamber two days postsurgery. Next to each ROI in the figure is the $60\text{-}\min$ time-averaged HbSat with data points taken at $1\text{-}\min$ intervals. With the exception of the ROIs marked with a “ ${}^{*}$ ,” there was no statistical difference in the average HbSat in this network of vessels (ANOVA and Newman-Keuls, $p<0.05$ ) when ROIs were compared to those grouped in the same branch and vessel segment.

Fig. 6

Image of a normal microvessel network in a nontumor bearing window chamber two days postsurgery ( $2.5\times$ objective, $2.2\times 3.1\mathrm{mm}$ field-of-view). The $\text{mean}\pm \text{standard}$ deviation appear next to each ROI along the vessels to indicate the time-averaged HbSat over $60\min$ . ROIs marked with a “ ${}^{*}$ ” were statistically different (ANOVA and Newman-Keuls, $p<0.05$ ) when compared to other ROIs along the same vessel segment, but not when compared as a group to other ROIs along the connecting branch segment.

3.2.

Oxygen Transport Fluctuations

Figure 7 is an example analysis of time series data of HbSat fluctuations. The data in the figure is for a 4TO7 tumor at seven days after tumor cell implantation. In Fig. 7a, the transmitted light image shows a region at the periphery of a tumor, with normal tissue on the left of the image and tumor tissue on the right of the image. ROI 1 is from an arteriole in an area of normal tissue, ROI 2 is from a venule in normal tissue, and ROI 3 is a representative vessel from the tumor microvessel bed seen in the image. Figure 7b is a plot of the HbSat fluctuations over time for the three ROIs in Fig. 7a. Even though the arteriole HbSat exhibits relatively high frequency HbSat fluctuations, low frequency fluctuations dominate the venule data. The tumor microvessel tends to have an intermediate average HbSat value between the arteriole and venule with low frequency HbSat fluctuations that follow the pattern of the nearby venule. The tumor microvessel fluctuations are larger in magnitude than the venule fluctuations. In Fig. 7c, it can be seen that a number of higher frequency components are present in the arteriole power spectrum, but lower frequency components dominate the power spectra of the venule and tumor microvessel. This trend was consistent in the 4T1 and 4TO7 tumors that were imaged and for normal arterioles and venules in nontumor bearing window chambers. Similar to the spectra in Fig. 7c, the largest peaks in the power spectra for venules and tumor microvessels occurred for frequencies below $0.2\mathrm{cpm}$ , and other frequency components generally had less than half the magnitude of the largest peak.

Fig. 7

Fluctuation analysis of HbSat in microvessels. (a) Transmitted light image of the edge of a 4TO7 tumor ( $5\times$ objective, $1.1\times 1.6\mathrm{mm}$ field-of-view). Normal tissue and vessels are on the left side of the image, while tumor tissue and the associated microvessel network are on the right side of the image. ROI $1=\text{normal}$ tissue arteriole, ROI $2=\text{normal}$ tissue venule, ROI $3=\text{tumor}$ microvessel. (b) Temporal fluctuations in HbSat for the ROIs. Data represent median values of a ROI and were acquired at $3\mathrm{cpm}$ ( $20\text{-}\mathrm{s}$ intervals). Interquartile ranges were omitted for clarity. (c) Normalized power spectra of the time series data for the ROIs.

3.3.

Anastomosis of Vessels with Different Oxygenations

Figure 8 depicts a direct connection between 4T1 tumor vessels with considerable differences in HbSat. Several smaller vessels with respective saturations of about 66% and 37% converged into one larger diameter vessel. There appeared to be laminar flow in the large diameter vessel, as the different flows propagated without mixing for a distance of at least $500\mu \mathrm{m}$ . The HbSat and direction of blood flow are indicated in the figure. Figure 9 is another example of this type of connection in a 4T1 tumor. Similar to Fig. 8, several vessels with different HbSats are seen merging together. The saturations of the merging vessels in this example, 61% and 39%, are similar to the values in the previous example in Fig. 8.

Fig. 8

Images depicting the direct connection of tumor microvessels with highly different HbSats ( $\text{left}=2.5\times$ objective, $3.6\times 4.8\mathrm{mm}$ field-of-view; $\text{right}=10\times$ objective, $0.9\times 1.2\mathrm{mm}$ field-of-view). The left image is a low magnification brightfield image, and the right image is a HbSat map of the boxed region in the brightfield image. The direction of blood flow is indicated by the arrows, and the HbSat of the two parallel flows are indicated in the HbSat map. The HbSat values are given as the $\text{mean}\pm \text{standard}$ deviation rounded to the nearest percent and 95% confidence intervals for the mean HbSat of the ROI.

Fig. 9

Another example of merging microvessels with very different HbSat values ( $\text{left}=2.5\times$ objective, $3.6\times 4.8\mathrm{mm}$ field-of-view; $\text{right}=5\times$ objective, $1.8\times 2.4\mathrm{mm}$ field-of-view). The HbSat values of the merging vessels are similar to those in Fig. 8. The HbSat values are given as the $\text{mean}\pm \text{standard}$ deviation (rounded to the nearest percent) and 95% confidence intervals for the mean HbSat of the ROI.

Figure 10 is an example of the effect, presumably due to AV anastomoses or similar structures, on oxygen transport in a 4T1 tumor nine days after implantation of tumor cells. One set of images shows the tumor, and the other set of images illustrates adjacent normal tissue (some tumor feeding vessels from the tumor images are also visible in the lower right of the normal tissue images). Table 3 contains HbSat values for the ROIs indicated in Fig. 10. Under normal air-breathing conditions, the HbSat in some of the venules immediately adjacent to the tumors (ROIs 4 and 5 in the Fig. 10 tumor) was slightly elevated compared to other nearby venule branches (ROIs 2 and 3 in the Fig. 10 tumor). In the switch to 100% oxygen breathing, there was an appreciable increase in HbSat of almost 30% in the venules that had elevated air-breathing values, which may indicate that some highly oxygenated red blood cells largely bypassed the tumor and were diverted into adjacent venules. Farther away from the tumor in normal tissue, there was no conspicuous indication of this effect in the microvasculature and only a minimal increase in venule HbSat with oxygen breathing. The arterioles in both the tumor (ROI 1) and normal tissue (ROI 2) images in Fig. 10 increased to about the same HbSat with oxygen breathing.

Table 3

HbSat values for the ROIs indicated in Fig. 10. The ROIs are indicated in the middle group of images (“Air”). The HbSat values are given as the Mean±standard deviation (rounded to the nearest percent) and 95% confidence intervals for the means HbSat of the ROI.