2.1.

Animals and Sample Preparations

In this study, the vascular images were segmented based on autofluorescence images of rat organs. All the animal studies and experiments were approved by the Institutional Animal Care and Use Committee (IACUC) at the Medical College of Wisconsin. The studies were performed using two rodent species, rats, and mice. Lungs, the lateral lob of liver, and kidneys were harvested from non-irradiated and irradiated adult female WAG/RijCmcr rats and hearts from non-irradiated male C57BL/6J mice. For lung sample preparation, the airway of the lungs was first inflated by gravity with 1 to 2 mL fluorescein isothiocyanate–dextran (FITC-dextran, MW 150,000, $100\mu \mathrm{M}$ solved in water). In addition to the airway injection of the lungs, the sample preparation was similar for all organs. They were immersed in chilled liquid isopentane for a couple of minutes before transferring them to liquid N2. All samples were stored in a $-80°\mathrm{C}$ freezer until optical cryo-imaging was performed.

2.2.

3D Fluorescence Metabolic Cryo-Imaging

The 3D fluorescence cryo-imager system was custom-designed in the Biophotonics Laboratory at the University of Wisconsin Milwaukee. The system captures 3D NADH and FAD fluorescent signals of frozen organs/tissues. The flash-frozen sample is stored in $-80°\mathrm{C}$ freezer to ensure the preservation of the metabolic state of the tissue. A complete description of the system can be found in our recent cryo-imaging studies.^26,40 Briefly, a mercury arc lamp (200 W lamp, Oriel, Irvine, CA, in the light source from Ushio Inc., Japan) is used as the light source. Appropriate optical filters at selected wavelengths are utilized to excite the specific fluorophores from the surface of the frozen tissue. For the NADH channel, excitation and emission filters were set at $350\pm 80\mathrm{nm}$ (UV Pass Blacklite, HD Dichroic, Los Angeles, CA) and $460\pm 50\mathrm{nm}$ (Chroma, Bellows Falls, VT), respectively. The excitation and emission filters for the FAD channel were set at $437\pm 20\mathrm{nm}$ (Omega Optical, Brattleboro, VT) and $537\pm 50\mathrm{nm}$ (Omega Optical), respectively. Lungs were also imaged using FITC specific optical filter sets: excitation at $494\pm 20\mathrm{nm}$ (Edmund Optics) and emission at $537\pm 50\mathrm{nm}$ (Omega Optical) for airway detection. For imaging Td-Tomato kidney, besides the regular NADH channel, we also used a red channel of imaging, with the excitation and emission filters set at $545\pm 25\mathrm{nm}$ (Chroma) and $645\pm 50\mathrm{nm}$ (Chroma), respectively. All filters are controlled by two motorized filter wheels (Oriental Motor Vexta Step Motor PK268-01B). The emitted fluorescent signals are captured with the image recordings system (CCD camera, QImaging, Rolera EM-C2, 14 bit).

The 3D NADH and FAD cryo-images, representing mitochondrial redox state of tissues, were analyzed using a code written in MATLAB. Calibration was performed using a flat-field image of both NADH and FAD channels. The 3D-rendered redox ratio (RR) image (NADH/FAD) was calculated voxel-by-voxel. Also attempts to understand the heterogeneity of the tissue were made to correlate the RR with the anatomy of the hearts⁴¹ and kidneys.⁴⁰ The following section describes how the autofluorescence images provide the structural features of the vascular network of the organs.

2.3.

Vascular Segmentation from Autofluorescence Images

Figure 1 shows the flowchart of the proposed algorithm that we used to segment the background vasculature from the foreground 3D autofluorescence images. A simple implementation steps in FIJI⁴² can be found in Table S1 in the Supplementary Material. The standard preprocessing normalization steps in fluorescence cryo-imaging, such as flat-field calibration is not needed before vascular segmentation because the intensity adjustment in step 1 normalizes between-sample variations, and the background subtraction in step 3 will remove the uneven illumination.

Fig. 1

Algorithm flowchart for background vasculature segmentation from fluorescence images. After loading the 3D stack of images, the brightness and contrast are adjusted to have an enhanced image. Then for each slice, the image is inverted, and the background is subtracted. Additional contrast enhancement and optional masking of the unwanted regions based on the tissue type is done on the 3D image. The reconstructed 3D vasculature can be fed to 3D vessel tracing algorithms for quantification purposes.

Below is the detailed sequence of steps carried out to obtain and reconstruct a vascular network from the inverted fluorescent image.

If the structure comes from high-intensity voxels such as airway in FITC airway injected lungs and red fluorescence images in TdTomato rat kidneys, the same segmentation algorithm without step 2 (image inversion) can be applied.

Fig. 2

Quantification of vascular markers. (a) A schematic view of a branch: a vessel tracing algorithm traces the path from the beginning point to the terminal points. The branch depths are increasing in each bifurcation. (b) A simple branch illustrated with its terminal points and (c) the branch color-coded with vessel diameter ($\mu \mathrm{m}$).

2.4.

Validating VMI Using TdTomato Rats

A genetically modified rat model expressing TdTomato primarily in vascular endothelial cells was utilized to image the vasculature in kidneys. Histological assessment of rat kidneys was also done to visualize TdTomato expression in endothelial cells of these rats using an antibody for TdTomato. The capability of the 3D fluorescence cryo-imaging to acquire images from multiple channels simultaneously allows us to have both red and NADH fluorescence of the kidney. We used the foreground vasculature extracted from red fluorescence as proof for examining the vasculature segmented from the NADH channel using VMI. The co-registration of the vascular network extracted from the two channels validates the proposed method of our vascular segmentation.

One of the common metrics for evaluating the quality of image segmentation is the Dice coefficient, which measures the overlap/merge between the ground truth and the test.⁴⁵ For calculating the Dice coefficient, we let the 3D volume be represented by the point set $X=\{x_1,\dots ,x_N\}$, where $N$ is the total number of voxels. We let the red vasculature be represented by the partition $V_{\text{red}}$ of $X$ with assignment function $f_{\text{red}}(x)$, i.e., voxel intensity at $x$, and we let the VMI vasculature be represented by the partition $V_{\mathrm{vmi}}$ of $X$ with assignment function $f_{\mathrm{vmi}}(x)$. Then the Dice coefficient is defined by

The branching structure of the VMI vasculature can also be compared with red fluorescence vasculature. Murray⁴⁶ proposed an optimization theory that the fundamental structure of a vascular tree should be such that it minimizes work. Murray’s law states that a branch that follows the “minimum work” hypothesis should also follow the equation:

After employing the tracing algorithm using filament tracing in Imaris software, we used the information on the depth of the vessels to define the parents and daughters. The depth of a vessel increases every time a bifurcation happens in the branch. Therefore, all vessels with depth $k+1$ are the daughter vessels of the parent vessels with depth $k$ and Murray’s law can be written as

Notably, using the depth to find the parent–daughter relationship in vessels can impose an unavoidable error by making the left side of Eq. (3) higher than the real value. The reason is that the terminal branches from lower depths [asterisks in Fig. 2(a)] are considered as parent vessel while there are no corresponding daughter vessels in the next depth.

3.1.

3D Vascular and Metabolic Imaging Using Autofluorescence

Figure 3 supports our hypothesis that a foreground fluorescence image can be inverted to reveal the vasculature of an organ like the kidney. The 3D raw NADH (excitation at 350 nm and emission at 460 nm) image of a kidney, a sagittal slice of the kidney, and the segmented vasculature on one slice are illustrated in Fig. 3. The stack of 2D vascular images was reconstructed to generate the 3D vascular images of the whole kidney.

Fig. 3

A background vasculature is segmented from a foreground autofluorescence image of a kidney. For a rat kidney, a 3D raw image of NADH fluorescence is shown. A sagittal slice view of raw kidney image is chosen, and the segmented vasculature from dark voxels is shown in red and merged with the raw slice to show the localization of the vascular pixels in the image. The 3D vasculature is reconstructed from all 2D segmented pixels. The 3D rendered images of NADH and vasculature can be found in Video S1 (Video S1, 685 kB, MP4 [URL: https://doi.org/10.1117/1.JBO.26.7.076002.1]) and Video S2 (Video S2, 471 kB, MP4 [URL: https://doi.org/10.1117/1.JBO.26.7.076002.2]), respectively.

VMI was also applied to other organs, such as heart and liver. Figure 4 shows selected representative slices for the step-by-step images of the algorithm for each organ: kidney, heart, and liver. In step 1, the contrast and brightness of the images are enhanced. The inverted images of one slice of each organ can be seen in step 2. Note that now, the feature of interest (vasculature) is bright in the image. A background-subtracted image of the slice can be seen in step 3. The resulting 3D vascular images are reconstructed from the stack of 2D images.

Fig. 4

VMI can be used to segment vascular networks of multiple organs, including kidney, heart, and liver. An example slice in each step of the algorithm is illustrated for each organ, following which the 3D vasculature is reconstructed. The scale bars for kidney, heart, and liver are 600, 400, and $1000\mu \mathrm{m}$, respectively. Videos from the 3D rendered vasculature images can be found in Videos S2 for kidney (Video S2, 471 kB, MP4), Video S3 for heart (Video S3, 470 kB, MP4 [URL: https://doi.org/10.1117/1.JBO.26.7.076002.3]), and Video S4 for liver (Video S4, 482 kB, MP4 [URL: https://doi.org/10.1117/1.JBO.26.7.076002.4]).

Vascular segmentation from the background of autofluorescence in lung tissues was not feasible because the vasculature, airway, and alveoli appeared dark in the images. Therefore, distinguishing the vasculature from these structures was not possible. To circumvent this problem, we injected an FITC solution into the airway and alveoli. Extrinsic fluorescence from FITC (excitation at 494 and emission at 537) and FAD (excitation at 437 and emission at 537) overlapped. This overlap and the injection of an FITC solution into the airway and alveoli enabled us to make the airway voxels bright in FAD images and keep the vascular structures dark. The same proposed segmentation algorithm was then applied to extract the inverted vasculature from the FAD images of the lungs. Figure 5 shows a 3D raw FAD image of the lung and a single slice of the lung. The airways, which are filled with FITC solution, are segmented from light (higher intensity) voxels in the FAD images. The 3D vasculature (in red) and airway (in green) are then reconstructed in 3D as shown in (Fig. 5). In the combined or merged images, the color of voxels that have an overlap between the segmented airway and vasculature should be yellow, but due to a very little intersection, no yellow voxels appeared in this figure. The Dice coefficient $<0.001$ also confirms that the airway and the vasculature did not overlap. These results demonstrate that the segmentation structures from inverted FAD images do not originate from airways but the vasculature. Note that now, the FAD images are originating from both FITC and FAD fluorescence. This helped us to lighten the airway, but the FITC fluorescence in the airway also interfered with the FAD signal. Therefore, on the downside, the RR is now NADH/(FAD + FITC), which is not an accurate representation of the mitochondrial RR (NADH/FAD).

Fig. 5

FITC airway injection helps to segment background vasculature from lungs. For a rat lung, FITC-dextran solution was injected into the airway. A raw 3D image of FAD fluorescence of lung is shown. A transverse slice view of raw lung image is chosen, and the segmented airway from light voxels of FAD image and vasculature from dark voxels are shown in green and red, respectively. Segmented airway and vasculature are merged with the raw slice to show the localization of the vascular and airway pixels in the segmenting image. Also the 3D vasculature and airway are combined and shown zero intersection (yellow voxel) with each other. Videos from 3D rendered images of FAD, airway, vasculature, and combined images can be found in Videos S5 (Video S5, 722 kB, MP4 [URL: https://doi.org/10.1117/1.JBO.26.7.076002.5]), Video S6 (Video S6, 753 kB, MP4 [URL: https://doi.org/10.1117/1.JBO.26.7.076002.6]), Video S7 (Video S7, 742 kB, MP4 [URL: https://doi.org/10.1117/1.JBO.26.7.076002.7]), and Video S8 (Video S8, 752 kB, MP4 [URL: https://doi.org/10.1117/1.JBO.26.7.076002.8]), respectively.

3.2.

Co-Registration with TdTomato to Confirm VMI Vasculature

The transgenic rat model expressing endothelial-specific TdTomato was used to validate the vascular segmentation by VMI. Figure 6 shows a histological assessment that illustrates the expression of TdTomato in endothelial cells in transgenic rat kidneys (upper row). The renal tubules and most non-endothelial cells of the transgenic TdTomato rat kidneys do not express TdTomato (Fig. 6). Both wild type and transgenic vascular endothelial cells are also stained with endothelial-specific antibody RECA-1 in sections adjacent to those stained for TdTomato. Though the TdTomato-staining in glomeruli was not always as distinct as that of RECA-1, the sections demonstrated co-registration primarily with blood vessels and not with renal tubules (open black arrows).

Fig. 6

Vascular endothelial cells in transgenic but not wild type rats express TdTomato. This figure shows stained sections of kidneys from wild type and transgenic (TdTomato) rat. The upper panel represents sections stained with DsRed antibody specific for TdTomato (stained in brown). The lower panel represents adjacent sections stained for the endothelial marker RECA-1 (brown stain in lower panel). Note positive staining for TdTomato in the endothelial cells of the glomeruli (dotted arrow) as well as large (red arrow) and small (black arrows) blood vessels in the upper panel of transgenic rats. Wild type and TdTomato transgenic vessels in glomeruli (dotted arrows) and large (red arrow), and small (black arrows) vessels in the kidney also stain brown with endothelial-specific antibody RECA-1 (lower panel) confirming the brown staining in the TdTomato transgenics co-register with blood vessels. Open black arrows point to unstained tubules, open dashed arrows point to unstained glomeruli and open red arrows show unstained vasculature. Note the absence of TdTomato brown staining in glomeruli of wild type rats as well as tubules of TdTomato transgenic and wild type rats.

Using the TdTomato transgenic rat model, the cryo-imaging was performed in the two channels of fluorescence, NADH (excitation 350 nm and emission at 460 nm), and red (excitation 545 and emission 645). The bright voxels in the red channel (segmentation algorithm without step 2) and the dark voxels in the NADH channel are segmented and reconstructed [Figs. 7(a) and 7(b), respectively]. In the kidney, the anatomy of the vasculature extracted from the NADH using VMI [Fig. 7(a)] is then combined with the vasculature segmented from red fluorescence [Fig. 7(b)] to make a hybrid image [Fig. 7(c)]. The overlap voxels between the two images [Figs. 7(a) and 7(b)] are displayed in yellow color [Fig. 7(c)]. The co-registration gives a Dice coefficient of 0.91, which shows a high degree of overlap/merge between the two segmented vasculatures.

Fig. 7

(a) TdTomato-transgenic rat kidney validates that VMI has high overlap with vessels expressing red fluorescence. The red channel is used as the ground truth for validating the vasculature extracted using VMI. (b) The segmented vasculature from light voxels of red channel and vasculature from the dark voxels of the NADH channel (VMI technique) are shown in red and green, respectively. (c) Segmented vasculatures are then combined to localize their intersection in yellow. The dice coefficient of 0.91 shows a great precision in vascular segmentation.

The branching of the vasculature between the two signals is also compared in Fig. 8. The relationship between the cubed diameter of the parent vessels to the summation of the cubed diameter of their corresponding branched daughter vessels is presented. Using linear regression, the two lines are fitted to each set of data points as shown in Fig. 8. According to Murray’s law, the data should be fitted to $y=x$ line, i.e., a line with a slope of 1 and $y$ intercept of 0. The $y$ intercept for both lines is $\sim 0$, and the slopes for both VMI and red channel are close to 1, indicating that the VMI branching like the red vascular branching follows Murray’s law of the “minimum work” hypothesis successfully. A single branch from the two signals is also evaluated for more insights into smaller vessel branches, and both VMI and red vascular branching follows Murray’s law on smaller branches as well (the data are provided in Fig. S2 in the Supplementary Material).

Fig. 8

VMI follows Murray’s Law. The parent vessel diameter cubed is plotted against the sum of the diameter cubed of their corresponding daughter vessels. The data from red fluorescence of TdTomato-transgenic rat kidney are shown as red circles, and the data from vasculature extracted using the VMI technique are represented as green stars. The two vascular branching data from VMI and red fluorescence have merged, and their linear regression fit is compliant with Murray’s law (identity line, $y=x$).

3.3.

VMI of Organs from Partial Body Irradiated Rats

Here we present an application of VMI to uniquely drive the topography of two sets of parameters simultaneously: mitochondrial redox state and the 3D vascular network of whole organs. Figure 9 illustrates the representatives of the 3D rendered vascular networks of the kidney, liver, and lung from rats exposed to different doses of irradiation. The corresponding 3D RR (NADH/FAD) of the kidney and liver are also presented in Fig. 9. The RR images of the lungs are not presented due to the interference of FITC with FAD.

Fig. 9

Increasing doses of PBI resulted in increased severity of vascular-metabolic damage in (a) kidneys, (b) livers, and (c) lungs. The dose of irradiation decreases the RR (NADH/FAD). The RR images are presented in pseudo-color with higher RR voxels shown in red and lower RR voxels in blue. The mean RR is also provided by the images.

The vascular networks in Fig. 9 illustrate the regression of the vessel networks after PBI. The vascular damage in kidneys and lungs also appears to qualitatively correlate with the dose of irradiation in the PBI rats. The RR images are presented in pseudocolor with higher RR voxels shown in red and the lower RR voxels in blue. The kidneys and livers exposed to a higher dose of irradiation show a greater decrease in the RR, representing a more oxidized mitochondrial redox state.