2.1.

Cells and Cell Culture

B16F10 melanoma cells were obtained from ATCC (Manassas, Virginia) and cultured in DMEM medium with 10% FBS, 1% Pen-strep, and 1% sodium pyruvate.

2.2.

Generation of Integrin α10 Knockout Mice

To generate integrin α10 knockout mice ($\alpha 10\mathrm{KO}$), an integrin $\alpha 10$ targeting vector containing an ES cell clone in C57BL/6J background was developed by European Mouse Mutant Cell Repository.⁷² The integrin α10 gene function was inactivated by splicing of upstream endogenous exons to a splice acceptor in the targeting cassette located between exon7 and exon8. This allele reports gene expression by the lacZ-reporter and can function as a null mutation. The ES cell clone was expanded and microinjected into C57BL/6J blastocysts to generate a chimeric integrin $\alpha 10$ knockout founder line at the Mouse Transgenic & Gene Targeting Core of Maine Medical Center Research Institute (MMCRI). Genotyping PCR was performed using genomic DNA extracted from tail tips as templates. PCR primers used were common forward (5′-ACACACCTGTTCACTTCCCC-3′), wild-type reverse (5′-AAGGACCGCAATCCCATAAC-3′), and mutant reverse (5′-TAGAGTTCCCAGGAGGAGCC-3′). Homozygous integrin $\alpha 10$ knockout mice were born at the expected mendelian ratio and were viable and fertile without obvious gross abnormalities. Mice were housed in an MMCRI pathogen-free air barrier facility, and animal handling and procedures were approved by the MMCRI Animal Care and Use Committee.

2.3.

Generation of B16F10 Melanoma Tumors and Control Tissues

Melanoma tumors were established essentially as previously described.³¹ Briefly, wild-type C57BL/6J mice (6 to 8 weeks old) were injected subcutaneously with $3.5\times {10}^5$ B16F10 cells. Tumors were allowed to form for 14 days. At the end of the 14-day growth period, subcutaneous tumors with the associated skin were dissected and fixed with formalin. For nontumor stimulated mouse skin tissue preparation, wild-type C57BL/6J mice and integrin α10 knockout mice ($\alpha 10\mathrm{KO}$) were shaved to remove hair and full thickness murine skin was dissected. All tissues were washed in PBS and fixed with formalin. Tissue sections were prepared from tumors or normal skin and stained by H&E.

2.4.

Imaging

SHG imaging of H&E-stained slides ($10\mu \mathrm{m}$) was performed using a custom-built two-photon microscope using an upright microscope stand (Olympus BX50WI, Olympus, Center Valley, Pennsylvania) with a laser scanning unit (Fluoview300, Olympus) that is coupled to a mode-locked titanium sapphire femtosecond laser (Chameleon Ultra II, Coherent, Santa Barbara, California). Laser power was modulated by an electro-optic modulator (ConOptics, Danbury, Connecticut), operated in a power range of 3 to 10 mW at the focal plane using a LUMPlanFLN $40\times$ 0.8 NA (Olympus, Center Valley, Pennsylvania) water immersion objective. Circular polarization was used for SHG imaging; it was verified at the focal plane by rotating a polarizer and experiencing no change in laser power. All SHG imaging used 890-nm excitation, and the epi-SHG signal was collected using a 448/20-nm bandpass filter (Semrock Rochester, New York) in a nondescanned geometry using a H7421 GaAsP PMT (Hamamastsu, Hamamstsu City, Japan) using a LUMPlanFLN $40\times$ 0.8 NA objective (Olympus, Center Valley, Pennsylvania) with $2\times$ optical zoom ($180\mu \mathrm{m}$ field of view) with a $512\times 512\text{pixels}$ laser scanning speed of $2.71\mathrm{s}/\text{frame}$ with Kalman 4 averaging. Representative bright field images were acquired using a 5.1-MP MU 500 AmScope eyepiece camera using both LUMPlanFLN $20\times$ 0.5NA and LUMPlanFLN $40\times$ 0.8NA (Olympus, Center Valley, Pennsylvania) water immersion objectives. Figure 1 shows some sample bright-field images of H&E-stained slides from which SHG images were taken. Table 1 shows the number of mice used for each cohort and the number of individual XY SHG images taken from different regions of interest from each animal.

Fig. 1

Sample bright-field and SHG images of the tissue samples. (a), (e), and (i) Cartoon depiction of the three mouse phenotypes. (b), (f), and (j) Tissue stained for H&E using the $20\times$ objective; (c), (g), and (k) tissue stained using the $40\times$ objective. SHG images $40\times$ objective $2\times$ optical zoom (selection shown in the middle column) of skin from (d) a normal wild-type mouse, (h) an alpha10 knockout, and (l) a mouse with melanoma.

Table 1

Description of the number of animals and SHG images for each data category.

2.5.

2D WTMM Anisotropy Method

In this implementation of the 2D WTMM method, the continuous wavelet-transform (WT), acting much like a “mathematical microscope,” is used to quantify image intensity fluctuations. Use of the continuous WT allows us to investigate a continuous range of size scales. The WT of an image is the gradient vector (strength and direction of largest intensity variation) of the image smoothed by dilated versions of a Gaussian (smoothing) filter. Specifically, consider the wavelets,⁷³

At a given scale $a>0$, the wavelet transform modulus maxima are the specific locations $\mathbf{b}$ within the image, where the modulus $M_{\psi }[f](\mathbf{b},a)$ is locally maximum in the direction of the argument $A_{\psi }[f](\mathbf{b},a)$. At each size scale $a$, these WTMM are automatically organized as “edge detection maxima chains.”^51–53,55 Additional algorithmic details can be found in the Appendix of Ref. 65. The maxima chains for three different size scales are shown in green in Fig. 2.

Fig. 2

Flowchart of the 2D WTMM anisotropy method. (a) The original image is convolved with the wavelet functions at various size scales (three scales are shown here). (b) Positions in the image where the intensity gradient is locally maximum are given by the WTMM, which form edge detection contours. Each WTMM vector points in the direction of maximal gradient (highest modulus value). (c) The distribution of the angles of these vector directions is tabulated in pdfs, $P_a(A)$, from which the anisotropy factor, $F_a$ is calculated.

An image having an anisotropic signature means that the intensity variation in the image will differ according to the direction considered. These images can be easily characterized from the directional information provided, at all size scales $a$, by the probability density functions (pdfs), $P_a(A)$, of the angles, $A$, associated to each WTMM vector (see Fig. 2). A flat pdf indicates unprivileged random directions of sharpest intensity variation (i.e., isotropy), while any departure from a flat distribution is interpreted as the signature of anisotropy. For the study of SHG imaging of collagen, a strong anisotropic signature is interpreted as a strongly structured collagen fiber lattice.⁷⁴

2.6.

Anisotropy Factor, $F_a$

To obtain quantitative information from the angle pdfs $P_a(A)$, they are compared with a theoretical flat distribution representing an ideal isotropic signature. The anisotropy factor, $F_a$, defined for each value of the scale parameter $a$, is given by the area between the curve corresponding to the observed pdfs and a flat normalized distribution (i.e., a constant value of $\frac{1}{2\pi }$):

Therefore, a theoretically isotropic surface will have a pdf of $P_a(A)=\frac{1}{2\pi }$ and therefore an anisotropy factor value of $F_a=0$, while any value greater than 0 quantifies a departure from isotropy. Thus, this formalism objectively provides a quantitative assessment of morphological structure. A step-by-step explanatory diagram is shown in Fig. 2.

2.7.

Synthetic Images

Batches of synthetic images were created using a Matlab script previously described in Ref. 74. In each synthetic image, 50 simulated collagen fibers, represented as individual rectangles along the $y$-direction, were generated. In this work, the diameter of individual fibers was either 2, 4, 6, or $8\mu \mathrm{m}$, which is representative of the range of collagen fiber diameters found in SHG imaging studies.^5,9 The length of the individual collagen fibers was uniformly sampled from the range of 40 to $60\mu \mathrm{m}$. Within a single image, the positions of the fibers were uniformly sampled across the entire simulated $512\times 512\text{pixel}$ area modeling the SHG $180\times 180\mu \mathrm{m}$ FOV corresponding images. The angle of each fiber, $\theta$, was uniformly sampled from a given range $[{\theta }_{\min },{\theta }_{\max }]$. Nineteen such ranges of angles were considered, from high alignment, $[{\theta }_{\min },{\theta }_{\max }]=[85\mathrm{deg},95\mathrm{deg}]$ to no alignment, $[{\theta }_{\min },{\theta }_{\max }]=[0\mathrm{deg},180\mathrm{deg}]$, in increments of 5 deg (i.e., [85 deg, 95 deg], [80 deg, 100 deg], [75 deg, 105 deg], …, [0 deg, 180 deg]). Noise was simulated using a combination of a 2D circular averaging filter with radius of 3 and motion filters in Matlab.⁷⁴ Fibers were allowed to intersect in all cases. In total, 7600 images were simulated to calibrate the WTMM anisotropy method and to compare its performance with a Fourier-based angular amplitude approach: 100 images per angle range and for each diameter ($100\times 19\times 4=7600$). Sample synthetic images are shown in Fig. 3(a).

Fig. 3

Calibration analysis on groups of 100 synthetic images, each containing 50 simulated fibers with $2\mu \mathrm{m}$ diameters. (a) Four sample synthetic images of fibers with alignment angles randomly taken from the ranges [0 deg, 180 deg] (green), [30 deg, 150 deg] (blue), [60 deg, 120 deg] (brown), and [85 deg, 95 deg] (red), respectively. (b) Average pdfs, $P_a(A)$, with $a=2.5\mu \mathrm{m}$. The dashed vertical lines mark the bounds of the ranges of angles theoretically expected (i.e., perpendicular to the fiber angles used in the simulations). (c) Corresponding boxplots representing the distribution of anisotropy factors, $F_a$, with $a=2.5\mu \mathrm{m}$, for each group of 100 simulations. (d) Comparative anisotropy analysis of the same datasets using the Fourier-based method yielding the $R_{\text{Computed}}$ factor.

2.8.

Noise Experiment

To test the robustness of the method and its capacity to properly evaluate the strength of the anisotropy in noisy conditions, increasing levels of white noise were added to a subset of synthetic images. In Fig. 4(a), a sample synthetic image as described in the previous subsection with fiber angles uniformly taken from the range $[{\theta }_{\min },{\theta }_{\max }]=[60\mathrm{deg},120\mathrm{deg}]$ is shown. In Figs. 4(b)–4(f), increasing levels of white noise were added to the original image shown in Fig. 4(a) as follows: the mean pixel value from the original image in Fig. 4(a) was calculated. Then, a white noise image obtained through a white noise random simulator was generated with a mean pixel value that was a multiple of the original image. These multiples were 1, 2, 5, 10, and 20, respectively [Figs. 4(b)–4(f)]. For example, if the original image had a mean pixel value of 100, then the white noise image that was added to it in Fig. 4(f) had a mean pixel value of $100\times 20=2000$. With high levels of white noise added to the original image, the fibers are buried under the very high pixel values from the noise that is added to it.

Fig. 4

(a) Sample synthetic image of fibers with alignment angles randomly taken from the range [60 deg to 120 deg]. (b)–(f) White noise was added to the sample image in (a) at levels $1\times$, $2\times$, $5\times$, $10\times$, and $20\times$, respectively.

2.9.

Fourier-Based Angular Amplitude Measurements: $R_{\text{Computed}}$

FT-based approaches have been used by several research groups to characterize morphological properties of collagen fibers.^45,74 Recently, such a Fourier approach based on an angular amplitude measurement eliminated the need for fitting or thresholding, thus providing a fast and robust collagen fiber morphology assessment.⁷⁴ The Fourier-based angular amplitude measurement is based on the characterization of fiber alignment in the Fourier domain. In short, a 2D image in the real domain can be Fourier transformed into the spatial domain, and by considering the amplitude of the 2D spatial frequency coordinates, the degree of collagen fiber alignment can be visualized. Images with high alignment will have nonuniform 2D spatial frequency amplitude patterns around the origin, whereas images with low alignment will have more uniform 2D spatial frequency amplitude patterns around the origin. To quantify the alignment in 2D Fourier space without any fitting or thresholding procedures, the content of the 2D spatial frequency amplitude is integrated at every angle from 0 deg to 180 deg. The sum of all vectors is then normalized by angular projection of all of the fibers within an image, which yields a value referred to as $R_{\text{Computed}}$, where 0 is random alignment and 1 is perfect fiber alignment. For more detail, see Ref. 74.

3.1.

Calibration Analysis

Groups of 100 synthetic images each containing 50 simulated fibers with diameters of 2, 4, 6, and $8\mu \mathrm{m}$ and with ranges of angles as described in Sec. 2 were generated. The 2D WTMM anisotropy analysis of four of these groups of images with fibers with $2\mu \mathrm{m}$ diameters and ranges of angles of [0 deg, 180 deg], [30 deg, 150 deg], [60 deg, 120 deg], and [85 deg, 95 deg] is presented in Fig. 3. Sample images for each angle range are shown in Fig. 3(a). Figure 3(b) shows the pdfs, $P_a(A)$ at an $a=2.5\mu \mathrm{m}$ scale for each angle range (averaged over all 100 synthetic images for each group). Given that the WTMM vectors point toward the direction of largest gradient values (given by the modulus of the WT), their angles are perpendicular to the angle of the simulated fiber. Thus we observe very strong preferences for such perpendicular angles. For example, for the angle range [85 deg, 95 deg] (red), we expect the WTMM vectors to point preferentially toward the perpendicular, i.e., toward $[-5\mathrm{deg},5\mathrm{deg}]$ and $[-175\mathrm{deg},175\mathrm{deg}]$. These expected perpendicular angle bounds are shown in dotted vertical lines in Fig. 3(b), which strongly coincide with the stepwise behavior of the distributions for each of the four angle ranges considered. For each group of 100 synthetic images, the distribution of anisotropy factors, $F_a$, is shown as box plots in Fig. 3(c), which shows a clear separation of groups from each of the four ranges of angles shown. As a comparison, the Fourier-based approach, which yields the $R_{\text{Computed}}$ [Fig. 3(d)], did not discriminate between the four groups of ranges of angles as well as the wavelet approach did.

3.2.

Sensitivity Analysis

To further investigate the sensitivity of the 2D WTMM anisotropy method to discriminating between synthetic images of simulated fibers with different ranges of directional angles, 19 groups of different ranges of angles were analyzed, as described in Sec. 2. This deeper analysis, presented in Fig. 5 (for scale $2.5\mu \mathrm{m}$), provides further demonstration of the robustness of the wavelet approach. The 2D WTMM anisotropy method is capable of discriminating between groups of synthetic images with consecutive ranges of angles (e.g., between images with fibers taken from the range [0 deg, 180 deg] versus images with fibers taken from [5 deg, 175 deg], etc.). At a $2.5\mu \mathrm{m}$ scale, the method was able to provide statistically significant discrimination 17 out of 18 times for simulated fibers with diameters of 2, 6, and $8\mu \mathrm{m}$, and 18 out of 18 times for simulated fibers with diameter $4\mu \mathrm{m}$ diameters. In stark contrast, the Fourier-based method did not perform as well: 11/18, 12/18, 12/18, and 8/18 for diameters 2, 4, 6, and 8, respectively. To be fair, one should investigate whether the Fourier-based method might perform better or worse depending on the spatial frequency used. Nonetheless, to provide a more in-depth exploration of the sensitivity of the wavelet approach, a similar analysis as shown in Fig. 5 was done using 40 different size scales, from 2.5 to $39.4\mu \mathrm{m}$ (data not shown). Across these 40 size scales, the median number of consecutive angle ranges for which the 2D WTMM anisotropy method was able to successfully discriminate is: 17/18 for simulated fibers with diameters of 2, 6, and $8\mu \mathrm{m}$, and 18/18 for fibers with $4\mu \mathrm{m}$ diameters. This provides further evidence that, even if a Fourier-based method was used at different spatial frequencies, it seems unlikely that it would perform as well as the wavelet approach.

Fig. 5

Deeper anisotropy calibration analysis on groups of 100 synthetic images each containing 50 simulated fibers. Compared with Fig. 3, where four ranges of fiber angles were considered, here 19 ranges were considered, from high alignment ([85 deg, 95 deg]) to no alignment ([0 deg, 180 deg]) in increments of 5 deg (i.e., [85 deg, 95 deg], [80 deg, 100 deg], [75 deg, 105 deg], …, [0 deg, 180 deg]). The boxplots representing the distribution of anisotropy factors, $F_a$, for each group of 100 simulations (left column) allow us to assess the sensitivity of the method to discriminating between consecutive ranges of angles. Similar results from the Fourier-based method are shown in the center column. In the right column, for each consecutive pair of angle ranges, each point plotted (red for the 2D WTMM anisotropy method, black for the Fourier-based method) corresponds to the $p$-value obtained from running a nonparametric Wilcoxon rank sum test. A $p\text{-}\text{value}<0.05$ indicates that the method was able to discriminate between groups of synthetic fiber images of consecutive angle ranges in a statistically significant manner. Also included in these plots are the total number (out of a possible 18) of consecutive angle ranges for which each method was able to successfully discriminate. The scale used here is $2.5\mu \mathrm{m}$.

3.3.

Noise Experiment

As shown in Fig. 6, as long as we consider a large enough scale (keeping in mind that pure white noise is a small-scale phenomenon), the analysis becomes blind to the noise and still adequately captures the anisotropic signature caused by the alignment of the fibers. This is still true even when white noise is added at $20\times$ the amplitude of the original fiber image, albeit with a lower anisotropy factor than when there is less noise. Interestingly, when we calculate the anisotropy factor for pure white noise only (which is known to be isotropic, so its associated anisotropy factor is close to zero), it is very significantly different than the fibers with $+20\times$ amplitude white noise at large wavelet scales ($p\text{-}\text{value}<{10}^{-16}$ for scale $39.4\mu \mathrm{m}$), even though visually it is hard to see the fibers under the $20\times$ noise.

Fig. 6

Application of the 2D WTMM anisotropy method on seven groups of 100 simulated images, at scales of $a=2.5$, 4.9, 9.8, 19.7, $39.4\mu \mathrm{m}$. The seven groups of data correspond to the [60 deg, 120 deg] synthetic fiber simulations (no noise, shown in green), the five levels of white noise added to these synthetic images (shown in blue, brown, yellow, orange, and red; also see Fig. 4), and a white noise only group (shown in gray). (a) Average pdfs, $P_a(A)$. The dashed vertical lines mark the bounds of the range of angles theoretically expected (i.e., perpendicular to the fiber angles used in the simulation: [60 deg, 120 deg]). (b) Corresponding boxplots representing the distribution of anisotropy factors, $F_a$, for each group of 100 simulations and for each level of noise.

3.4.

Anisotropy Analysis of Mouse Skin SHG Images

Cohorts of images from the mouse skin SHG images listed in Table 1 were analyzed with the 2D WTMM anisotropy method. The box plots shown in Figs. 7(a) and 7(b) correspond to the anisotropy factors, $F_a$, for these three cohorts at scales of $a=2.5$ and $a=39.4\mu \mathrm{m}$, respectively, i.e., the smallest and largest scales considered. Drastically different anisotropy signatures are observed at these two extreme scales. Interestingly, there seems to be a somewhat monotone transition between these two extreme behaviors as a function of scale, as shown in Fig. 7(c), where all scales between $a=2.5$ and $a=39.4\mu \mathrm{m}$ are shown. To provide a relative comparison of the melanoma stimulated and the integrin $\alpha 10\mathrm{KO}$ skin collagen fibers when they are each compared with normal skin, the data in Fig. 7(c) were normalized by dividing every value by the value obtained for normal skin, as shown in Fig. 7(d). Also shown in Fig. 7(d) are the analogous normalized values for $R_{\text{Computed}}$, represented only as horizontal lines (red for $\alpha 10\mathrm{KO}$ and blue for melanoma) because only one spatial frequency was used with this method.

Fig. 7

Multiscale anisotropy analysis on SHG images of mouse skin: $\alpha 10\mathrm{KO}$ (red), normal (green), and melanoma (blue). Box plots representing the distribution of anisotropy factors for the three groups of skin samples at (a) the smallest scale considered ($2.5\mu \mathrm{m}$) and b) the largest scale ($39.4\mu \mathrm{m}$). All of the intermediate scales in between are shown in (c). To observe similarities and differences between the $\alpha 10\mathrm{KO}$ and melanoma skin images versus the normal skin images, the median values in (c) were normalized by dividing by the normal skin median values (d). Also shown in (d) are the analogous normalized values for $R_{\text{Computed}}$, represented only as horizontal lines (red for $\alpha 10\mathrm{KO}$ and blue for melanoma) because only one spatial frequency was used with this method. Exaggerated cartoons depicting proposed collagen fiber structures are shown in (e).

4.1.

Importance of Multiscale Analyses

Critically important observations can be missed when using a tool that only considers a single scale (or a single frequency, if using a Fourier-based method). Clearly, it would be impossible to observe the drastically different behaviors at different size scales that are presented in this study. For example, if a single-scale analysis had been performed on these data at an $a\sim 15\mu \mathrm{m}$ scale [Figs. 7(c) and 7(d)], then it would have been impossible to discriminate between the melanoma-stimulated and integrin $\alpha 10\mathrm{KO}$ skin collagen fibers. This would have forced the analyst to report inconclusive results. But perhaps even more importantly, suppose two different analysts were studying these data at two different individual scales, say analyst 1 at one given scale $a<10\mu \mathrm{m}$ and analyst 2 at a given scale $a>20\mu \mathrm{m}$. Then, the two analysts would have reached contradictory conclusions. Analyst 1 would conclude that melanoma-stimulated skin collagen fibers are statistically significantly more aligned (higher anisotropy factor $F_a$) than integrin $\alpha 10\mathrm{KO}$ skin collagen fibers and analyst 2 would report the exact opposite.

We hypothesize that the superiority of the 2D WTMM anisotropy method is due to the wavelet transform’s ability to analyze both space and scale (analogous to time and frequency for signal processing), compared with Fourier techniques being limited only to scale (frequencies). There are drawbacks that come with this additional analytical power associated with the WTMM method, namely algorithmic complexity and higher computation time, versus the simpler and faster FT.

4.2.

Biological Interpretation

Given the sensitivity of the 2D WTMM method to detecting specific differences in collagen fiber orientation over multiple size scales, this important technique allows for a more precise analysis of subtle changes in collagen organization that may play functional roles in governing cellular behavior in vivo. To this end, previous evidence suggests that collagen fiber spacing, density, and orientation play important roles in regulating tissue morphology, cell shape, and integrin-mediated signaling that collectively helps control cell behavior.^{10,26,76–79} For example, changes in collagen architecture have been shown to modulate integrin binding and promote both normal and pathological processes such as wound healing, angiogenesis, and tumor growth and metastasis.^{10,26,76–79} However, while considerable insight is available as to how collagen binding integrins can control gene expression,^24,25,29 much less is known regarding whether collagen-binding integrins play roles in regulating the orientation of collagen fibers. In this regard, while integrin $\alpha 10$ appears to play a role in regulating collagen network abundance in murine bone,³⁵ our new studies now suggest that $\alpha 10$ integrin may also contribute to the control of collagen fiber orientation, as distinct differences were observed between collagen fiber orientation within normal skin and skin from mice in which integrin $\alpha 10$ was knocked out.

Integrins including $\alpha 1\beta 1$, $\alpha 2\beta 1$ $\alpha 10\beta 1$, and $\alpha 11\beta 1$ are known to serve as functional cell surface receptors for collagen.²⁴ Integrin-mediated binding of collagen can result in activation of multiple downstream signaling events leading to activation of effector molecules that control cytoskeletal dynamics including small GTPases such as RhoA and Rac as well as the transcriptional coactivator “Yes Associated Protein” (YAP), which collectively regulate the ability of fibroblasts and melanoma cells to apply tensional forces to collagen, leading to contraction and reorganization of fiber orientation.^80–82 In fact, studies have revealed a potential role for the collagen-binding integrin $\alpha 11$ in controlling collagen orientation and contraction, as collagen fibers associated with breast tumors growing in transgenic mice lacking integrin $\alpha 11$ exhibited a more curved and disordered fiber orientation as compared with a straighter and more aligned collagen orientation found in mice that express this integrin.⁸³ While little evidence is currently available concerning the ability of integrin $\alpha 10$ to alter collagen fiber contraction, studies have indicated that integrin $\alpha 10$ can be expressed in fibroblasts along with alpha-smooth muscle cell actin ($\text{α}\text{SMA}$) which exhibit characteristics of a highly contractile phenotype.³⁰ Interestingly, studies have also indicated that inhibiting $\alpha 10$ integrin-mediated melanoma cell binding to denatured collagen can reduce the levels of active YAP³¹ and that reduced levels of activated YAP in fibroblasts can alter their ability to contract collagen.^81,82 Given these studies, it would be interesting to speculate that the changes in the orientation of the skin collagen fibers observed in mice expressing integrin $\alpha 10$ as compared with mice lacking $\alpha 10$ might be associated with the altered ability of fibroblast lacking $\alpha 10$ integrin to efficiently contract the skin collagen into a straighter and more aligned orientation. In addition, given that both melanoma cells and fibroblasts can express $\alpha 10$ integrin^30,31 and that these different cell types may bind to collagen at different positions on the fibers, along with studies indicating that these different cell types exhibit distinct levels of collagen contraction,⁸⁴ when taken together, these parameters may contribute in part to the differential orientations observed between collagen from normal skin and collagen from skin stimulated with melanoma over the size scales studied.

The exact mechanism by which integrin $\alpha 10$ regulates collagen fiber orientation is currently not known, and further studies will be needed to elucidate the mechanisms that contribute to this surprising observation. However, developing more precise methods and strategies to accurately identify, characterize, and quantify changes in collagen orientation over multiple size scales in vivo may not only provide a better understanding of how integrins contribute to changes in collagen structure but also how biophysical changes in collagen organization in vivo regulate cellular behavior. Taken together, our studies, in conjunction with existing insight into how collagen orientation contributes to physiological and pathological processes, may ultimately not only help identify new therapeutic strategies for the treatment of diseases controlled by ECM remodeling but also help optimize the use of engineered collagen containing tissue scaffolds for multiple applications.