2.1.

Hemisection and Contusion Injury

The Department of Laboratory Animal Resources and the Committee for the Humane Use of Animals at SUNY Upstate Medical University approved this study following Association for Assessment and Accreditation of Laboratory Animal Care guidelines. Adult, female Sprague-Dawley rats (average weight 255 gram) were used for all experiments. Rats were anesthetized with Ketamine and Xylazine (80 and 10 mg/kg, respectively) and body temperature was maintained using a heating pad. Aseptic surgery was performed using the following technique. The skin over the upper thoracic area was shaved and cleaned with a Betadine solution. The skin was incised, and then the connective and muscle tissue were bluntly dissected to expose the thoracic ninth (T9) vertebral body. A T9 laminectomy was completed, taking care not to damage the spinal cord during removal of the dorsal lamina. A lateral hemisection was performed at T9; initially a surgical needle punctured the spinal cord dorsoventrally at the midline avoiding the dorsal spinal artery; angled microscissors were then used to cut the right half of the spinal cord followed by an angled needle scraping the vertebrae ventrally and laterally surrounding the lesion to ensure completeness of the hemisection. Finally, the lesion was closed in layers with individual sutures. A schematic of the hemisection surgery is shown in Fig. 1. For the contusion injury, rats were surgically prepared as described. After exposing the spinal cord at T9, animals were injured using the Infinite Horizons (IH) impaction device (Precision Systems and Instrumentation, LLC). The details of the method of operation of the IH impactor can be found elsewhere.¹¹ Briefly, a contusion injury to the exposed spinal cord was achieved by rapidly applying a force-defined impact with a stainless steel-tipped impounder. The impounder, controlled by a stepping motor, applied a user-defined force to the spinal cord (200 kDyne). This load has previously been known to produce a moderate injury. ^{6, 11}

Fig. 1

A schematic of the hemisection surgery. Also shown is a transverse cross section of the hemisected spinal cord. The light micrograph shows an injured spinal cord. The arrow points toward the injured portion. A representative cross section of the injury site is also shown.

The animals were sacrificed at various time points, post-injury. Injured spinal cords were excised using a posterior approach. Euthanasia, by perfusion of deeply anaesthetized rats with physiological phosphate buffered saline (PBS), was followed by a laminectomy between the first cervical vertebra (C1) and the fourth lumbar vertebra (L4). The nerve roots were carefully severed and the spinal cord was cut at levels C1 and L4, then carefully removed and placed in a bath of isotonic PBS. No attempt was made to remove the dura mater. The cords were then stored in a refrigerator at 4°C until testing. All testing was performed within 4 h of excision of the spinal cords. A total of 30 animals were used. Six animals were used for each time point post-injury for the hemisection model (n = 18 total). Six animals were used for the contusion group, and sacrificed, two weeks post-SCI, and six animals were used as controls. The control animals did not receive any injury and the excised spinal cords were termed as healthy. Two spinal cords from the 4 day post-injury group were not used due to (a) the laser burning the cord due to excess blood and (b) one cord was too fragile and proved difficult to excise in its entirety.

2.2.

Chondroitin Sulfate Solutions and the Effect of cABC

Chondroitin 6-sulfate (CS6, Sigma Aldrich) was used for in vitro experiments. The glycosaminoglycan powder was used as received. To study the effect of cABC (Seikagaku Corporation) on CS6, 100 μl of cABC solution (2 Units/ml) was added to a 5 ml solution of 5mg/ml CS6 at room temperature and spectra were collected in a specially designed quartz cuvette at 20-min intervals for a period of 1 h.

2.3.

Raman Spectroscopy

A custom built Raman spectroscope was used for this study.¹² The laser wavelength was 785 nm (Process Instruments, Salt Lake City, Utah) and spectra were collected in backscatter mode. The laser power at the sample was 450 mW and the spot size was 300 μm determined using Zap-it™ paper. A fused silica cuvette was employed for in vitro experimentation on liquid samples. The dispersed light was detected using a Roper Scientific/PAR CCD camera and a Kaiser f/1.4 Holospec spectrograph. The spinal cord was oriented similar to its orientation in vivo with the dorsal side facing up. The cord was placed in a specially designed quartz holder that was placed on a 3-D stage controlled with micrometers and care was taken to keep the spinal cord hydrated with PBS during the course of an experiment. For hemisected cords, spectra were obtained focusing on the scar (12 spectra, 5 min each). For contused cords, as shown in Fig. 2, spectra were sequentially obtained beginning at an uninjured region at the rostral end of the cord and moving through the injury epicenter toward uninjured regions caudally (5 min per spectrum, spectral locations spaced 500 μm apart). For healthy cords, spectra were obtained from random locations on the cord (at least ten spectra per cord, 5 min each). Further, spectra were also obtained focusing on the injury epicenter (12 spectra, 5 min each). All experiments were done at room temperature.

Fig. 2

Schematic of spectrum collection from a contused spinal cord. Spectra were collected moving from healthy uninjured portions at the rostral end through the glial scar, back to uninjured regions at the caudal end.

2.4.

Data Processing and Analysis

The separation of the fluorescence and Raman contributions to the observed Stokes shifted emission from near-infrared (NIR) excited ex vivo or in vivo tissue samples cannot be deterministically executed, unless it is actually known how much of each emission is present. As is often the case with biological samples, this information is not available. The primary goal in this study was to determine if healthy spinal cord can be discerned from injured spinal cord using Raman spectroscopy so an approach that can be potentially automated is utilized. Further, this approach is statistically unbiased in the sense that only deterministic, well-characterized, and reproducible artifacts are induced. ^{15, 16} This allows for comparison of spectra of different samples to each other visually and quantitatively, using statistical measures of differences such as integrated peak areas, in order to deduce whether changes occur due spinal cord injury.

Unless the fluorescence and Raman contributions can be separated by some independent method, all baseline correction procedures ^{15, 16} for biological samples contain assumptions and therefore induce distortions. For example, all procedures assume that the shapes of fluorescence spectra are featureless with respect to having relatively narrow peaks, e.g., choosing arbitrarily ≈200 cm⁻¹ or less. To the extent that Raman features, including congested clumps of Raman features, are significantly less than 200 cm⁻¹ wide, the smoothing procedure used produces a baseline that roughly bisects only the Raman contribution to the observed spectrum above the contribution formed by the sum of the fluorescence component and any Raman contribution that cannot be resolved due to spectral congestion. This is true because the wide smoothing window we use is 101 points wide corresponding to ≈200 cm⁻¹. Thus in general, features that are much narrower than 200 cm⁻¹ are reproduced in the baseline corrected spectrum without distortion. To the extent resolved Raman features are roughly 200 cm⁻¹ wide and wider, the 101 point smoothed spectrum obtained from the raw data contains a contribution from those features and so their contribution will be decreased in, i.e., be subtracted from the raw data to obtain, the ultimate baseline corrected Raman spectrum. The baseline corrected spectrum thus contains only contributions from the resolved Raman spectrum without making any unjustified assumptions about the fluorescence spectrum or unresolved Raman components. However, the position of zero counts is completely unknown and the baseline corrected spectrum has positive and negative excursions from zero. One can properly correct for this when integrating peak areas by a proper choice of zero taken from the wings of the peak as described below.

Note that the cropping of the low Raman shift end of the spectrum precludes the net spectral shape of the fluorescence and the laser line rejection filter from making any contribution in shape or strength to the baseline corrected spectrum up to at least ≈1200 cm⁻¹ Raman shift. Interestingly, relatively strong and wide features, e.g., amide I (1670 cm⁻¹) and CH₂ deformation (1450 cm⁻¹), separated by a couple hundred cm⁻¹, also make a contribution to the smoothed baseline and so these two features are also distorted, i.e., decreased by roughly 10% relative to all the other Raman features in the baseline corrected spectrum. The large clump that includes amide III (≈1200–1300 cm⁻¹) is not affected as much but it certainly is systematically decreased.

Indeed we have previously shown ^{17, 36} that essentially the same “quantitative” results can be obtained (from spectra of phantoms, tissues in vivo, and simulations having very similar characteristics as those presented in this manuscript) with or without the same 101–7 smoothing baseline correction described below. For our spectral resolution, 101 points is sufficiently wide and in observing the spectra with the naked human eye, the relative heights of Raman features are compared with each other essentially as an implicit internal standard and it is these comparisons that have been highlighted, i.e., not “absolute” variations. Also, the seven point smoothing portion of this procedure (after spike removal) has no effect whatsoever.

Using Matlab^® custom written routines, spectra were cropped in the fingerprint region (400 –1800 cm⁻¹) and were corrected for background fluorescence using an arbitrary but unbiased 101-point moving window-averaging scheme. ^{12, 17, 18, 19} This 101-point moving average was subtracted from the raw data and these background corrected spectra were smoothed using a seven-point moving window algorithm and where noted, spectra were standard normal variate (SNV) transformed by mean centering and dividing by the standard deviation. ^{13, 14} Fluorescence in the fingerprint region was calculated by integrating the raw counts to obtain the area under the curve and assuming the contribution of the Raman to the total emission is negligible. ^{12, 18, 20} In order to make quantitative comparisons, seven distinct Raman peaks were indentified in the spectra and areas under each of the peaks were calculated from the baseline corrected, SNV transformed data. For quantitative analysis each of the peaks was assigned an individual baseline¹⁷ by subtracting the minimum intensity, i.e., from the wings of the peak in order to avoid negative values and subsequently summing the intensities across the wavenumbers defining the peak to obtain an area measure. The SNV ^{13, 14} transform was used to normalize for baseline shift and gain differences. Currently, we cannot discern that it produces any modification to the spectra or to the relative strength of the Raman features at all. The purpose of using the SNV transform is to allow averaging of spectra taken at different times from comparable samples with equal statistical weighting.

A one way analysis of variance (ANOVA) was performed followed by a Tukey's least significant difference (LSD) post-hoc test to determine differences in peak areas between groups with 95% confidence. Further, as an independent measure of classifying the spectral peak areas, we implemented the k-nearest neighbor (KNN) algorithm in Matlab. Data were arranged as peak areas corresponding to the injury type (healthy, hemisected four days, two weeks or eight weeks PI, and contused cords). The training set was 50 randomly selected peak areas from the different injury conditions (healthy, hemisected, or contused) and the classification set was the remaining dataset. The algorithm based all classifications on a k-value of five nearest neighbors. This process was repeated 5000 times to obtain an average classification rate.

2.5.

Immunofluorescence and Histological Staining

A total of 16 animals (four per group) were used for immunohistochemical and histological staining including healthy/uninjured controls, and four days, two weeks, and eight weeks after spinal hemisection injury. Animals were euthanized with an intraperitonial injection of sodium pentobarbital (50 mg/kg), followed by transcardial perfusion with PBS (500 ml) prior to perfusion with 4% paraformaldehyde in PBS (500 ml). Excised spinal cords were left overnight in 4% paraformaldehyde and then placed in 20% Sucrose solution in PBS for two days. Next an 8–10 mm segment of each spinal cord including the injured region, were frozen in optimal cutting temperature media and 20-μm serial sections were cut in the transverse plane, using a cryostat. Series of serial sections separated by 100 μm were mounted on glass slides. One series was immune-stained for CSPGs (CS-56, mouse monoclonal antibody for intact CSPGs, 1:100, Sigma C8035). This series was first incubated in blocking media (10% normal goat serum, 3% bovine serum albumin in PBS, pH 7.4) overnight. Next, sections were incubated with a secondary antibody [AlexaFluor^® 555 (red) IgM goat anti-mouse, Invitrogen Inc.] for one hour at room temperature and stored at 4°C until imaged. A histological stain for myelin using the eriochrome cyanine staining technique was also employed on a series of sections. Both fluorescent and brightfield imaging was done using a Zeiss Axioskop microscope (Carl Zeiss, Oberkochen, Germany). Digital images were captured using a Spot RT Slider camera (Diagnostic Instruments, Sterling Heights, MI) and the accompanying Spot Advanced software (v.3.3.4 for Macintosh, Diagnostic Instruments).

3.1.

Raw Data of Healthy and Injured “Hemisected” Spinal Cords and Variation in Fluorescence

Figure 3a shows the unprocessed data obtained from various locations on a single healthy spinal cord and Fig. 3b shows the representative time course on a single location of a particular cord for successive 5-min CCD acquisitions, respectively. The integrated fluorescence values are shown on the ordinate. It can be seen that there is background fluorescence but the Raman peaks are evident and that the relative magnitude of these peaks is similar (for each injury condition), irrespective of the fluorescence, although eight of ten locations on the healthy cord produced data within ±3% of the average. The other two locations produced a large bivariate deviation from average and the variation is likely due to uneven drying of the tissues, the presence of surface wrinkles or creases, or the penetration of the excitation laser through the tissues to the quartz holder or some combination of one or more of these factors.

Fig. 3

(a) Raw data from various locations on a single spinal cord. Also shown on the right are integrated fluorescent counts for the raw spectra. Note that although background fluorescence varies between locations, the Raman features are fairly constant. Exposure time per location was 5 min. (b) Raw data from extended irradiation of a single point on a spinal cord. Integrated fluorescence is shown on the right and photobleaching of the fluorescence is apparent. Note that the Raman features are similar irrespective of the fluorescence. Exposure time per spectrum was 5 min.

For most of the data, the fluorescence variation with time is commensurate with the variation observed by sampling spectra from different locations on the same cord. Further, it is also clear that the spectra contain Raman features that are relatively constant with respect to the fluorescence, indicating that the location did not affect the relative Raman spectral features. In order to see this more clearly, Fig. 4a shows the average and the 95% confidence intervals of all of the raw spectra taken either on healthy tissue or different locations within the SCI zone for each injured cord at each time. The fluorescence for the healthy cord was substantially larger (p<0.005) and spectrally narrower than for any of the injured rats. The fluorescence of the injured cords monotonically decreased with time post-injury but no significant difference in fluorescence between injured cords was observed. This is shown in Fig. 4b.

Fig. 4

(a) Raw data for the hemisected and healthy spinal cords. Data are plotted as the mean raw data obtained from all animals per group and the upper and lower 95% confidence interval. Note the different scales on the ordinate. (b) Fluorescence measures for each group after hemisection injury. Data are plotted as mean ± standard deviation. Asterisk (*) indicates that the healthy cords had significantly higher fluorescence (p<0.005) than injured cords. No significant differences in fluorescence between injured cords were noted.

3.2.

Raman Spectra of Healthy and Hemisected Spinal Cord at Different Times Post-injury

Figure 5 shows the mean baseline-corrected, smoothed, and transformed spectra of control and hemisected spinal cords. Although these data are averages over many animals, the variability in the data was small and very systematic variation could be perceived. Certain features such as the 1450 cm⁻¹ CH₂ deformation mode and the 1670 cm⁻¹ amide I modes were very prominent in all the spectra. The peak at 700 cm⁻¹, corresponding to cholesterol, decreases at four days PI and is absent at two and eight weeks PI, in comparison with the healthy cords. Important peaks are highlighted and peak assignments are listed in Table 1. In order to clearly see the variation, we produced difference spectra using the healthy cord as reference (Fig. 6), which revealed very systematic variations that were often, but not always, monotonic. We observed in Fig. 6 that relative to the healthy spinal cord tissue, the tissue in and around injured spinal cord showed a decrease in the peak intensity at 700 cm⁻¹ (cholesterol), and monotonic decreases in the peaks at 1005 cm⁻¹ (white matter, phenylalanine breathing mode in proteins), 1065 and 1086 cm⁻¹ (C–C stretching and PO₂ stretching of lipids and phospholipids), 1299 cm⁻¹ (CH₂ twisting and wagging), and 1450 cm⁻¹ (CH₂ deformation from lipids and proteins) features and a monotonic increase in the peak intensities at 490 and 640 cm⁻¹ (CSPGs and GAG aggregation). This indicates that the tissue damage and the subsequent wound healing response post-SCI, led to a change in the concentration of the biomolecules that contributed to these spectral peaks. Other features, notably near 800–1000 cm⁻¹ and above 1500 cm⁻¹, had quite complex behavior with respect to time post-injury. Uninjured rats were used as opposed to sham-operated rats as controls. This was done in order to compare uninjured cords from healthy animals to cords from injured animals. Also, the injury model used was not a novel model of SCI. It was further observed, that spectra from uninjured portions of healthy cords showed features similar to those from control animals (data not shown), thereby precluding the use of sham animals.

Fig. 5

The mean background corrected, SNV-transformed spectra of hemisected and healthy spinal cords. Important bands are highlighted in gray. Please refer to Tables 1, 3 for peak assignments.

Fig. 6

Difference spectra of the hemisected cords using healthy cords as a reference. Important peaks are highlighted with gray bands. Note that there appear to be systematic changes that occur with progression of injury. Important changes corresponding to GAGs and proteoglycans (450–540 cm⁻¹, 541–690 cm⁻¹, and 940–995cm⁻¹), demyelination (lipid loss, 1398–1519 cm⁻¹), and carbonyl chemistry (above 1650 cm⁻¹) are highlighted. Please refer to Tables 1, 3 for peak assignments.

Fig. 9

Time course of cABC digestion of chondroitin 6-sulfate in PBS at room temperature. Spectra were obtained at intervals of 20 min for up to an hour. The spectrum of the CS-6 solution is also shown. Note the evolution of the peaks around 1750 cm⁻¹ upon digestion with cABC.

Table 1

Peak assignments to Raman spectra of spinal cord tissue. Question marks indicate unassigned peaks.

Table 3

Classification matrix output from KNN algorithm. The numbers indicate percentage of correct classification.

3.3.

Immunofluorescence and Histology of Hemisected Spinal Cords

Figure 7 shows the results from the immunofluorescent staining of spinal cord sections for CSPGs (top panel) and the histological staining of sections for myelin (bottom panel). The immuno-staining for CSPGs, indicates that the intensity of labeling is maximal at two weeks PI, but is above background levels at four days and eight weeks PI relative to uninjured controls. Myelin staining indicate that at four days PI, demyelination is present but is incomplete on the side of injury. By two weeks PI, there is intense demyelination on the ipsilateral side of injury, which continues at eight weeks PI. The Raman spectra complement the results of the histological and immunohistochemical assessment of the post-injury lesion site.

Fig. 7

Representative Immunohistochemical photomicrographs of healthy and hemisected spinal cords. Photomicrographs are temporally arranged from left to right as healthy, four days, two weeks, and eight weeks PI. Top row: Eriochrome cyanine stain for myelin. Healthy spinal cord stains blue and the gray and white matter are distinctly different. At four days PI, demyelination is evident but not complete, at two weeks PI, there is complete demyelination on the ipsilateral side of injury which continues at eight weeks PI. Bottom row: CS-56 immunofluorescent stain for CSPGs. The healthy cords have a uniform background staining intensity. At four days PI, there is a slight increase in intensity of staining. At two weeks PI, the intensity of staining is clearly maximal on the ipsilateral side and at eight weeks PI, there is increased intensity in comparison with controls, but not as intense as the two weeks PI staining.

3.4.

Raw Data of Healthy and Injured “Contused” Spinal Cords and Variation in Fluorescence

Spectra from contused cords showed reduced fluorescence in comparison to healthy cords, similar to the spectra obtained for hemisected cords. Upon scanning across the spinal cord from uninjured portions, through the injured glial scar locations, back to healthy uninjured regions, the fluorescence decreased in the injured regions and increased back to normal levels with greater distance from the lesion site (data not shown).

3.5.

Raman Spectra of Contusion Injured Spinal Cords

Figure 8a shows the mean baseline-corrected, smoothed, and transformed spectra of data obtained from the contusion epicenter. In order to see the variation with respect to uninjured controls, we produced difference spectra using the healthy cord as reference [Fig. 8b], which revealed difference spectra very similar to hemisected cords. Figure 8c shows the Raman spectra from scanning across a contused cord from healthy to injured to healthy regions. Important peaks are highlighted and peak assignments are listed in Table 1. It can be observed that spectral features considerably change as one approaches the injury site and return to the appearance of uninjured controls as one moves to uninjured regions. Figures 8d and 8e show the changes in peak areas corresponding to spectra in Fig. 8c. The peak areas corresponding to GAGs and lipids are shown and it was noted that at the injury site, the peak areas for GAGs increase whereas those corresponding to lipids decrease.

Fig. 8

(a) Mean SNV transformed spectrum of contused spinal cords. (b) Difference spectrum of contused cords using healthy cords as a reference. (c) Raman spectra from a scan across a contused spinal cord. Rostral is top and caudal is bottom. Spectra in black are from healthy uninjured regions and those in red are from the scar site. Important bands are highlighted in gray. Note the change in spectral features around the 400–800, 940–1000, and 1400–1500 cm⁻¹ regions. Please refer to Tables 1, 3 for peak assignments. Peak areas corresponding to (d) aggregated CSPGs and GAGs and (e) lipids from a scan across a contused spinal cord. Scan numbers 1–8 and 17–27 are on uninjured locations and scan numbers 9–16 are on the scar site. The distance between each scan location was 500 μm. For GAGs and CSPGs, the peak area increases upon approaching the scar and for lipids, decreases.

3.6.

Raman Spectra of Chondroitin Sulfate Glycosaminoglycans

Figure 9 shows the room temperature time evolution of the PBS subtracted, baseline-corrected Raman spectrum of chondroitin-6 sulfate in PBS (pH 7.4), treated with cABC (2 Units/ml). Chondroitin sulfate consists of a long polysaccharide chain with free carboxylic acid, sulfate, and methyl amide groups as well. The spectrum of the undigested material displays well-known features at 600 cm⁻¹ and below, corresponding to ring vibrations of the individual saccharides, and at 1064 cm⁻¹, corresponding to the sulfate group. In addition, we observe overlapping Raman features near the anomeric carbon linkage mode near 800–900 cm⁻¹, as well as a variety of features about 1600 cm⁻¹ corresponding to the amide linkages and carboxylic acid carbonyl localized modes. ^{10, 21, 23}

3.7.

Peak Area Quantification and Classification using k-Nearest Neighbor Algorithm

Peak areas were calculated as described in Sec. 2 and the results of the one-way ANOVA, followed by a Tukey's LSD post-hoc test are shown in Table 2. The groups having the same letters are not significantly different. As shown in Table 2, peak 3 (1210–1396 cm⁻¹) and peak 5 (1021–1113 cm⁻¹) showed no significant differences. However, many significant differences between groups were observed for the remaining peaks. Peaks corresponding to healthy cords were significantly different than injured spinal cords although fewer significant differences within injury groups were noted. The KNN algorithm was used as an independent measure to classify the spectra and the resultant classification matrix from the KNN algorithm is shown in Table 3. The overall true classification rate observed was 83%. The classification matrix indicates the percentage of correctly classified data sets. For example, reading across the second row of Table 2, 97% of the healthy cords were correctly classified as healthy and the remaining 3% were incorrectly classified as hemisected cords. The four-day hemisected cords had a low classification rate of 48%, whereas the two-week hemisected and contused cords had classification rates of 94 and 96%, respectively. The eight-week hemisected cords had a classification rate of 63% but importantly, the percentage of injured cords that got classified as uninjured was extremely low for the two- and eight-week cords, with the four-day cords having a false classification of 12%.

Table 2

Peak widths for quantitative calculations and results of post-hoc testing. Groups with the same letters are not significantly different (p>0.05).