Wrapped up: advancements in volume electron microscopy and application in myelin research

2.1 Sample preparation

For any microscopy method, sample preparation is one of the most crucial steps. Here we briefly summarize a variety of sample preparation approaches for vEM. A detailed introduction into the methods is available in this review [6]. To our experience, the so-called rOTO (reduced Osmium-Thiocarbohydrazide-Osmium) protocol [18] (see below) is suitable for myelin visualization in mouse and human tissue [19]. In general, the targeted structure of interest and the applied imaging method are dictating the sample preparation method. In this section we describe which modifications were applied to meet specific requirements of a given sample or animal species.

Due to the inherent properties of the image formation inside the SEM and the sample properties of biological specimens, protocols were optimized with the aim to achieve homogenously stained large sample blocks generating enough contrast for high resolution imaging using backscattered electron detection. In general, the smaller the sample the easier and quicker fixation and staining can be achieved. In the 1960s the (OTO) protocol was developed as a staining method for lipid-containing membranes and droplets in osmium tetroxide-fixed tissue [20]. This was developed further by Deerinck and colleagues [18] for serial block-face scanning electron microscope (SBEM) volume imaging, since the original protocol suffered from inhomogeneous staining in larger samples. A combination of potassium ferrocyanide with aqueous OsO₄ solution, followed by thiocarbohydrazide and OsO₄ (rOTO) as well as lead aspartate staining enhanced the signal for backscattered electron imaging of resin embedded tissue at low acceleration voltages. Potassium ferri- or ferrocyanide were used in combination with OsO₄, making it more reactive [21]. However, this was not applicable in samples above 200 µm of sample thickness. In the lab of Winfried Denk a dense en-bloc staining protocol was developed replacing thiocarbohydrazide with pyrogallol (BROPA, brain-wide reduced-osmium staining with pyrogallol-mediated amplification) and using Spurr’s resin for the resin infiltration of bigger samples [22]. This was further developed for faster processing of Zebrafish brain (called fBROPA) to reduce incubation times from months to days [23]. Here, the addition of lead aspartate increased conductivity and prevented charging during SEM imaging. In another approach contrast and diffusibility were enhanced by a two-step uranyl acetate incubation protocol with the first incubation at low temperature and the second at higher temperatures, followed by lead aspartate staining [24]. Attempting to process even larger samples, the same lab modified the protocol according to Mikula to extend the sample size from 8 mm³ up to a full mouse brain (∼1 cm³ in size) [25]. Most of the published protocols are variations of the common rOTO-scheme and were developed for brain samples of specific species. For the preparation of other organs or brain tissue from different model organisms, reagents and incubation times might require optimization.

If chemical fixation deteriorates the structure of interest, or samples are small enough (max. 200 µm thickness) high-pressure freezing followed by freeze substitution can be used to preserve samples closer to their native state. Here, cryo-immobilization is used to keep all structures in their native context, followed by slowly introducing different stains while warming up the sample and finally embedding it in a resin. Most freeze substitution protocols for TEM imaging use only a few staining steps. To be able to get enough contrast for focused ion beam – scanning electron microscopy (FIB-SEM) or SBEM, especially for larger tissue, additional contrast enhancement steps (using OsO₄ or uranyl acetate) can be added at the end of the freeze substitution also utilizing microwave-assisted processing for thicker specimen [26].

In the choice of the embedding resin, viscosity for good infiltration into large tissue blocks, rigidity of the sample to vacuum conditions and resistance to electron beam damage as well as the cutting properties in SBEM need consideration. An exhaustive comparison of a variety of resins regarding viscosity and behavior in FIB-SEM and SBEM was recently performed by Borghgraef and colleagues [27]. Tegethoff and Briggman also evaluated resin formulations for their polymerization uniformity, hardness and cutting properties to identify the best protocol for vEM [28].

2.2 vEM data acquisition

In the recent years various model systems and disease mechanisms have been studied using isotropic datasets generated by the FIB-SEM [29], [30], or larger datasets with lower resolution in z using the SBEM [31]. Lastly array tomography has been used to acquire datasets on sections collected on tape or wafer, which allows a rescanning of the same area at different resolutions [32], [33], [34]. Further developments on the microscope end allowed image collection not just with one electron beam, but an array of 61 or 91 beams [35], [36] and up to 196-beams at a time [37] leading to fast high-resolution mm-scale data collection. Here the lab of Jeff Lichtman in Harvard is pioneering efforts to map petascale fragments of the human brain [9], [38].

SEM samples need to be conductive to allow charge dissipation from the electron beam. Due to a better balance between negatively charged electrons and positively charged gallium ions inside the FIB-SEM chamber less charging occurs on the surface of the sample and smaller amounts of heavy metal suffice to dissipate charges. Larger specimens with open space between cells suffer from charging artefacts especially in SBEM. To prevent this, conductive components, for example silver particles, can be introduced into the resin embedding during the last infiltration step, surrounding the specimen with metal particles and dissipating charge from the freshly cut surface [39], [40]. Another approach to circumvent charging from the microscope point of view is a focal charge compensation device in the chamber flowing nitrogen gas across the sample surface and reducing surface charging [41].

Volume data can also be acquired under cryogenic conditions with a FIB-SEM or Plasma FIB-SEM (PFIB-SEM) if the samples need to be kept as close as possible in native conditions. PFIB-SEM utilizes different ion sources such as argon, xenon or nitrogen instead of gallium to allow faster milling of bigger specimen for volume imaging [42].

3.1 Ilastik

Ilastik is a Python-based “interactive learning and segmentation toolkit” (www.ilastik.org) designed to enable users without image processing expertise to easily analyze biological images. It offers different workflows that implement segmentation and classification tasks in an easy-to-use interface. The most widely applied workflow is pixel classification using a random forest classifier that learns from user-provided labels through a paint interface. The classifier uses intensity, edge and textural features, which are computed for each pixel based on its local neighborhood. The pixel classification workflow performs semantic segmentation, separating the image into classes such as foreground or background, instead of performing an instance segmentation, where individual objects are created. This approach provides real-time feedback, facilitating interactive segmentation and correction. By combining machine learning with interactive user annotation, pixel classification is able to handle complex segmentation tasks. Ilastik allows users to view overlays with predictions and an uncertainty map to guide through difficult regions within the image. Once trained on a representative set of images, Ilastik can batch process large image sets automatically. Pixel classification supports both 2D and 3D data and is capable of automatic and semi-automatic segmentation of EM volumes and other microscopy data. Other Ilastik workflows provide functionality for automatic and semi-automatic instance segmentation, object detection and object tracking. Among these, the “carving workflow” is especially useful for vEM data, as it can be used for interactive 3D segmentation based on seeded region growing. The implementation of Ilastik can be particularly useful for the segmentation of myelinated axons across nodes of Ranvier [55] or the detection of myelin in general ([66] and Supplementary within). It is also integrated within other tools for myelin research such as AimSeg (see below).

Overall, Ilastik, more specifically its pixel classification workflow, offers an excellent starting point for many analyses due to its user-friendly design, requiring no coding experience, and its capability to efficiently perform standard image processing tasks. Although it excels in handling local cues such as brightness, color and texture, it is not optimized for capturing global information. Nonetheless, it provides a practical solution for many image segmentation challenges that would otherwise require a custom algorithm. Ilastik is open-source and can be used and extended without restrictions [61].

3.2 Semi-automated and automated segmentation and quantification of myelin properties on 2D images

Introduced in 2021, MyelTracer is a Python-based, open source, semi-automated software focusing on g-ratio quantification (https://github.com/HarrisonAllen/MyelTracer) [62]. The g-ratio quantification is based on the Open Computer Vision Library (OpenCV) and PyQt5’s GUI. The developers of MyelTracer place a high priority on user-friendliness to make their tool accessible to users with little to no coding experience. Their software uses intensity thresholding to create contours around axons and myelin, but also includes a manual tracing function, allowing for correction of abnormal/difficult myelin sheath shapes and axons. The efficacy of MyelTracer was tested across different tissues and ages, including remyelination scenarios, and samples after optic nerve crush. MyelTracer can be used for 2D segmentation, to measure areas, calculate g-ratios, and determine the percentage of myelinated axons, reducing quantification time by 40–60 %.

AimSeg is a machine learning-based tool for the assessment of myelin thickness and myelin properties in EM images. It uniquely focuses not only on compact myelin, but also on the innermost non-compacted myelin layer, known as the inner tongue (https://github.com/paucabar/AimSeg) [63]. Although critical for myelin function and pathology, the inner tongue has been largely overlooked in past analysis. AimSeg segments axons, inner tongues, and compact myelin hierarchically as three objects, which in combination enables precise calculation of parameters like g-ratio and inner tongue area.

The tool offers both, automated segmentation and an assisted workflow, allowing for correction of results with minimal human intervention. It was trained on TEM images of mouse corpus callosum undergoing demyelination and remyelination. Thus, making it versatile for analyzing and quantifying compact and non-compacted myelin under various conditions. Its core pipeline utilizes a Fiji script to process input files generated by supervised machine learning classifiers from Ilastik (for pixel and object classification). Users can either use the pre-trained classifiers and enhance them with their own data and annotations, or train their own classifiers based on the provided user documentation. In addition, AimSeg includes optional post-processing steps for automatic correction of axon segmentation and user editing steps for correcting inaccuracies or common annotation errors before quantification. When assessing their computation time, the authors could show a reduction of 1 h per image manual annotation time (validation of ground truth) to 6–7 s on average by using automated processing with AimSeg including postprocessing. AimSeg provides a bioimage analysis tool, combining the user-friendliness of Ilastik with the versatility of Fiji. However, a potential limitation might be the spatial resolution when applied to other imaging modalities.

AxonDeepSeg (https://github.com/axondeepseg/axondeepseg) is an open-source tool using supervised deep learning for the automatic segmentation of axons and myelin sheaths on 2D images [43]. Featuring a CNN architecture, the model is tailored for the semantic segmentation of histological images and provides two pre-trained models optimized for SEM and TEM data. Trained on mouse brain samples and macaque corpus callosum, the authors achieved high accuracy in segmenting axons and myelin, as demonstrated on a full rat spinal cord slice. AxonDeepSeg is implemented in Python using the TensorFlow deep learning framework. It is important to note that AxonDeepSeg has been rigorously trained and tested on healthy tissue. However, the study did not explore its efficacy on pathological samples.

3.3 Automated segmentation and quantification of myelin properties on 3D volumes

ACSON (AutomatiC 3D Segmentation and morphometry Of axoNs) supports automated segmentation and morphometric analysis of white matter ultrastructure in 3D [64] (https://github.com/aAbdz/ACSON). This tool segments various components, including myelin, myelinated and unmyelinated axons, mitochondria, cells, and vacuoles, providing a detailed morphological analysis. By applying their pipeline to SBEM image stacks of the corpus callosum from sham-operated and brain-injured rats, the authors revealed that cross-sections of myelinated axons were elliptical rather than circular. In addition, they showed that axonal diameters varied significantly along their longitudinal axes, underscoring the limitations of 2D studies. The ACSON pipeline comprises several steps. Initially, the images are denoised, after which the bounded volume growing (BVG) technique is employed for segmentation. This combines seeded region growing and edge detection algorithms in three dimensions. Subsequent refinement of the segmentation is achieved through supervoxel techniques. Finally, the pipeline segments subcellular structures and cells, with the aim of annotating myelinated and unmyelinated axons. Their morphometry pipeline offers an automatic quantification of the segmented myelinated axons. Although ACSON largely automates the segmentation process, the user is still required to tune parameters such as similarity or intensity thresholds. Of note, the annotation of mitochondria presents a particular difficulty and still requires more human intervention.

gACSON (Graphical User Interface for Automatic Segmentation of axONs), released in 2022, is an open source Matlab-based software designed for visualizing, segmenting, assessing, and analyzing the morphology of myelinated axons in 3D EM volumes of brain samples [65] (https://github.com/Andrea-Behanova/gACSON). Aside from the description of the software the publication also provides a useful summary of other available tools and their properties. gACSON allows for both automatic segmentation and manual correction. It automatically segments the intra-axonal space and corresponding myelin sheaths, enabling the calculation of various metrics such as axonal diameter, eccentricity, myelin thickness, or g-ratio. Building on their previous pipelines ACSON and DeepACSON [67], gACSON integrates additional segmentation algorithms and features. While supervised learning-based methods such as DeepMIB [68] or AxonDeepSeg require manually annotated training data, gACSON uses unsupervised segmentation without the need for such data. By performing instance segmentation, gACSON does not require any assumptions about axon shape, which differentiates it from other techniques such as MyelTracer [62] (2D, semantic segmentation), where axons are assumed to be approximately circular in cross-section. Additionally, gACSON automatically separates adjacent myelin sheaths, a task that MyelTracer requires user input for. Another novelty is an interactive approach for reviewing and correcting automatically generated segmentations. The gACSON pipeline includes denoising with the block-matching 4D filtering algorithm (BM4D) and semantic segmentation of myelin. A distinctive feature is the instance segmentation of myelin, which accurately maps myelin sheaths to the appropriate intra-axonal space instances, allowing precise calculation of parameters like myelin thickness and g-ratio. The tool was applied to six SBEM datasets from the somatosensory cortex of one sham-operated and two traumatic brain injury (TBI) rats, demonstrating its capability for automatic segmentation of brain tissue ultrastructure and correction of automatic segmentations.

Despite its advantages, gACSON has some limitations. The dataset size cannot exceed the available RAM, necessitating the tiling of large datasets into subvolumes. Furthermore, the instance segmentation can encounter inaccuracies when there are substantial differences in myelin sheath thickness of adjacent fibers, due to the watershed algorithm dividing myelin content equally between adjacent axons. Also, for more difficult conditions, e.g. due to pathologies, the unsupervised segmentation quality may be worse compared to approaches that learn from annotated data.

3.4 Validation approaches

Validation and quality control is an important aspect of (semi-) automated analysis workflows to ensure consistent analysis quality. The most common approach is visual inspection of segmentation results using an image viewer with segmentation overlay. All tools presented here support this feature. However, this approach still requires manual effort to inspect results. Consequently, for analyses of large datasets, e.g. large volume EM or high-throughput TEM, only samples of the segmentation results can be validated manually. Moreover, the analysis tools listed here often lack functionality to efficiently visualize such large datasets, requiring use of another specialized tool such as WebKnossos [69], Catmaid [70] or MoBIE [71]. An alternative to manual inspection is validating the segmentation results against manual or proof-read annotations of a sample of the dataset to be analyzed. Here, different metrics to evaluate semantic or instance segmentation can be used, see [72] for a comprehensive overview. Most of the tools listed here support such an automated evaluation, except Ilastik and MyelTracer. This approach offers a more objective evaluation and can also be used to compare different segmentation approaches. However, it is more labor intensive, due to the fact that it requires additional manual annotations. Furthermore, the annotated data has to be chosen such that it covers the variety of the data to be analyzed, otherwise risking an underestimate of the segmentation error. The annotated data also has to be strictly separated from annotations used for training any machine learning based approach to avoid such an underestimate.

In conclusion, the evolution of segmentation methods, from manual to automated deep learning approaches, is changing the field of myelin and axon analysis. While challenges remain, particularly in terms of precision and handling pathological variations, the advancements in AI and deep learning hold great promise for more accurate, efficient, and comprehensive analyses in the future.

4.1 Developmental myelination

During development myelin is generated as a specialized plasma membrane domain of oligodendrocytes. By spirally wrapping around axons, a sheath composed of dense layers of compacted membranes is formed (Figure 1). Details of this wrapping process were unclear and different models were proposed. According to the “carpet crawler” model a sheet-like structure of a broadened oligodendrocyte process contacts and then wraps the axon at full length [75], while the “liquid croissant model” proposed a small cell process spiralling around the axon before growing laterally [76]. Snaidero and colleagues have shown with a combination of methods including FIB-SEM and three-dimensional reconstruction that the leading edge of the growing myelin sheath was the inner tongue. This is the innermost myelin wrap that directly contacts the axon [77]. According to their model, this inner tongue expands and spirals underneath the continuously added layers of myelin membrane, while these layers are growing laterally along the axon towards the forming nodes of Ranvier with their non-compacted edges always in contact with the axon. These non-compacted edges finally formed the paranodal loops. This model highlighted the inner tongue as the growth zone and the role of non-compacted openings, also called cytoplasmic channels within the myelin sheath [78], as supply pipelines for myelin growth in development.

For correct developmental myelination, target recognition, myelin thickness, internode length and the formation of nodes of Ranvier must be tightly regulated. Although the mechanisms are not entirely clear, bidirectional interactions by immunoglobulin cell adhesion molecules (CAMs) like Cadm3/4 play important roles [79], [80]. The Ig-CAM myelin-associated glycoprotein (MAG) is located at the adaxonal surface of the internodal myelin sheath and excluded from compact myelin [81], [82]. Its extracellular domain is facing the periaxonal space and is involved in regulating axon/myelin apposition [83], [84]. Paranodal adhesion proteins contactin-1 (Cntn1) and contactin-associated protein (Caspr) at the axonal surface and neurofascin 155 (Nfasc155) at the glial side [85], [86], coordinate together with MAG myelin growth and paranode formation in development [87]. The latter study analyzed developmental myelination by live imaging in Zebrafish and by vEM using automatic tape-collecting ultramicrotome (ATUM) and FIB-SEM in young mice under double and triple knock-out conditions. In these mutants, three-dimensional reconstruction of EM data stacks revealed complex phenotypes. Double myelin sheaths were formed by either two sheaths growing on top of each other or by the formation of two leading edges in one myelin sheath with one wrapping around the axon and the other growing into the forming sheath. Taken together, the study established that both, MAG at the internode and the paranodal adhesion molecules Cntn1, Caspr and Caspr2 function synergistically to target the growing edge and keep the developing myelin sheath connected to the axon during the wrapping process [87].

Using the mouse optic nerve as model tissue, Djannatian and colleagues investigated the ultrastructural changes in myelin during development by SBEM [88]. In this developmental phase, myelin membranes were produced in excess in the form of complicated three-dimensional aberrant myelin structures, so-called outfoldings that extend between neighboring fibers and are formations of redundant myelin. The 3D data stacks at different ages revealed that many of these structures were engulfed and phagocytosed by microglia cells. These findings were complemented by live-imaging in zebrafish showing microglia activity and the dynamics of myelin formation and remodeling. This important study demonstrated that developmental myelination is a dynamic process that involves myelin overproduction, myelin retraction, remodeling and refinement by microglia.

A recent study by Iyer and colleagues addressed the question by which mechanism myelin sheath length is determined during developmental myelination and added another important detail to the current concept of the cell biology of myelination [89]. They showed that calcium signalling is specifically required for myelin sheath length determination, but not for the number of myelinated axons or myelin thickness. To determine myelin sheath length in the mouse optic nerve after genetic calcium attenuation they performed SBEM and 3D reconstruction. This revealed multiple myelin outfoldings and dysmorphic, shorter sheaths in the mutants. Taken together, their study demonstrated that calcium-regulated actin filament assembly in nascent sheaths is a mechanism to elongate and refine the myelin sheath in response to neuronal activity.

4.2 Properties and functions of myelin

The inner tongue, paranode and juxtaparanode of the myelin sheath were found to play an important role in myelin maintenance. As Meschkat and colleagues reported, in adulthood, newly generated myelin membranes were continuously added at the inner tongue in proximity to the paranodal loops [90]. In this investigation, myelin basic protein (MBP) – which is the only essential protein required for the compaction of myelin in the CNS, was used as a structural indicator of resident myelin. After the inducible conditional knock-out of exon 1 of the classical Mbp locus in adult mice, any newly formed myelin membrane was incapable to form compact myelin due to the absence of MBP. This resulted in a slow disappearance of the developmentally formed myelin sheath due to replacement by non-compacted myelin membranes. This structural change was visualized with FIB-SEM at different time points after knock-out induction thereby providing information about the slow turnover of a myelin internode with a half-life of 20 weeks in the optic nerve.

vEM was also used to address the fundamental question whether individual oligodendrocytes show a preference for axons of a specific diameter [91]. Single oligodendrocytes myelinate multiple distant axons whereby each myelin sheath is connected to the cell body by a fine process. Tanaka and colleagues applied SBEM to samples from mouse corpus callosum to investigate oligodendrocyte heterogeneity. They classified interfascicular oligodendrocytes into two subtypes depending on their arrangement in rows or as isolated cells within the white matter tract and analyzed the shape of single oligodendrocytes and their myelin sheath. This revealed some fundamental properties: These oligodendrocytes generally myelinated distinct axons and hardly formed multiple sheaths on the same axon. They also showed a cell-specific preference for axonal diameters and a characteristic myelin thickness indicating heterogeneity within the oligodendrocyte population.

Myelin function might differ between white matter tracts and cortical regions and also between different types of myelinated axons. Myelin sheaths in white matter tracts speed up action potential propagation and fidelity by reducing the local internodal capacitance [92], [93] and support axonal energy metabolism [94], [95]. Different to this, myelination in grey matter regions can be discontinuous with large unmyelinated gaps, as shown for pyramidal neurons in superficial layers by high-throughput vEM using the ATUM and scanning serial sections [16]. Array tomography and vEM further revealed that most myelinated axons in the cortex belong to inhibitory interneurons [17].

To address myelin function heterogeneity in cortical regions, Kole and colleagues investigated the role of myelination on axons of parvalbumin-expressing (PV⁺) basket cells [96]. For this study on structure-function relationship in fast-spiking inhibitory interneurons, they took advantage of publicly available vEM datasets of mouse visual cortex from the MICrONS Consortium (https://www.microns-explorer.org/ [97]) and complemented it with electrophysiology, live Ca²⁺ imaging and immunofluorescence data. In the EM dataset, the distribution of axonal mitochondria in myelinated, unmyelinated and intermittent myelinated axon branches was visualized in relation to the myelin sheaths of the axon. Their study revealed that axonal mitochondria show a cell-type specific distribution pattern and clustered in myelinated segments of axons of these PV⁺ interneurons. Mitochondria were also longer, larger, showed a higher aspect ratio and occupied a larger volume than in non-myelinated parts of the axon. Different to this, in excitatory layer 2/3 cortical pyramidal neurons, such a volume difference of mitochondria between myelinated and unmyelinated segments of the axon was not found. By investigating Ca²⁺ fluxes, Kole and colleagues detected low mitochondrial Ca²⁺ transients in axons and axonal mitochondria along myelinated internodes and large action-potential evoked Ca²⁺ responses at non-myelinated branch points and nodes of Ranvier. Myelin function in PV⁺ interneurons might therefore include the restriction of mitochondrial Ca²⁺ buffering to the non-myelinated areas and support of ATP production along the internodes to sustain high firing frequencies and synchronization with fast gamma oscillations typical for these neurons [96].

Previously undescribed myelin proteins found in the myelin proteome might have unexplored functions towards axonal integrity. Knock-out mouse models offer an opportunity to study these and learn more about the complex axon-glia relationship. For example, the tetraspan-transmembrane protein CMTM5 (chemokine-like factor-like MARVEL-transmembrane domain containing protein 5) was found enriched in oligodendrocytes and myelin [98]. Although CMTM5 is a minor component in the myelin proteome, Cmtm5-deficient mice developed an early-onset progressive axonopathy despite the formation of normal appearing myelin [99]. Axonopathy in this mutant was evaluated by FIB-SEM and three-dimensional reconstruction. The pathology was characterized as axonal degeneration visible in the form of myelin whorls that consist of membranes with myelin-typical periodicity lacking any indication of an axon. Three-dimensional analysis was required to determine that these whorls were indeed spherical membranous objects and not parts of larger assemblies like myelin outfoldings. These whorls were interpreted as remnants of axons which underwent Wallerian degeneration with preservation of neuronal cells bodies [100]. The observed amelioration in the presence of the Wallerian degeneration slow (Wld ^S ) mutation which can protect axons from several types of insult [101] further supported this interpretation [99].

Ultimately, these studies were performed to elucidate properties of normal myelin and discover functions of myelin proteins. In the following section we will discuss how vEM contributed to the understanding of pathological changes of myelin.

5.1 Inherited myelin disease

Myelinated axons exist in a close metabolic relationship with the myelinating oligodendrocyte [95], [102]. The importance of an intact axo-glia unit is highlighted by the impact of genetic defects affecting the myelinating glial cell on neuronal health [103]. One example is proteolipid protein (PLP), the major myelin protein in the CNS of mouse and human encoded by the PLP1/Plp1 gene [98], [104], [105]. Point mutations, gene duplication or deletion are the genetic cause of a spectrum of inherited X-linked dys-myelinating leukodystrophies from severe Pelizaeus Merzbacher disease (PMD) to the milder late-onset spastic paraplegia type 2 (SPG2) [106]. The Plp ^−/Y mouse is a model of SPG2 which develops large axonal swellings and neuropathy leading to the view that the performance of the ensheathing oligodendrocyte directly influences the health and survival of the myelinated axon [107], [108], [109]. Although the exact disease mechanism is unclear, length-dependent axonal degeneration is secondary to the glial defect and might be a consequence of focal axonal transport impairment [107]. Revisiting the phenotype of the Plp ^−/Y mouse by FIB-SEM yielded additional interesting information: The myelin sheath appears normally compacted [73], but is characterized by many large cytoplasmic/myelinic channels that might play a role in enhanced axonal support [110]. However, sustained metabolic alterations in this myelin mutant affect axonal mitochondria [109]. Swollen and shorter mitochondria were found along the axon and within axonal swellings that sometimes consisted entirely of mitochondria. In addition, so far undescribed accumulations of organelles at the inner tongue were observed.

Recently, in an extended vEM study, the axo-myelin unit in the Plp-deficient mouse was carefully scrutinized with manual 3D reconstruction of FIB-SEM stacks [46]. The Mag-deficient mouse model was as well included into this study. MAG was mentioned in the preceding section as an important adhesion protein in ensuring correct myelin wrapping in development. Despite the differences in location and functions of PLP and MAG, their deficiencies share the myelin feature of outfoldings [81]. This formation of redundant myelin occurs transiently in normal development (see preceding section and [88]) but becomes pathological in adulthood. The three-dimensional ultrastructural investigation now revealed the spatial dimensions and complex shapes of these outfoldings (Figure 2 and movie) [46]. Unexpectedly, this study found that lack of these structural myelin proteins not only impacted the myelin sheath itself, but also entailed axonal pathology characterized by diameter increase, axonal sprouting and anastomosis. Descriptions of such secondary axonal pathologies have been emerging only recently.

Figure 2:

Myelin pathologies: reconstruction of myelin (yellow) and axon (blue) in optic nerves of myelin mutants analyzed at postnatal day 75 based on a focused ion beam – scanning electron microscope (FIB-SEM) dataset (openly available, see section data availability). (A), (B), (C) 3D model; (A′), (B′), (C′) EM image with highlighted features (myelin yellow, blue axon). (A), (A′) from a Mag-deficient (Mag ^−/−) mouse. (B), (B′), (C), (C′) from a Plp-deficient (Plp ^−/Y) mouse. Scale bar 1 µm. Adapted from [46] (under Creative Commons Attribution License).

5.2 Models of white matter injury

While inherited leukodystrophies like PMD are rare, multiple sclerosis (MS) is one of the most prevalent causes of neurological impairment manifesting in young adulthood [111]. MS is a chronic inflammatory demyelinating disease characterized by lesion formation in white and grey matter areas of the CNS ultimately leading to axonal loss and disability [112]. In animal models of MS, amyloid precursor protein (APP)-positive axonal accumulations of organelles are often used as marker for acute neuroaxonal damage [113]. Using the mouse model of toxic demyelination with cuprizone [114], content and volume of axonal swellings in the corpus callosum were studied by SBEM and immunohistochemistry confirming that alterations in axonal transport leading to accumulations of APP⁺ vesicles and mitochondria are a hallmark of early axonal damage [115]. To elucidate the role of myelin and demyelination in axonal vulnerability, Schäffner and colleagues investigated axonal pathology in relation to the status of myelination in human MS lesions and mouse models of MS [116]. By 2D EM of MS lesions, they found that axonal pathology was characterized by either organelle accumulation in axonal swellings or by condensed axoplasm. Time-course studies in experimental autoimmune encephalomyelitis (EAE), cuprizone and lysophosphatidylcholine (LPC) mouse models of white matter injury revealed that axonal swellings were an early and reversible characteristic of axonal pathology. By SBEM they could show that these occurred only in axonal segments with residual myelin while adjacent demyelinated sections of the same axon appeared normal. In contrast, condensed axoplasm as a sign of late stage damage was only found in myelinated axons and less prevalent in a hypomyelinating mouse mutant which exhibits more unmyelinated axons due to a low expression of MBP. Based on these results a hypothesis was proposed that under inflammatory conditions, myelin becomes dysfunctional promoting axonal damage that could be prevented by efficient demyelination [116].

Currently, MS is incurable. Supporting endogenous remyelination to restore function is one important therapeutic goal. Therefore, a better understanding of this process is crucial for the development of possible therapeutic interventions. A wealth of evidence from mouse models indicates that neural progenitor cells (NPCs) and oligodendrocyte progenitor cells (OPCs) play a major role in remyelination after recruitment to the lesion site and differentiation into myelinating oligodendrocytes (reviewed in Ref. [117]). To investigate the remyelinating potential of residual and viable mature adult oligodendrocytes after damage and partial loss of their internodes Duncan and colleagues studied two large animal models of demyelination [118]. The first model is feline irradiated food-induced demyelination (FIDID) which resulted in spinal cord demyelination and concurrent remyelination. During recovery in the ventral and lateral spinal cord, a mosaic pattern of unaffected and remyelinated axons was observed which were identified based on their different g-ratios. Meticulous analysis by SBEM showed single oligodendrocytes with cytoplasmic connections to unaffected mature myelin sheaths and remyelinating myelin sheaths at different stages of wrapping. Similar findings were obtained in a non-human primate model of vitamin B₁₂ deprivation in the same study. These are strong indications that mature oligodendrocytes might participate in remyelination under certain conditions. The difference from previous findings could be due to the different model species as well as the type of injury or lesion formation that might be specific for the affected CNS region.

Response to white matter injury and remyelination requires coordinated multicellular tissue reactions [119]. To study functional and structural changes in a mouse model of LPC-induced demyelination, Androvic and colleagues developed spatial transcriptomics-correlated electron microscopy (STcEM) which combines two unlikely compatible techniques to visualize morphology and function of cells in a complex tissue environment [120]. This technique allowed to combine large-area vEM of lesions and spatial transcriptomics with multiplexed error-robust fluorescence in situ hybridization (MERFISH) [121] on adjacent tissue sections revealing functional heterogeneity in expression responses. For validation, immunocytochemistry and lipid staining-coupled single-cell RNA sequencing were applied. This type of method development and combination is exemplary for future possibilities to visualise complex processes with the ultimate goal of understanding disease mechanisms as a basis for meaningful therapeutic intervention.

6.1 Instrumentation and sample preparation

A variety of instruments and different approaches are currently available for vEM. Further technical development can be expected in the field of FIB-SEM with the invention of ‘enhanced FIB-SEM’ [122] or PFIB-SEM improving acquisition time, resolution in X, Y and Z, stability of the sample during image acquisition and homogeneity of the milling procedure. In the field of fast image acquisition by multibeam-SEM, FAST-EM (Fast, Automated Scanning Transmission Electron Microscopy) utilizes a new optical transmission detection technique on sections mounted on a scintillator that converts electrons into photons [123], [124]. Kievits and colleagues now implemented an array tomography workflow for vEM on an early adopter FAST-EM system [125].

Sample preparation procedures must be adapted to the respective properties of the sample. In principle, these protocols are variations of a heavy metal-rich formulation like the classic rOTO protocol, at least in neuroscience. New protocols were developed to overcome side effects of fixation like cell swelling and preserve the extracellular space in brain tissue [126]. It was also shown in this study that this protocol increased the efficacy of heavy metal infiltration. Indeed, as sample sizes increase, effective infiltration methods will be required, possibly with application of microwave-assisted processing [26] or agitation methods [127]. There might be development in the field of resins depending on how the existing formulations cope with new milling procedures as applied in PFIB-SEM. For serial section imaging, recent methods development aiming at high-throughput collection of thousands of sections directly on wafer greatly improved the efficacy and quality by abolishing the need for serial section collection [128], [129]. Both studies applied addition of a thin layer of magnetic particle containing resin on the side of the sample. After sectioning, a dense cloud of freely floating sections was collected with the help of a permanent magnet and deposited onto wafer. The serial order of the sections was retrieved after imaging, e.g. by light microscopy when the magnetic resin also contained fluorescent particles [129]. This method is compatible with correlative 3D light and electron microscopy workflows (vCLEM).

Correlative workflows facilitate imaging across the scales and help targeting the structure of interest within a larger tissue context. A review worth reading by Chris Guerin and Saskia Lippens tells the history of moving CLEM into the three dimensional space [130]. From this active community many innovative developments are expected in the future. For example, embedding procedures which are compatible with fluorescence imaging facilitate easier correlation in vCLEM [131]. Wide-field fluorescent image acquisition integrated in the SEM using cathodoluminescent markers was used to map the regions of interest for high-resolution SEM image acquisition on serial sections [132]. Combining imaging modalities like intravital fluorescence microscopy or functional imaging with X-ray tomography for targeting in vEM helped to find rare events [133] or to link function to structure [134].

6.2 Data handling

Large datasets produced by vEM are extremely valuable if they meet the FAIR (Findability, Accessibility, Interoperability, and Reuse of digital assets) principles of data policy. Well described datasets including metadata should be openly available via repositories such as Nanotomy (http://www.nanotomy.org/OA/), EMPIAR (Electron Microscopy Public Image Archive, https://www.ebi.ac.uk/empiar/), neurodata.io (https://neurodata.io/) or the Image Data Repository (https://idr.openmicroscopy.org/). To enable further analysis in the spirit of open science, these data should be accessible on demand with a tool such as CATMAID [70], MoBIE [71] or WebKnossos [69]. Such datasets are not only required for potential scientific re-use [135], but manually annotated and segmented open datasets are also needed to train neuronal networks.

The automatic quantification of myelinated axons becomes increasingly important due to the larger availability and throughput of vEM for myelin research. Several user-friendly tools for automated analysis of myelin and electron microscopy exist thanks to the wide adoption of machine and deep learning in bio-medical image analysis. However, these tools are still limited in their generalization to different imaging conditions and hampered by difficulties in training deep neural networks on newly annotated data. Two recent trends promise to remove these hurdles and make advanced image analysis fully available to biologists. First, the recent development of a common model format and a resource for sharing deep neural networks, BioImage.IO [136], enables re-use of trained networks within user-friendly software such as Ilastik [61] and Deep MIB [68] as well as integration with ZeroCost4DL [137] a tool for easy network training. Due to these trends researchers will be able to further train networks on their EM myelin data and share the results with the community. Second, the emergence of “foundation models” for microscopy data. Foundation models are powerful deep learning models trained on massive datasets to solve a number of different tasks of a specific category. While these models require a large data and computational footprint for initial training, the availability of segmentation foundation models, in particular Segment Anything [138] has enabled their adoption to microscopy [139] and enables fast interactive and automatic segmentation of complex structures in EM and other microscopy modalities. Further developments of these methods promise powerful segmentation approaches that can be adapted to new data from very few examples. Given the further automation and ease-of-use of 3D segmentation tasks, the development of better models for 3D morphology that go beyond simple measures like the g-ratio will also become increasingly important. Many parameters can be extracted from any segmented object emphasizing the importance to convert 3D data into meaningful metrics. One example can be found in the analysis of 3D axonal properties shown by Steyer and colleagues [46] extracting curvature, torsion, circumference changes and eccentricity. This exemplifies the need for interdisciplinary research bringing together neuroscience, computational science and mathematics.

6.3 Further perspectives

Currently, large 3D imaging of brain tissue and automated reconstruction from of an increasing number of species will inevitably reveal a plethora of unknowns that have been overlooked. Besides this, systematic comparisons of mutant versus wildtype in smaller datasets although deemed less spectacular, are essential to test exciting hypotheses. However, of particular interest are human diseases leading to neurodegeneration e.g. MS and natural processes like ageing. A deep-learning workflow specifically developed for myelin segmentation in human autopsy samples is available [140]. Multimodal approaches as pioneered by Androvic and colleagues [120] offer unique insight into cellular activity, morphology and distribution in a disease context. Other fields of research beyond neuroscience will profit from such methods development. Although many questions have been investigated using conventional 2D TEM, there is much more to explore in 3D to improve our understanding of the interplay between different components in the central and peripheral nervous system. In light of the current technical advance in the field of instrumentation, sample preparation, imaging and data analysis, the future holds a wide range of opportunities.