1.

Introduction

Optical technologies capable of real-time quantitative analysis of tissue structure and biochemical composition have the potential to revolutionize medical diagnostics. Histology is the gold standard for clinical diagnostics, but it can often involve painfully invasive biopsies, long turn-around times, and qualitative decision making. Established noninvasive imaging technologies, such as computed tomography, magnetic resonance imaging, and ultrasound, offer faster, reliable, and easily interpreted anatomical data with some compositional information. However, these modalities suffer from a variety of limitations including long acquisition times, ionizing radiation exposure, poor contrast between soft tissue types, the need for use of exogenous contrast agents, and large and expensive hardware.^1–3 Alongside these methods, a range of optical techniques that utilize custom-engineered exogenous fluorescent markers of disease have emerged; these include small molecule organic fluorophores,⁴ quantum dots,⁵ carbon nanotubes,⁶ and various other nanomaterials.^7,8 However, considerable scope remains for developing minimally invasive spectroscopic devices [e.g., based on autofluorescence (AF) or diffuse reflectance (DR)] that can be used in the absence of specific probes or dyes.

AF spectroscopy involves exposure of tissues to monochromatic ultraviolet (UV) or visible light and the collection of the Stokes-shifted spontaneous emission spectrum that originates from endogenous fluorophores. Naturally fluorescent biological molecules include a variety of amino acids, structural proteins, enzymes, coenzymes, vitamins, lipids, and porphyrins, each of which possesses distinctive excitation/emission spectra.^9,10 Their distributions vary between tissue types, and their individual fluorescence properties can be affected by their molecular environment,¹¹ pH,¹² and temperature,¹³ making AF spectroscopy a powerful diagnostic tool. This may be complemented by DR spectroscopy, which uses broad-spectrum illumination and detection of diffusely reflected photons over the same wavelength range. DR spectroscopy is sensitive to absorption by endogenous chromophores, including hemoglobin and myoglobin,¹⁴ as well as light scattering from micron-scale structures, such as cell membranes, nuclei, and mitochondria.¹⁵

In this study, we aimed to investigate the utility of AF/DR spectroscopy in the context of muscle disease. Myopathies represent a major disease burden in which early diagnosis can be immensely helpful in terms of future clinical management. Myopathies have numerous characteristic histological features that are currently difficult to detect in situ and can only be reliably identified using invasive biopsies. Evidence of a congenital myopathy can be seen in the presence of nemaline rods,¹⁶ fiber-type disproportion,¹⁷ intramyocellular lipid,¹⁸ Z-line streaming, and sarcomeric disruption.¹⁹ Muscular dystrophies often involve adipocyte or fibrotic invasion, the presence of which can correlate with progressive muscle weakness.²⁰ Both quantitative and qualitative histology techniques exist;^20,21 however, both are time consuming, and biopsy collection is associated with tissue morbidity, pain, infection risk, long turn-around times, and error-prone logistics.²²

There is reason to expect that a spectroscopic approach will be effective in characterizing muscle tissue. In skeletal muscle, unique AF signatures of individual myofiber types have been demonstrated, enabling classification of whole muscle types with unique heterogeneous myofiber compositions.^23–25 Differences in their metabolic biochemistry are thought to alter the fluorescence properties of mitochondrial fluorophores including nicotinamide adenine dinucleotide (NADH) and flavin adenine dinucleotide (FAD), as well as some essential amino acids including tyrosine and tryptophan. This technique may also be applicable to detecting changes in muscle composition with disease and aging. Nakae et al.²⁶ demonstrated the accumulation of fluorescent lipofuscin granules from oxidative breakdown of cellular macromolecules in human and murine dystrophic and aged myofibers detectable with visible wavelength AF microscopy. Further, the sensitivity of AF spectroscopy to muscle fat and collagenous tissues^25,27 may be relevant to the diagnosis of muscle diseases for which intracellular adiposity and fibrosis are phenotypes.^18,28

A current barrier to widespread use of this technique is a lack of economical, accessible, and easy-to-use platforms that integrate AF/DR spectroscopy with relevant and easily interpretable spectral analyses. In principle, such an approach is feasible and supported by a range of studies characterizing tissue AF in the gut, arteries, and myocardium^29–31 and in tumors.^32–35 AF spectroscopy has also been employed in conjunction with DR spectroscopy in a range of settings.^36–39 Athough there is currently a lack of information on the use of AF/DR spectroscopy within the specific context of diseased muscle tissue, AF/DR probes have been developed to identify commercial meat cut quality based on intermuscular collagen and adipose content.^14,25,27

In this study, we describe a fiber-optic AF/DR system and speculate that both modalities may be capable of real-time marker-free classification of healthy and diseased skeletal muscle tissues. The system was trialed using several established preclinical models of muscle pathology. The mdx mice are a widely used model of muscular dystrophy that feature a mutation in the murine dystrophin gene.⁴⁰ Like in the clinical scenario, mdx mice exhibit a refractory period in which they are asymptomatic before developing muscle weakness. The $Nf{1_{Prx1}}^{-/-}$ mouse line is a limb-specific knockout mouse line in which the NF1 gene is deleted in mesenchymal tissues including muscle.⁴¹ NF1 is critical for muscle development and function, and NF1-deficient muscle features intramyocellular lipid accumulation and fibrosis.^18,28 Data preprocessing and classification techniques were first optimized with the goal of classifying optical measurements from different muscles groups in healthy wild-type mice before these techniques were applied to the classification of muscle tissues from mouse models of muscle pathology. These results were compared with standard histological techniques for muscle assessment.

2.

Materials and Methods

2.1.

Optical System

A custom AF/DR system was built using commercially available components. A 200-mW 405-nm laser diode (LD) was used for AF measurements, coupled via a servo-actuated shutter, bandpass filter (FB405-10, Thorlabs), and microscope lens into one of seven multimode optical fibers ($200\text{-}\mu \mathrm{m}$ core diameter and 0.39 NA) within a fan-out style bundle (BF72HS01, Thorlabs) terminating distally in a six-around-one configuration [Fig. 1(a)]. The shutter allowed the laser power to ramp-up and stabilize before each measurement, while illumination power, set to 20 mW, was controlled by adjusting the coupling between the fiber and LD. Tissue AF emission was collected using the bundle’s central fiber and long pass filtered at 425 nm (DMLP425, Thorlabs) to remove the illumination wavelength before detection by a USB spectrometer with a wavelength range of 345 to 1041 nm (USB 4000, Ocean Optics).

Fig. 1

(a) Schematic of the fiber-optic AF/DR system. 405-nm laser (violet) and halogen lamp (yellow) light sources were coupled into proximal ends of two multimode (MM) optical fibers via servo-mechanical shutters and a bandpass (BP) filter to further narrow the AF excitation wavelength range. Tissue DR and AF light was collected via a central fiber at the probe’s distal end (red) and passed through a 425-nm long-pass (LP) filter before detection at the spectrometer. Additional fibers within the fiber-optic probe (white) were unused. (b) The fiber-optic probe head was positioned directly onto the muscle surface during optical measurements, here showing a DR measurement on the murine gastrocnemius.

A stabilized tungsten–halogen light source (SLS201L/M, Thorlabs) was used for DR measurements, coupled via a servo-actuated shutter and microscope lens into a fiber adjacent to the 405-nm LD [Fig. 1(a)]. Diffusely reflected light was collected using the same central fiber as for AF measurement and returned to the same spectrometer via the same optical path. The four remaining fibers within the bundle were unused. System control and data collection were performed on a PC laptop using MATLAB (Mathworks) via direct USB connection to the filter wheel and spectrometer. An Arduino was used for transistor–transistor logic and pulse-width modulation signals for control of the LD and shutter servos, respectively.

3.

Results

3.1.

AF and DR Spectra Enable Separation of Different Muscle Types in Wild-Type Mice

It was hypothesized that different muscle groups that contain different fiber types and cytoskeletal proteins would produce different AF and DR spectra. Measurements were made from several murine hind limb muscles (gastrocnemius, quadriceps, and tibialis anterior) as well as from cardiac muscle from the same animals. Statistical analysis of normalized spectra was applied to generate 95% confidence intervals (CIs), which indicated the separability of each muscle group at each detection wavelength. For AF measurements, overlapping CIs were observed across all wavelengths for skeletal muscle groups, whereas cardiac muscle showed significantly different spectral shape to skeletal muscles across the 500 to 600 nm range [Fig. 2(a)]. For DR measurements, overlapping CIs were observed across all wavelengths for all skeletal muscle groups, whereas cardiac muscle was separable from only quadriceps and gastrocnemius groups at $\sim 560\mathrm{nm}$, and from tibialis anterior only between 630 and 895 nm [Fig. 2(e)].

Fig. 2

(a)–(d) AF and (e)–(h) DR spectroscopy enabled classification of muscle types (Quad, Gastr, TA, and heart) in healthy B6 mice using PCA. (a), (e) Average normalized spectra (solid lines) show spectral differences between quadriceps, gastrocnemius, tibialis anterior, and cardiac muscle (95% CIs shown by dotted lines and shaded areas). (b), (f) Coefficient weightings of each principal component describe sources of variance across all wavelengths, including each PC’s total percentage explained variance (see legend). (c), (g) PC1 and PC2 scores show clustering of muscle types, which enabled their classification using decision boundaries (shaded regions) designed by a QDC. (d), (h) Classification ROC curves provide a quantitative measure of muscle type separability, where an AUC of 1 describes a perfect classifier and 0.5 a weakest classifier (see legend for muscle-specific ROC-AUCs).

To enhance muscle group separability and resulting classification accuracy, spectra were dimensionally reduced using PCA. The coefficient weightings for the first three components of AF and DR spectra are shown in Figs. 2(b) and 2(f), respectively, whereby large (non-zero) weightings indicate sources of variance between observations. For AF spectra, prominent weightings were observed at 475, 550, and 650 nm and for DR spectra at 490 and 570 nm.

These spectral differences caused muscle-type specific clustering of AF and DR PC scores as shown in PC1 versus PC2 score space [Figs. 2(c) and 2(g)]. Skeletal muscles formed mildly overlapping clusters with intercluster variance across both PC1 and PC2, whereas cardiac muscle formed a separate cluster best distinguished from skeletal muscles through PC1 in AF measurements, and PC2 in DR measurements. PC3 variance is not visualized in these figures but was also used for the purpose of classification.

The classification of muscle types was performed using a QDC classifier with fivefold cross validation. The classifier’s performance for each muscle type was quantified using the area under the curve (AUC) of its ROC curve [Figs. 2(d) and 2(h)], while its decision boundaries were observed in the PC scores plots [Figs. 2(c) and 2(g)]. The cardiac muscle was well separable from all skeletal muscles in both AF and DR analyses ($\mathrm{AUC}=1$ and $\mathrm{AUC}=1$ for AF and DR, respectively), whereas skeletal muscles (Quad, Gastr, and TA) were less separable with AUCs ranging between 0.79 and 0.99.

3.2.

AF/DR Spectra Enable Separation of Healthy Muscle and Muscle Containing Intramyocellular Lipid

It was next tested whether AF/DR spectra could separate healthy versus pathological muscle tissue in murine models of muscle disease. NF1 has shown features of a metabolic myopathy,⁵² and histological staining of quadriceps muscle in $Nf{1_{Prx1}}^{-/-}$ revealed an accumulation of intramyocellular lipid.^18,28 Cardiac tissue was included as a negative control as the Prx1 promoter does not allow for expression in the heart. Analysis was performed on B6 and NF1 mice (10 weeks) using AF and DR measurements (Fig. 3).

Fig. 3

(a)–(d) AF and (e)–(h) DR spectroscopy of quadriceps muscle enables classification of NF1 mice from B6 controls (both 10 weeks) using PCA scores; however, when applied to cardiac muscle, the same analysis yields no separability. Normalized AF spectra of (a) quadriceps and (c) cardiac muscle in B6 and NF1 mice with [(b), (d)] corresponding PC score and QDC decision boundary plots. Normalized DR spectra of (e) quadriceps and (g) cardiac muscle in B6 and NF1 mice with [(f), (h)] corresponding PC score and QDC decision boundary plots.

Normalized AF spectra from B6 and NF1 quadricep muscles showed significantly different means ($p<0.05$ between $\sim 450$ and 720 nm) [Fig. 3(a)]. NF1 DR spectra showed a more pronounced absorption feature between 500 and 600 nm, whereas both mouse groups showed different means across the entire wavelength range ($p<0.05$). Both AF and DR demonstrated strong quadriceps separability as indicated by distinct PC score clustering [Figs. 3(b) and, 3(f)] and QDC classifier AUC’s of 0.99 and 0.97, respectively. B6 and NF1 cardiac muscle showed similar spectral shapes and the same mean normalized intensity across almost all wavelengths for both AF and DR ($p>0.05$) [Figs. 3(c) and 3(g)]. This similarity resulted in overlapping PC score clusters [Figs. 3(d) and 3(h)] and poor QDC classification performance with ROC-AUCs of 0.67 and 0.63 for AF and DR, respectively. NF1 muscle histology (Fig. 4) showed endomysial fibrosis and intramyocellular lipid by H&E and Oil Red O staining, respectively, consistent with prior findings.¹⁸

Fig. 4

Histopathology of quadriceps muscle from 10-week-old NF1 mice differs from that observed in B6 mice. H&E staining shows endomysial fibrosis, and Oil Red O staining illustrates the intramyocellular lipid seen in NF1 mice. Scale bars indicate $200\mu \mathrm{m}$.

3.3.

AF/DR Spectra Provide Separation of In Predystrophic and Dystrophic Muscle

A similar analysis was performed on B10mdx mouse groups at two different time points using the B10WT groups as controls. At 3 weeks, mdx mice are known to be asymptomatic and show no major histological features of dystrophy. By 10 weeks, mdx mice show centralized nuclei and fibrosis characteristic of dystrophic muscle. It was queried whether any fundamental differences in the AF/DR spectra would be apparent in mdx mice before histological features were apparent, as well as at later stages of muscle degeneration/regeneration.

AF and DR measurements were made on the quadriceps of B10WT and B10mdx mice at 3 and 10 weeks [Figs. 5(a) and 5(d)]. The PCA-QDC method was again used to test separability of different pairs of these groups. First, B10WT groups at 3 and 10 weeks were compared. Both groups showed similar AF spectral features [Fig. 5(a)] with overlapping PC Score clusters [Fig. 5(b)] making them indistinguishable to the QDC [Fig. 5(c)] with an ROC $\mathrm{AUC}=0.58$ (Table 1). DR measurements provided better separability, $\mathrm{AUC}=0.83$. In contrast, B10mdx at 3 and 10 weeks showed differences in AF spectral features resulting in strong classifier performance, $\mathrm{AUC}=0.96$, whereas DR spectra were only modestly separable, $\mathrm{AUC}=0.77$. Further comparisons were made between control B10WT and B10mdx mice, both at 3 weeks, with strong AF separability, $\mathrm{AUC}=0.98$ and with more modest DR separability, $\mathrm{AUC}=0.73$. Finally, B10WT and B10mdx mice were compared at 10 weeks, with similar results to that of 3 weeks, with stronger separability using AF, $\mathrm{AUC}=0.87$ and modest separability using DR, $\mathrm{AUC}=0.77$.

Fig. 5

(a)–(c) AF and (d)–(f) DR spectroscopy on quadriceps muscle enabled classification of mdx mice from B10 controls using PCA at both 3 week and 10 week time points. Normalized AF and DR spectra (top row) indicate differences in optical properties between these groups. These differences are highlighted by clustering in PCA space (middle row) as well as in their classification ROC curves (bottom row).

Table 1

DC ROC-AUC metrics for pairwise classification of B10WT and B10mdx mouse groups at 3 and 10 weeks. AUCs were used as a measure of class separability, where AUC>0.9 = strong separability, 0.9>AUC>0.7 = modest separability, and AUC<0.7 = poor separability.

Compared groups	QDC ROC-AUC
	AF	DR
B10WT 3 weeks versus 10 weeks	0.580	0.826
B10mdx 3 weeks versus 10 weeks	0.956	0.774
B10WT versus B10mdx at 3 weeks	0.982	0.793
B10WT versus B10mdx at 10 weeks	0.868	0.774

Histology of B10mdx mouse quadricep (Fig. 6) muscle found features of dystrophy consistent with the literature.⁵³ This included centralized nuclei and degenerating/regenerating muscle fibers as seen using H&E histological staining. 3-week-old B10mdx mouse muscle showed no pathological features, comparable to B10WT controls.

Fig. 6

Representative histopathology of quadriceps muscle from 10-week-old mdx mice revealed several dystrophic features, including centralized nuclei and muscular atrophy by H&E not evident in 3-week-old mdx mice and aged B10 controls. Scale bars indicate $200\mu \mathrm{m}$.

4.

Discussion

In this study, we describe the construction and testing of a fiber-optic AF/DR probe with significant potential for in situ analysis of biological tissues. The system relies on the native tissue optical properties in contrast to a growing series of studies that use custom-engineered exogenous markers including quantum dots,⁵ carbon nanotubes,⁶ and other nanomaterials,⁸ chiefly operating in the near-infrared wavelengths. At visible excitation wavelengths, AF and DR spectroscopies are more limited by tissue scattering in comparison with NIR. However, in applications such as skeletal muscle in which tissues are easily accessible, independence from exogenous dyes is advantageous.

UV light sources are typical in AF spectroscopy. Strong fluorophore absorption in this spectral region, predominantly by amino acids, ensures strong excitation efficiency and a broad fluorescence emission spectrum for analysis. However, given a monotonic decrease in tissue scattering coefficient with wavelength and a rapid drop in absorption coefficient across the visible spectrum (400 to 750 nm) in most tissues,^54,55 deeper penetration of illumination light can be achieved using longer wavelengths.^56,57 Visible laser sources are in addition safer, cheaper, and more efficiently guided by standard optical fiber in comparison with UV wavelengths. As such, a suitable balance between the variables of excitation efficiency, tissue penetration, cost, and safety were achieved in this study using 405 nm. At this wavelength, the exogenous fluorophores most efficiently excited include porphyrins, phospholipids, collagen, and flavins, such as FAD.¹⁰

Spectral examination of different skeletal muscles and myocardium showed a separability of these muscle types. This is consistent with the data of Chagnot et al.²⁴ and Schilders and Gu²³ who showed intrinsic AF differences between muscle fiber types using deep ultraviolet and visible wavelength excitation microscopy, respectively. A major source of variance between muscle types was observed between 500 and 600 nm for both AF and DR measurements, as is evidenced by principal components weightings in this region [Figs. 2(b) and 2(f)]. For DR measurements, this can be attributed to the combined absorption by Mb and Hb in both their oxygenated and deoxygenated states,^58,59 whereas an additional absorption feature at $\sim 760\mathrm{nm}$, evident only in cardiac muscle DR spectra [Fig. 3(e)], can be attributed to deoxygenated forms of both Mb and Hb.^60,61 Deoxy-Mb and Hb exist in all muscle tissues, so this absorption may be preferentially elevated in cardiac muscle due to a higher concentration of these chromophores. Notably in these studies, we analyzed AF and DR data separately to compare the value of both modalities. Fiber-optic AF spectra are inherently a hybrid signal affected by AF and diffuse scattering and photon absorption, so integrating its analysis with DR would be questionable and certainly unconventional.

Two murine models of muscle disease that featured different phenotypes were investigated. The first was a model displaying limb-specific deletion of Nf1, which has been reported to develop intramyocellular lipid as well as weakness.¹⁸ The second model was the mdx mouse, which progressively develops a muscular dystrophy phenotype caused by structural deficiencies in the myofiber membrane.⁵³ NF1 mice were separable from wild-type mice using DR and AF spectroscopy on the quadriceps, whereas measurements on cardiac muscle were inseparable [Figs. 3(d) and 3(h)]. This result is consistent with histological findings of increased intramyocellular lipid, likely having different optical properties to muscle without lipid droplets. Fibrotic tissues contain a higher density of collagen, making them more fluorescent in comparison with healthy muscle tissue in which AF in the visible region is derived by weaker concentrations of mitochondrial fluorophores, such as NADH.¹⁰ Differences in DR spectra for NF1 measurements in skeletal muscles could be caused by increased adiposity as adipose tissue is a weak absorber of visible light,⁶² resulting in brighter reflectance spectra and weaker fluorescence spectra. Importantly, and regardless of the underlying differences in optical properties, these muscle types were separable when approaching the spectral data with PCA-based dimensionality reduction and machine learning techniques.

The mdx mice again showed a separability based on both age and genotype, which was best observed using AF spectroscopy. This finding is consistent with those of Nakae et al.²⁶ who attributed this phenomenon to the build-up of fluorescent lipofuscin granules in dystrophic murine and human samples. Notably, the B10 3-week and 10-week mouse quadriceps were not optically separable by AF/DR, whereas the mdx mice were. This indicates that the progressive dystrophic phenotype does shift the AF/DR spectra over time, which in a translational setting may enable disease progression to be followed longitudinally. An increase in lipofuscin accumulation has been demonstrated with age in both mdx and control mice, the latter of which only presents after 20 weeks.²⁶ Importantly, from this time point, dystrophic muscle remains optically separable from controls due to a net higher concentration of the fluorophore throughout all time.

A limitation of this study was that the B10 controls were not littermate controls, and there may have been genetic drift with the mdx strain to confound a direct comparison. Furthermore, the use of recently euthanized mice rather than live mice and the biochemical and morphological differences between human and murine muscle tissue⁶³ promote the need for further in vivo validation studies. Such studies may be supported by a more streamlined fiber-optic bundle with fewer fibers allowing long-term percutaneous access.

More robust methods of spectral calibration and analysis may also be beneficial. AF proved more diagnostically accurate than DR across all studies. When using fiber-optic geometries in optically turbid media, intrinsic fluorescence emissions are elastically scattered and absorbed as the photons traverse the tissue volume between sites of fluorophore emission and fiber-optic collection. As such, fiber-optic AF spectra are hybrid in nature and sensitive to the same optical processes that dominate DR. It is not obvious from our data whether this sensitivity enhances or degrades AF diagnostic efficacy, but different approaches to spectral analysis may prove more informative. It is worth noting that the PCA-based spectral feature extraction and noise rejection used in this study do not necessarily facilitate the understanding of the physical processes underpinning our results as complex PC weightings were difficult to assign biological significance. Other approaches that rely on detailed modeling of physical processes have enabled intrinsic AF extraction using a well-calibrated DR spectrum in combination with light transport models based on tissue phantom data, diffusion theory,⁶⁴ Monte Carlo simulations,⁶⁵ and other simpler numerical methods.^66,67 Such an approach may also be complemented by the use of functional PCA⁶⁸ instead of its more conventional form described in this article. If implemented in the context of this study, such methods may realize a more robust diagnostic system that could function across a more diverse range of hardware and sample populations in which factors such as tissue blood volume, probe geometry, contact pressures, and other hardware specifications are not as easily controlled for.

5.

Conclusion

We describe the design of a minimally invasive marker-free method of optical spectroscopy suitable to the real-time in vivo analysis of muscle tissue. The system was implemented in healthy and diseased mouse tissues, including models of muscle myopathy seen with NF1-deficiency in mice and muscular dystrophy caused by dystrophin deficiency. Statistically transparent spectral analysis methods based on feature extraction by PCA and a QDC proved capable of distinguishing between healthy skeletal muscle types and diseased tissues. These data show the feasibility of using a fibre-optic AF and DR spectrometry system to characterize muscle tissue and underlying disease. This study serves as a strong proof-of-principle for further development of hardware and data analysis methods for clinical translation. To date these findings are largely empirical, and further study of autofluorescent macromolecules will be needed to appreciate the mechanisms underlying these spectral changes.