Table 1 reports densities and counts of myelinated fibers from various species (rats, cats, monkeys, humans). One point of uncertainty when creating the following table arose from the terminology used in some articles. Sometimes, when employing the word “fibers,” authors did not explicitly mention if they were referring to myelinated fibers only or to all fibers (myelinated and unmyelinated). When applicable this distinction was indicated.

One notable observation emerging from this table is the variability in densities and counts from one specimen to another [e.g., density of 152,000 myelinated axons per mm² vs. 125,000 (Biedenbach et al., 1986)], as well as from one study to the next [e.g., for the human pyramidal tract, Demyer (1959) counted 1,000,000 myelinated fibers against 49,000–67,000 for Wada et al. (2001)].

As a general trend, different fiber counts are associated with different spinal levels. For example, the pyramidal tract of the rat in the brainstem presents roughly between 90,000 and 140,000 fibers, depending on the study, whereas the CST at the cervical 2 level seems to be constituted by about 43,000 fibers (Leenen et al., 1985).

Axon diameter has been measured in several species. The numbers are reported in Table 2 along with the myelin sheath thickness. Note that in previously-published studies the word “fiber diameter” sometimes included the surrounding myelin sheath and sometimes not, without clear mention of it.

The average diameter in the pyramidal tract of the rat varies between 0.72 μm (Leenen et al., 1985; Brösamle and Schwab, 2000) and 1.182 μm (Harding and Towe, 1985). Variations could be due to different shrinkage (i.e., different protocols), inter-individual differences, or due to whether the myelin sheath was included in the measurement. The distribution of calibers is broad (ranging from 0.1 to 6 μm) but more than 90% of axons are smaller than 1.5 μm. In the corticospinal tract, the average diameter is 0.72 and 0.68 μm (respectively for the main and the ipsilateral ventral components) (Brösamle and Schwab, 2000).

In the cat, the only average axon diameter available is in the pyramidal tract, with a value of 1.98 μm (Biedenbach et al., 1986). In this tract, diameters found were as low as 0.25 μm and up to 23 μm (Biedenbach et al., 1986), but the pyramidal tract is generally composed of thin fibers with the majority of axons smaller than 2 μm (van Crevel and Verhaart, 1963; Biedenbach et al., 1986). Note that a maximal diameter of 23 μm is quite high when comparing to the largest human myelinated fibers [20 μm with myelin sheath in the highest cases (Standring, 2008)], or to the results obtained for the monkey. One has to keep in mind the possibility of histological distortion or the possibility that authors included the myelin sheath in the measurement (this was not specified).

In the monkey, the average diameter in the pyramidal tract is 0.91 μm (Firmin et al., 2014). Axon caliber distribution is broad, ranging from 0.04 to 9.48 μm. There is also a large presence of small fibers, with 52% of myelinated fibers presenting axons smaller than 1 μm (Firmin et al., 2014). In addition, the tract is reported to be heavily myelinated (99% of axons) (Firmin et al., 2014).

Axons of the pyramidal tract in humans were found to be as thin as 0.3 μm (Graf and Schramm, 1983) and as large as 20 μm (Verhaart, 1947), with a majority of thin fibers (about 84%) smaller than 2 μm. This maximal diameter is high, even for human fibers, and possibly explained by inclusion of the myelin sheath. In the lateral CST, predominance of small fibers over large ones was shown to be more pronounced at cervical level 6 and lumbar level 4 than at thoracic level 7 (Terao et al., 1994). In a similar way, according to Ohnishi et al. (1976) in the fasciculus gracilis of humans, axon diameter is smaller at C3 than at T5.

Since the 80's, the presence of a high number of unmyelinated axons in the WM of the SC is often reported, particularly in cats and rats (Langford and Coggeshall, 1981; Leenen et al., 1982, 1985; Chung and Coggeshall, 1983; Chung et al., 1985; Harding and Towe, 1985).

Several studies about the effect of aging in different nervous tissue have been published and are summarized in Table 5.

In the human SC, fiber counts and mean axonal diameter (Ohnishi et al., 1976; Zhou et al., 1997) as well as fiber density (Terao et al., 1994; Nakanishi et al., 2004) are believed to decrease with age. Whether this decrease in density is preferential to small vs. large fibers is debated (Terao et al., 1994; Nakanishi et al., 2004). With respect to human nerves, results are diverse and often contradictory. For the human superior laryngeal nerve, axonal counts have been reported as decreasing with age (Jacobs and Love, 1985; Mortelliti et al., 1990) or remaining constant (Tiago et al., 2007). The same trends were observed in animals, with decrease of fiber density with age. However, large discrepancies were found, for examples in monkeys (Morrison et al., 1990; Fortune et al., 2014), and rats (Jeronimo et al., 2005; Watanabe et al., 2007; Soltanpour et al., 2012). The tissue (e.g., PNS vs. CNS) and the species of origin may influence the evolution of morphometric parameters in regard to aging.

The effect of aging in the myelin sheath is poorly studied. One comment by Jacobs et al. suggested that thin or medium sized axons showed a disproportionately thick myelin sheath in older humans (>60 years old), consisting in loss of the expected relationship between axonal diameter and myelin sheath thickness (Jacobs and Love, 1985).

Table 6 describes the effect of gender on axon morphology. Of particular interest, the study of Yang et al. (2008) showed that in young mice, males presented a higher total volume of myelinated fibers and myelin sheath than females did. A significant decrease in those parameters in male with aging was reported, whereas they increased with age in females. No sex difference for mean myelinated fiber diameter was found. They hypothesized that the increase in mean fiber diameter in males could be caused by the loss of the smaller fibers. In females, the increase in myelin sheath thickness could be explained by split of myelin lamellae, because of a shorter life span of oligodendrocytes. In a similar fashion, a group showed the importance of age on the SC but later (Zhou et al., 2000) refined their findings by specifying that the age-related reduction in axon diameter was seen in males only (Zhou et al., 1997, 2000). In the mice SC (cervical level), Cerghet et al. (2006) showed smaller levels of oligodendrocytes in females than in males as well as reduced myelin basic protein, hinting at a lesser degree of myelination in females (Cerghet et al., 2006). Interestingly, other studies did not find a significant effect of gender on axon morphometry (van Crevel and Verhaart, 1963).

Sexual dimorphism has already been established in the brain (particularly in the hippocampus) so even if to this day this issue has not been much addressed in the SC, it is likely that sexual dimorphism might be present in there too (Bear et al., 2007). For example, the sacral level in the SC differs from males to females since the former have supplementary motoneurons in the gray matter in order to control bulbocavernous muscles involved in ejaculation (Breedlove and Arnold, 1980; Forger and Breedlove, 1986; Wagner and Clemens, 1991; Bear et al., 2007). In addition, it seems that aging and sex would be interdependent factors for morphometry.

The number of fibers in the WM will necessarily vary according to the rostro-caudal location in the SC, with descending fibers reaching their targets and ascending fibers entering progressively (Standring, 2008). Ohnishi et al. (1976) have indeed reported a decrease in the number of myelinated fibers of the fasciculus gracilis in human between C3 and T5 (respectively 25,267 and 23,069 per mm²). In general, more studies covered cervical segments than thoracic and lumbar ones.

The variation of axon diameter regarding the vertebral level is not clear so far: Leenen et al. (1985) reported no difference in the rat pyramidal tract, whereas differences were reported in the human lateral CST (Terao et al., 1994) and in the human fasciculus gracilis (Ohnishi et al., 1976).

Intra-species variability seems to be high, as noted by Verhaart (1947), van Crevel and Verhaart (1963), Jacobs and Love (1985), and Morrison et al. (1990). It is informative to look at the number of animals used in each article. For instance, Harding's results (Harding and Towe, 1985) are based on a total of two rats; Biendenbach's (Biedenbach et al., 1986) on three cats; Firmin et al. (2014) used two monkeys for the neuroanatomical part of his article. Leenen's study (Leenen et al., 1985) in rats was realized on five animals, similar to Ralston's study (Ralston et al., 1987) studying six monkeys. In addition to the relatively low number of animals per study, because most studies used TEM, in general only 1–6% of the area of the tract is considered in the analysis. Partial sampling can introduce variability and error in addition to not being representative (Fortune et al., 2014).

Table 7 exemplifies some protocols used for characterizing axon morphology. The fairly large variability of protocols can partly explain the variability observed in the reported results across studies. Indeed, the choice of protocol can have several impacts on shrinkage, distortions and other artifacts. Another typical example that often occurs when imaging deep tissue, is that imperfect fixation of the myelin sheath with osmium (due to poor penetration) can induce lamellae separation (Harding and Towe, 1985).

Even with an optimal protocol, issues remain such as microscope resolution. The theoretical resolution with light microscopy is ~300–350 and ~200–250 nm with numerical aperture of 1 (in air) and 1.5 (in oil), respectively (Bart, 2006; Madou, 2011; Nadeau, 2012; Bradshaw and Stahl, 2015), although in practice a resolution below 1 μm is hard to obtain (Madou, 2011). As a result, approximately 20% of fibers are not detected when imaged with light microscopy (Arbuthnott et al., 1980; Biedenbach et al., 1986; Ralston et al., 1987; Firmin et al., 2014). To reduce the mismatch between neuroanatomical and electrophysiological measures of fibers diameter, Firmin et al. (2014) used electron microscopy to show that many more small axons (with a diameter of 0.5 μm and less) were detected in comparison with previous studies that used light microscopy (Arbuthnott et al., 1980; Harding and Towe, 1985; Hildebrand et al., 1993; Firmin et al., 2014). The difficulty in detecting small fibers (less than 1 μm) comes with the challenge of accurately obtaining morphometry information such as diameter and myelin sheath thickness (Biedenbach et al., 1986; Firmin et al., 2014).

Lastly, most studies have employed a manual fiber identification and manual selection of the profile of interest. This procedure may induce a bias against small or ill-defined fibers (Biedenbach et al., 1986). Furthermore, the time needed to do this manual identification produces exhaustion-related bias and more importantly limits the area studied (More et al., 2011; Reynaud et al., 2012; Isaacs et al., 2014). For instance, Firmin et al. only sampled 0.015 mm² of the tract in EM when the complete acquisition window was of 3.782 mm² (Firmin et al., 2014).

There is a need for a thorough investigation of SC morphometry with high statistical power. To achieve this goal, analysis of the complete section of the SC (as opposed to samples) from several animals is necessary to provide a good overview of the inter-individual variability within species when all factors are accounted for (sex, weight, etc.). Due to the large area of the SC however (~1 cm² in human) and the small size of fibers (~1 μm), a fast imaging technique at high resolution is necessary. To analyze these images, automatic and accurate segmentation softwares are needed.

The tissue can be fixed in two ways. The first method, called immersion fixation, consists of cutting the tissue into small pieces which are then submerged into fixatives until the tissue is set in place. The cutting of the sample into small pieces should be done with a razor blade by one quick slashing motion instead of several back-and-forth movements in order to avoid any tissue damage. The second method, called perfusion fixation, involves delivering the fixatives through the systemic circulatory system. Perfusion flowrate should be adjusted according to the blood pressure: a pressure too high could lead to vessels exploding, while a low pressure could have a detrimental effect on fixation speed.

It is generally accepted that vascular perfusion fixation is a necessary step for tissues that undergo rapid lysis after removal and for providing well-preserved axons and myelin (Palay et al., 1962; Williams and Jew, 1975; Langford and Coggeshall, 1980; Lamberts and Goldsmith, 1985; Chang and Slikker, 1995; Fix and Garman, 2000). Thus, if perfusion cannot be done (e.g., human tissue), the tissue needs to be immersed in the fixative solution as soon as possible after death in order to minimize tissue damage.

A mixture of glutaraldehyde (Ga) and paraformaldehyde (PFA) is recommended for ultrastructural studies as it combines the high cross-linking ability of Ga and the fast action of the PFA (Williams and Jew, 1975; Lamberts and Goldsmith, 1985; Chang and Slikker, 1995; Fix and Garman, 2000; Brancroft, 2008; Hayat, 2012). The original paper by Karnovsky (1965) recommends a mixture of 5% glutaraldehyde and 4% formaldehyde, which is a higher concentration than what people typically use (Hayat, 2012). Other empirically-determined mixtures that are widely used for both structural and immunocytochemical studies consist of a modified Karnovsky solution by using it in half-strength formulation; 2% PFA and 2.5% Ga in 0.1 M Sodium Cacodylate (or phosphate buffer with a pH of 7.4) (Karnovsky, 1965; Dykstra and Reuss, 2003). The optimal concentration of Ga and PFA is difficult as the chemical interactions of these two fixatives with each other and with the tissue are still not completely understood (Glauert and Lewis, 2014). A relatively wide range of Ga concentrations (about 1.5–4%) has been used in animal tissue fixation (Hayat, 2012). In general, a low concentration of fixative requires a longer duration of fixation, but long Ga fixation (over a week) can cause tissue shrinkage (Hayat, 2012). High concentrations (>4%) are also unsuitable since they destroy the ultrastructure (Bozzola and Russel, 1992). The presence of Ga affects the process of protein cross-linking and the osmolarity. High Ga concentrations (30% and more) can lead to severe shrinkage whereas low Ga concentration (0.5% and below) can cause severe cell components extraction if osmolarity is not controlled (Dykstra and Reuss, 2003; Hayat, 2012). Such extreme osmolarity of the fixative leads to myelin shrinkage, swelling or vacuolation.

Note that Ga should not be used with coherent anti-stokes Raman spectroscopy (CARS) imaging because it discards the anti-stokes effect (Bélanger et al., 2009). In this case, Ga can simply be discarded or replaced with acrolein.

The penetration of fixatives into the tissues is described by Fick's second law (Yong-Hing et al., 2005; Thavarajah et al., 2012): ∂ci∂t=D∂ci∂x where i is fixative considered, c_i the fixative concentration, x is the axis of penetration, t the time of fixative penetration and D the diffusion coefficient of the fixative. Assuming that the volume of a fixative is large enough and the nominal concentration c₀ of the solution is unchanged at the boundary of the tissue, the fixative concentration is:

Let's define d, the depth of penetration at which 50% of the nominal concentration (c_i(d, t) = 1/2c₀) is obtained. We can write d ≈ 1/2Dt.

The diffusion coefficient D is influenced by temperature (T), fluid viscosity (η), the size of the chemical compounds (r) according to the Stokes-Einstein law for solutions with low Reynolds numbers: D=kBT6πηr where k_B is Boltzmann's constant (Mehrer and Stolwijk, 2009). Thus, high temperatures, low fixative solution viscosity and small chemical compounds will increase the rate of diffusion.

The diffusion values for osmium tetroxide (OsO₄), Ga and formaldehyde are respectively D = 0.2, 0.34 and 2.0 10⁻⁹ m²/s (Johannessen, 1977). Thus, using d≈1/2Dt, at room temperature, Ga and PFA penetrate the tissue by 1 and 3 mm respectively after 4 h, or by 3 and 7 mm after 24 h.

Regarding Osmium penetration, the reaction with unsaturated lipids drastically reduces the permeability of the membranes and thus the diffusion within the tissue. Osmium staining is thus strongly reduced after 200 μm. Strategies have been developed to improve Osmium penetration, by reducing temperature during the osmification process (osmium reaction is more greatly reduced than diffusivity) or/and by using a reducing agent to improve contrast (Mikula et al., 2012; Hua et al., 2015)

At low temperatures, the penetration rate is slowed down, as well as the reaction of the fixatives (which affect the permeability of the tissue membranes). At high temperatures autolytic changes occur more rapidly (Hunter et al., 1993) which cause the formation of fixation artifacts (e.g., cell shrinkage, cell volume and shape changes) (Hayat, 2012). Some investigators demonstrated that formaldehyde fixation at room temperature induces poor preservation (Cross et al., 1990). However, depending on the type of tissue, Ga can be used at temperatures ranging 0–25°C with little morphological differences in the ultrastructure (Hayat, 2012). Regarding perfusion, cold temperatures affect some metabolic processes and can induce loss of these structures (Hayat, 2012). In vascular perfusion, using a Ga solution below body temperature may cause vasoconstriction (Fix and Garman, 2000).

When cells die during the fixation process, the lysosomes release their contents and make the pH of the tissue acidic. The optimal pH is close to that of the tissue and has to be maintained during the entirety of the tissue preparation process. Indeed, a change in the tissue pH could severely alter its ultrastructure. Since the average pH of most animal cells is 7.4 it is recommended to adjust the pH of the fixative solution to the physiological range of 7.2–7.5, especially when using Ga or osmium tetroxide (OsO₄) (Peters, 1970).

The osmolarity of a fixative solution has a direct effect on the myelin morphology. Axon size and shape may be affected by the buffer or the fixative solution's osmolarity. Osmolarity depends on the type of buffer and the concentration of the fixatives (Schultz and Karlsson, 2006). At low osmolarity (≤320 mOsmol), the fixation is very poor and causes cell swelling. The osmolarity of the fixative solution should be slightly hypertonic (400–600 mOsmol) to prevent blood vessels from bursting during perfusion (Fix and Garman, 2000; Schultz and Karlsson, 2006). Higher osmolarity (>1,000 mOsmol) could induce considerable tissue shrinkage and widening of the extracellular space (Ohnishi et al., 1976). As the resistance in blood vessels increases during perfusion, the fixative flow stops rapidly through the specimen, resulting in a very poor fixation (Ohnishi et al., 1976; Schultz and Karlsson, 2006). Electrolytes, especially CaCl₂, when added to isotonic fixative solutions can overcome the issues of osmolarity and decrease fixation artifacts such as swelling and shrinkage (Hayat, 2012). To increase the osmolarity, saccharose or tannic acid can be added to the fixative (Hayat, 2012).

Osmium tetroxide (OsO₄) has two interesting effects: it fixes the lipids and it stain the membranes for optical or electron microscopy. Without osmium, lipids are dissolved in the dehydration process (Robertson, 1957; Langford and Coggeshall, 1980; Dykstra and Reuss, 2003; Bruce-Gregorios, 2006; Fujimoto et al., 2012; Glauert and Lewis, 2014). In electron microscopy osmium improves the back-scattered electron density thanks to its large atomic number (Z = 76) and the conductivity of the sample's surface (Bozzola and Russel, 1992). As discussed earlier, osmium has a short penetration depth, and it is thus recommended to use a small sample size in one dimension or to image a slice close to the surface (<200 μm). Note that other electron-dense stains exist, such as uranium salts, potassium permanganate and ruthenium tetroxide (Hayat, 2012).

After being post-fixed in osmium tetroxide, the specimen needs to be dehydrated at room temperature. The purpose of dehydration is to remove water from the tissue so that the embedding media can infiltrate the tissue. Ethanol (mixed with propylene oxide) or acetone solvent are commonly used. Some electron microscopy textbooks suggest that ethanol with propylene oxide present better infiltration than acetone and would be recommended for large samples (Bozzola and Russel, 1992; Glauert and Lewis, 2014). However, propylene oxide is more toxic than ethanol and acetone appears to cause less shrinkage than ethanol (Dykstra and Reuss, 2003; Rana, 2008).

As acrylic resin is harder than paraffin, the use of this embedding media allows for cutting thinner sections than with paraffin (0.1–2 vs. 3–6 μm), in turn allowing greater resolution for optical microscopy (Sheehan and Hrapchak, 1980; Bancroft and Gamble, 2002; Lowe and Anderson, 2014).

In order to cut the hard acrylic resin, diamond or tungsten-carbide knifes can be used. Diamond knives are used to cut ultra-thin sections for TEM (<100 nm), but they are usually only a couple of millimeters long and can only cut samples of that size range. Tungsten-carbide blades can cut larger samples however produces thicker slices (~1 μm).

The choice of epoxy resin is determined by its hardness, its viscosity (and thus its penetration), its cutting behavior (very homogeneous resins are easier to cut), and its resistance under the electron beam. “Embed 812” is a standard epoxy resin that satisfies the above-mentioned criteria.

Currently, the observation of axons and myelin sheath relies on the conventional stains mentioned above. However, the preservation of myelin ultrastructure is strongly sensitive to fixation artifacts. Therefore, new myelin imaging techniques have been developed during the last decades that can be used without fixing the tissue or using immunostaining with myelin basic proteins (MBP) and other agents that can identify the ultrastructure in situ (fluorescence) (Wang et al., 2010; Bajaj et al., 2013).

Coherent Anti-Stokes Raman Scattering (CARS) is a recent imaging technique that has proven useful for imaging myelin (Wang et al., 2005; Bélanger et al., 2009). This modality is sensitive to the biochemical nature of tissues. When the frequency difference between two laser beams targeting a sample is resonant with the Raman vibrations of lipid (CH2 at 2,850 cm⁻¹), the lipids that make the myelin membrane emit light. Axial and lateral CARS resolution are respectively about 0.70 and 0.28 μm (Wang et al., 2005). As CARS relies on the endogenous chemical contrast provided by lipids in the myelin sheath it abolishes the need for staining and dehydration/embedding (the critical step for myelin damage), therefore allowing for more realistic measures of g-ratio and myelin sheath thickness (Wang et al., 2005; Bélanger et al., 2009; De Vito et al., 2015). Additionally, this technique can image myelin in vivo, allowing researchers to follow demyelination in animal models of MS or nerve crush injury (Henry et al., 2009; Bélanger et al., 2011; Imitola et al., 2011). An example of a CARS image is shown in Figure 3.

Optical Coherence microscopy uses the backscattered light due to refractive index variations in the tissue as a source of endogenous contrast. Myelin produces a high intrinsic backscattering signal due to the high refractive index of the lipid-rich myelin sheath. Just like CARS, OCM performs virtual sectioning by selecting a slice (through interferometry) at a depth of up to one hundred microns. Preventing the sectioning step prevents tissue damage and allows in vivo imaging (Ben Arous et al., 2011). Axial and lateral resolution can reach respectively 1.15 and 0.5 μm (Ben Arous et al., 2011). OCM provides not only quantitative measurements of myelin optical properties but also novel optical markers of cell viability and myelination (Srinivasan et al., 2012). The limited resolution, however, prevents fine measurement of axon morphometry.

Third harmonic generation (THG) microscopy is a coherent, nonlinear, dye-free imaging modality. THG can reach a micrometer resolution and has been used to monitor myelin loss and recovery in the mouse CNS (Farrar et al., 2011). Just like the two previous techniques, THG does not require any exogenous dye or fluorescent proteins and provides 3D structural images. It can be combined with other non-linear optic such as the two-photon excitation fluorescence (2PEF) and second harmonic generation (SHG). The resolution can be improved down to 5 nm if a super-resolution technique is used (Sandkuijl et al., 2013; Dashtabi et al., 2017).

Immunostaining is an antibody-based method to detect a specific protein in the tissue. Immunostaining based on myelin basic proteins (MBP) in combination with confocal microscopy provides a sensitive and reliable method to observe the fine structure of myelin sheath (Xing et al., 2012).

Confocal microscopy has a pixel resolution around 0.3 μm (Conn, 2010). Using laser-scanning confocal microscope, axial and lateral resolution can reach 38 and 122 nm respectively (Potrusil et al., 2012), and can image a slice a couple of hundred microns deep in the tissue, allowing in vivo imaging. For electron microscopy, antibodies are labeled with gold particles that can be observed with TEM. However, immune-EM can be technically challenging, expensive and requires rigorous tissue fixation (Xing et al., 2012). Note that confocal microscopy does not necessarily require fluorescence labeling. Using multiple confocal lasers tuned at different wavelengths, Spectral Confocal Reflectance microscopy (SCoRe) can be specific to myelin, making the application of this technique easier on animals other than rodents (Schain et al., 2014).

Imaging large sections of a sample, such as a complete SC transection, has become possible through the use of a motorized stage combined with an algorithm that stitches together small field-of-view images acquired at high resolution. These large-scale images can then be observed and analyzed on a virtual slide viewer (Weinstein et al., 2009; Pantanowitz et al., 2011).

Alternatively, whole-slide scanners provide a high-speed standardized acquisition with predefined parameters, with the ability to image up to 200 slides in one night, in bright field or with fluorescence, at 20 or 40× (Weinstein et al., 2009; Nederlof et al., 2011; Pantanowitz et al., 2011; Ameisen et al., 2012; Laurent et al., 2013; Gallas et al., 2014). A recent example of the use of whole-slide scanners is the BigBrain: a whole human brain scanned at 20 μm resolution (Amunts et al., 2013).

Automated analysis has become important for large-scale morphometric studies, as manual identification of axons is long, tedious and subject to user bias (More et al., 2011). Several algorithms exist for axon and myelin segmentation (Romero et al., 2000; More et al., 2011; Bégin et al., 2014; Mesbah et al., 2016; Zaimi et al., 2016, 2017). Following segmentation of axon and myelin, metrics such as distribution of axon diameter, mean myelin thickness or myelin volume fraction can be more readily computed (Bégin et al., 2014).

The use of axon segmentation software has shown encouraging results in terms of sensitivity and accuracy, especially when compared to manual identification (Reynaud et al., 2012; Isaacs et al., 2014). For example, Bégin et al. (2014) segmented 60% of the complete section of a mouse SC in 4 h, counting a total of 32,000 axons. More et al. (2011) have identified 84.3% of axons in a whole SEM slice of peripheral nerve (4 ms/axon) and were able to measure myelin sheath thickness and axon diameter. Figure 4 illustrates a semi-automatic segmentation on SC tissue.

A fully automatic approach with perfect accuracy is, however, difficult to achieve. For example, algorithms can sometimes fail to discriminate fibers that are too packed, or to distinguish between neuronal fibers and other tissue components such as blood vessels. Hence, manual intervention is often required (More et al., 2011; Bégin et al., 2014; Isaacs et al., 2014; Zaimi et al., 2016). Machine learning methods (e.g., deep learning algorithm), could potentially be trained to overcome these issues and avoid human intervention (Cireşan et al., 2013; Zaimi et al., 2017).

In addition to the segmentation issues, the metrics extracted are potentially affected by the quality of the image. The inherent point spread function of the imaging system, a bad focus, or an inhomogeneous contrast/illumination (e.g., due to inhomogeneous staining) can lead to over- or under-estimated morphological values (e.g., myelin thickness, myelin volume). An additional challenge is thus to use an imaging system that is highly stable on the entire slice.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.