Unsupervised clustering of AxonDeepSeg morphometrics uncovers early patterns of axon-myelin secondary degeneration following neurotrauma

Unsupervised clustering of AxonDeepSeg morphometrics uncovers early patterns of axon-myelin secondary degeneration following neurotrauma | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Unsupervised clustering of AxonDeepSeg morphometrics uncovers early patterns of axon-myelin secondary degeneration following neurotrauma Parth Patel, Brittney R. Lins, Sarah C. Hellewell, Terence McGonigle, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8951825/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Reliable quantification of axon and myelin ultrastructure is essential for understanding white-matter pathology, yet conventional manual methods are subjective, labour-intensive, and limited to compactly myelinated fibres. Here, we adapted and validated a custom AxonDeepSeg model for automated segmentation of transmission electron micrographs from the rat optic nerve following partial transection or sham surgery. The trained model achieved accuracy and Dice similarity indices comparable to published benchmarks, and automated axon diameter measurements closely matched manual quantification. Beyond reproducing traditional morphometrics, AxonDeepSeg detected subtle yet significant increases in g-ratio, solidity, and axon-to-fibre area ratio two weeks post-injury-changes that were undetectable by manual analysis-indicating early alterations in axon-myelin geometry preceding overt demyelination. Unsupervised clustering of multidimensional morphometric data identified six distinct axon–myelin phenotypes, two of which (Clusters 1 and 2) exhibited pronounced injury-related remodelling not apparent on assessment of the full, un-clustered dataset. These findings demonstrate that automated morphometry not only enhanced throughput and reproducibility but also increased sensitivity to otherwise undetectable microstructural changes in subsets of myelinated axons. This integrative workflow provides a scalable framework for identifying early and subtle indicators of axonal pathology and secondary degeneration in neurotrauma and other demyelinating conditions. Transmission electron microscopy Axons Myelin Machine learning AxonDeepSeg Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Myelin, the lipid-rich sheath encasing axons, is essential for rapid and efficient transmission of action potentials within the nervous system [ 1 ]. Damage to myelin - whether caused by neurotrauma or neurodegenerative disease - contributes to the cognitive, sensory, and motor deficits characteristic of these conditions [ 2 – 4 ]. Following neurotrauma, a cascade of metabolic and cellular disturbances renders uninjured axons adjacent to the primary lesion vulnerable to progressive secondary degeneration [ 5 – 7 ]. Myelinating oligodendrocytes are particularly susceptible to secondary degeneration due to their high metabolic demand and low antioxidant capacity [ 8 , 9 ], making them prone to oxidative stress and lipid peroxidation. The resulting dysmyelination and myelin decompaction can exacerbate axonal dysfunction and impair repair processes, underscoring the need for reliable methods to quantify myelin integrity and assess therapeutic strategies that preserve myelin structure [ 10 , 11 ]. Transmission electron microscopy (TEM) remains the gold standard for quantifying axonal and myelin morphology at ultrastructural resolution. Conventional morphometric analyses assess parameters such as axon diameter, myelin thickness, and g-ratio, with injury typically inducing reductions in myelin compaction and increased variability in axon calibre. For example, the extent of myelin decompaction can be quantified by measuring the percentage of the axonal circumference surrounded by decompacted laminae [ 12 , 13 ]. However, these analyses are typically performed manually using image-processing software such as ImageJ [ 14 , 15 ], which are labour-intensive, subjective, and limited in throughput. Hundreds of axons across multiple regions must be measured to achieve statistical power, and such manual approaches are prone to user bias and inconsistent criteria across studies. This has contributed to inter-study variability and limits reproducibility when comparing injury models or therapeutic interventions [ 12 , 16 – 23 ]. An automated approach capable of segmenting, measuring, and classifying axonal and myelin morphology, including irregularly compacted sheaths, would therefore provide a more objective and scalable method for ultrastructural analysis. Automated morphometric analysis relies on accurate image segmentation; the partitioning of microscopic images into regions corresponding to biological structures such as axons and myelin [ 24 ]. Traditional semi-automated segmentation techniques such as thresholding [ 25 ], watershed [ 26 , 27 ]. or region growing [ 28 , 29 ] require manual pre- or post-processing, introducing user-dependent subjectivity. These approaches are also often restricted to specific imaging modalities or contrast conditions, limiting their broader applicability. In contrast, convolutional neural networks (CNNs) overcome many of these limitations by learning contextual features directly from training data, enabling fully automated segmentation without user intervention. Once trained, CNN-based models can segment images in seconds, offering major improvements in reproducibility, throughput, and efficiency. A range of CNN-based segmentation frameworks have been developed for electron microscopy data, each varying in usability, accessibility, and analytical depth. While tools such as Microscopy Image Browser [ 30 ], AxonSeg [ 31 ], and others [ 32 – 34 ] have achieved high-quality segmentation, most are limited by one or more of the following factors: lack of an intuitive graphical interface, incomplete open-source availability, or morphometric outputs which are restricted to basic geometric features. Many of these models also require advanced coding expertise or provide only two-dimensional functionality, limiting adoption by experimental neuroscientists. AxonDeepSeg [ 35 ] overcomes these constraints by combining full open-source accessibility, compatibility with multiple imaging modalities (electron-, optical-, and light-microscopy), and a streamlined command-line or graphical user interface-based workflow that requires minimal computational experience. Critically, AxonDeepSeg provides an expanded suite of morphometric parameters, from classical metrics (axon diameter, g-ratio) to higher-order geometric descriptors such as solidity, eccentricity, and orientation, enabling multidimensional quantification of axon and myelin ultrastructure. These features make AxonDeepSeg particularly suited for objective, high-throughput, and reproducible quantification of axonal pathology. In this study, we trained a custom AxonDeepSeg model using TEM images of the rat optic nerve following partial transection injury - a well-established model of secondary degeneration [ 36 ]. Beyond validating the performance of AxonDeepSeg for automated segmentation, we leveraged its diverse morphometric output to quantify both conventional features and additional geometric parameters such as solidity, eccentricity, and orientation. We further introduced the axon/fibre area ratio as a novel metric to capture subtle changes in myelin decompaction. Finally, by applying unsupervised clustering to these multidimensional morphometric data, we identified previously unrecognised axon–myelin phenotypes and revealed distinct patterns of structural change following injury. This framework fills a key methodological gap by enabling objective, high-throughput, and reproducible analysis of axonal ultrastructure, providing a scalable basis for cross-study comparison of injury- and disease-related degeneration. Materials and Methods Animals Female Piebald Virol-Glaxo rats (n = 10; 3 months old; 160–200 g; Animal Resource Centre, Murdoch, Western Australia, Australia) were randomly allocated to receive either optic nerve partial transection or sham control procedure (n = 5/group), an adequate sample size for electron microscopy analyses [ 37 , 20 , 12 , 13 ]. The experimental unit was individual rats. Randomisation was achieved at a cage level by allocating animals to groups in the order in which they were received from the supplier. Additional confounders were not controlled. Rats were housed two to three per standard polypropylene cage under specific-pathogen-free conditions, with food and water available ad libitum . Following arrival at the facility, rats were left to acclimatise for a minimum of one week prior to initiating procedures. All procedures were conducted in accordance with the Australian Code for the Care and Use of Animals for Scientific Purposes (National Health and Medical Research Council, 2013) and approved by both The University of Western Australia Animal Ethics Committee (RA/3/100/1485) and Curtin University Animal Ethics Committee (ARE2017-4). Partial optic nerve transection Partial transection of the right optic nerve was performed as previously described [ 38 ] (Fig. 1 a). Rats were anaesthetised via intraperitoneal injection of xylazine (10 mg/kg) and ketamine (50 mg/kg). An incision was made above the right eye to expose the optic nerve, and a 200 µm-deep dorsal incision was made 1 mm posterior to the globe using a diamond radial keratotomy knife. The incision above the eyelid was sutured post-operatively, and animals received analgesic (Norocarp, 2.8 mg/kg) and antibiotic (Neomycin, 10 mg/kg) treatment. Sham-operated rats underwent identical surgical procedures without the optic nerve incision. Each rat received one procedure, one time. Following recovery from anaesthesia rats were returned to their home cage. Optic nerve partial transection was used because it produces distinct and reproducible primary and secondary injury sites, ideal for investigating secondary degeneration [ 36 ]. No animals were excluded from the study due to pre-defined inclusion or exclusion criteria (i.e. intolerability to anaesthesia, premature death, welfare concerns, surgical errors or complications, or anatomical abnormalities). Transmission electron microscopy Tissue processing and imaging were conducted as previously described [ 38 , 37 ]. Two weeks after injury or sham surgery, rats were euthanised with pentobarbitone sodium (850 mg/kg) and phenytoin sodium (125 mg/kg; intraperitoneal) and perfused transcardially with 0.9% saline followed by 2% paraformaldehyde/2.5% glutaraldehyde/2% sucrose in 0.1 M phosphate buffer (pH 7.2). Optic nerves were dissected, separated from the dural sheath, stored in 0.13 M Sorenson’s phosphate buffer (pH 7.2) and post-fixed in 1% osmium tetroxide (Electron Microscopy Sciences, ProSciTech, Townsville, QLD, Australia; Cat# C011). Samples were dehydrated through an ethanol gradient, transitioned to propylene oxide, and embedded in epoxy resin (Araldite Procure, ProSciTech; Cat# 039). Ultrathin cross-sections (100 nm) were prepared using an LKB Nova ultramicrotome (Bromma, Sweden), stained with uranyl acetate and lead citrate, and imaged using a JEOL 2100 transmission electron microscope (Tokyo, Japan) at 4000× magnification. Ten images per nerve were captured from the ventral region of the optic nerve, spatially distant from the primary injury site but vulnerable to secondary degeneration, using Olympus iTEM Soft Imaging Solution software. A 3.05 mm copper support grid mounted beneath the sections ensured uniform segmentation and facilitated random sampling across the ventral optic nerve. Training and evaluation of the AxonDeepSeg model The AxonDeepSeg model (version 3.30) was developed on an NVIDIA GTX 1070 GPU using a dataset of 10 TEM images (see Fig. 1 a for a representative original image). Model performance was assessed using 5-fold cross-validation [ 39 ]. The dataset was randomly partitioned into five subsets, each containing two images. For each fold, eight images were used for model training, and the remaining two images were held out for evaluation. As AxonDeepSeg requires a validation dataset during training, one of the two held-out images was designated as a validation input. This validation image was used solely for satisfying the software workflow requirements and was not used for any hyperparameter tuning, model selection, or early stopping. Both held-out images were therefore independent of the training process and were used to compute performance metrics. As a result, five independent models were trained, and performance metrics were averaged across folds to obtain an overall estimate of model performance. We reported pixel-wise metrics, including False Negative Rate, False Positive Rate, Accuracy and Dice Similarity Coefficient and the ability of AxonDeepSeg to correctly identify axon versus myelin regions was compared using Student’s t-tests. Additionally, a two-way ANOVA was used to compare AxonDeepSeg-derived axon diameter measurements in sham and injured optic nerves with manually obtained measurements, which demonstrated consistency in morphometric estimation (Fig. 2 ). Following cross-validation, a final model was trained on the full dataset (all 10 images) using identical training parameters and used for the subsequent segmentation. Ground-truth labels (Fig. 1 b) were generated using the default AxonDeepSeg model, manually corrected in Adobe Photoshop CC, and reviewed independently by a second researcher before training. Blinding was achieved at the level of outcome assessment and data analysis by a third party assigning coded image identifications prior to manual measurements, with true identification revealed at completion. Blinding during allocation and conduct of the experiment was not possible due to the surgical requirements of the study. Schematic overview of the optic nerve partial transection model in Piebald Virol-Glaxo rat highlighting the surgical site behind the right eye. A pop-out TEM image of an injured optic nerve cross section showcases the characteristic ‘mushroom’ morphology at the primary injury site on dorsal aspect. A red frame in the ventral region represents the site of image collection and a second pop-out window shows an example of the high magnification TEM micrographs of axon and myelin cross sections. Myelin appears electron dense (dark) while axons appear lighter (a) . The workflow for training AxonDeepSeg illustrated by a ground truth label generated by manual segmentation of the original TEM micrograph and the corresponding automated segmentation generated by AxonDeepSeg.. The ground truth labels contain 3 colour values: white (axon), grey (myelin), and black (background). The trained segmentation images contain 3 colour values: blue (axon), red (myelin), and black (background; b ). A suite of established and new morphometric features was assessed in AxonDeepSeg and manually as outcome measures (Table 1 ). To ensure a robust morphological representation, all axon and myelin profiles within the selected frames were included in the training process, regardless of shape or size. The images were split into 256 × 256-pixel patches and training was performed with a batch size of 4 for 300 epochs, taking approximately 2 hours per fold, for a total training time of approximately 10 hours. The final trained model was then used to perform automated segmentation (Fig. 1 b) on 68 images containing 1,213 total axons to generate the morphometric dataset. The segmented images and their corresponding morphometric outputs were visually quality checked. Decompacted interlamellar spaces that were incorrectly classified as axons were identified using their assigned object IDs, and the corresponding morphometric data were removed from the raw output prior to analysis. Table 1 Definitions of morphometric features quantified by AxonDeepSeg as outcome measures Features Definition g-ratio Ratio of the axon diameter to the total fibre diameter (axon + myelin). Axon area (µm²) Cross-sectional area of the axon compartment. Axon perimeter (µm) Length of the boundary enclosing the axon region. Myelin area (µm²) Difference between the fibre area and the axon area, representing the myelin sheath and peri-axonal gap. Axon diameter (µm) Diameter of the axon, measured along the minor axis when modelled as an ellipse. Fibre area (µm²) Cross-sectional area of the entire fibre (axon + myelin). Solidity Compactness of the axon, defined as the ratio of axon area to the area of a convex polygon containing the axon. Lower values indicate irregular or decompacted morphology. Eccentricity Measure of axonal circularity or flattening, where 0 = perfect circle and values approaching 1 indicate elongation. Orientation Angle between the major axis of the axon and the horizontal axis of the image. Axon/Fibre area ratio * Ratio of axon area to fibre area, indicating the relative thickness of the myelin sheath. Note. Definitions adapted from the AxonDeepSeg documentation ( https://axondeepseg.readthedocs.io/en/latest/documentation.html#morphometrics-file) . * indicates metrics not exported by AxonDeepSeg but calculated subsequently. Analysis of axon and myelin morphology To determine whether AxonDeepSeg-derived morphometric data could detect injury-related changes in axon and myelin structure, unpaired Student’s t -tests were used to compare sham and injury groups for each independent morphological parameter. Normality and homogeneity of variance were verified using the Shapiro–Wilk and Fligner–Killeen tests, respectively. All assumptions were met except for the orientation parameter (W = 0.7525, p = 0.004), for which a non-parametric Mann–Whitney U test was applied. Unsupervised clustering of AxonDeepSeg-derived data To explore fibre-level heterogeneity and identify morphologically distinct axon-myelin phenotypes, an unsupervised clustering analysis was performed on the complete morphometric dataset derived from AxonDeepSeg. The dataset comprised quantitative descriptors for each segmented axon, including both traditional ultrastructural features (axon diameter, g-ratio, myelin area) and non-traditional geometric parameters (solidity, eccentricity, orientation, and axon/fibre area ratio). These features were chosen to capture both the physical and topological variability of axons and their ensheathing myelin. All morphometric variables were standardised (z-score normalisation) prior to dimensionality reduction. Principal component analysis (PCA) was applied using the scikit-learn library [ 40 ] to reduce data dimensionality while preserving the major sources of variance in the dataset. The first few principal components, which cumulatively explained most of the variance, were then used as inputs for spectral clustering. This method was selected for its robustness to non-linear relationships and ability to identify subgroups within datasets containing correlated features [ 41 ]. Spectral clustering was implemented in scikit-learn [ 40 ], which computes a similarity matrix of samples based on feature proximity and transforms it into a low-dimensional embedding of the data, the labels are then assigned using the discretize method. The number of clusters was determined empirically by evaluating model stability and silhouette scores across different cluster counts. Six stable clusters were identified and retained for subsequent analyses. To verify the distinctness of each cluster, post-hoc statistical comparisons were conducted across all morphometric features using the Games–Howell multiple comparisons test, which is appropriate for unequal sample sizes and heterogeneous variances. Bootstrapping was applied to confirm stability of the resulting group means. Clusters 1–6 were then characterised by their mean g-ratio, axon diameter, and associated morphometric profiles. Comparison of sham and injury groups within AxonDeepSeg-derived clusters To examine injury effects within the phenotypic clusters defined by AxonDeepSeg, unpaired Student’s t -tests were used to compare axon and myelin features between sham and injury groups. Clusters 3–6 were excluded from these comparisons due to insufficient sample sizes and incomplete animal representation. Clusters 1 and 2 contained sufficient samples to allow reliable statistical comparison. Results Automated segmentation of optic nerve ultrastructure using AxonDeepSeg Electron micrographs of optic nerve cross sections were used to train the AxonDeepSeg model. A randomized dataset of 10 images of the ventral optic nerve collected 2 weeks after partial transection or sham surgery was used for model training and evaluation. Model performance was assessed using 5-fold cross-validation, and a final model was trained on all 10 images for downstream segmentation of axons and myelin in TEM images from the same optic nerves. Visual inspection confirmed that the trained AxonDeepSeg model effectively delineated axon and myelin boundaries across varying fibre morphologies, enabling reliable extraction of morphometric features for downstream analysis (Fig. 1 b). Validation of AxonDeepSeg model performance The performance of the model was evaluated on the test image set. Pixelwise accuracy represents the proportion of correctly classified pixels within each compartment. The model achieved comparable pixelwise accuracy for axon segmentation (92.15%) and myelin segmentation (91.26%; t (18) = 1.440, p = 0.1671; Fig. 2 a). Dice similarity achieved for axon segmentation (91.87%) was significantly higher than for myelin segmentation (85.74%; t (18) = 9.429, p < 0.0001; Fig. 2 b). In accordance with this, false-positive rates for myelin classification (7.67%) were significantly higher than axon classification (3.62%, t (18) = 3.282, p = 0.0041; Fig. 2 c). In contrast, false-negative rates were not significantly different regarding axon (11.84%) and myelin (11.47%) classification (t (18) = 0.2024, p = 0.8419; Fig. 2 d). Importantly, axon diameter measurements derived from the AxonDeepSeg model did not differ from manually obtained values (F (1,16) = 1.446, p = 0.2466; Fig. 2 e), validating the accuracy of the automated model against an established manual morphometric analysis. AxonDeepSeg identified injury-associated changes in axon and myelin morphology Automated morphometric analyses revealed injury-specific alterations in optic nerve ultrastructure two weeks after partial transection. Automated analysis compared sham and injured optic nerve micrographs and revealed subtle but significant effects of injury, such as increased g-ratio (t (8) = 2.661, p = 0.0288; Fig. 3 a), solidity (t (8) = 3.040, p = 0.0161; Fig. 3 g), and axon-to-fibre area ratio (t (8) = 2.915, p = 0.0194; Fig. 3 j), demonstrating sensitivity to subtle ultrastructural alterations and shifts in axon-myelin relationships following partial optic nerve injury. No significant differences were detected in axon diameter, axon perimeter, axon area, myelin area, eccentricity, orientation, or fibre area (all p > 0.05; Fig. 3 b,c,d,e,f,h,i). Unsupervised clustering reveals distinct axonal phenotypes To capture fibre-level heterogeneity, unsupervised clustering was applied to the morphometric dataset (Fig. 4 a). This approach, integrating both traditional (e.g., axon diameter, g-ratio) and non-traditional (e.g., solidity, eccentricity) morphometric descriptors, identified six distinct axon–myelin phenotypes. Each cluster displayed a unique combination of quantitative features corresponding to recognisable ultrastructural profiles. Representative TEM images of each cluster (Fig. 4 b) illustrate visual differences between clusters. Cluster 1 (n = 1080) comprised of compactly myelinated axons with moderate diameter and low g-ratio, representing structurally intact fibres. Cluster 2 (n = 92) included myelinated axons of larger diameter with partially decompacted myelin and higher g-ratios. Cluster 3 (n = 6) had reduced g-ratio and smaller fibre areas, consistent with axons surrounded by thickened, decompacted myelin. Cluster 4 (n = 9) was characterised by large axon and fibre areas, extensive myelin disruption, and lower solidity. Cluster 5 (n = 20) contained small, unmyelinated axons, while Cluster 6 (n = 5) were large-diameter axons with moderately decompacted myelin and high g-ratio values. The mean values of morphometric features for each cluster (Table 2 ) provide a practical reference for identifying similar phenotypic classes in independent datasets. Importantly, these values include routinely measurable parameters such as axon diameter and g-ratio, allowing researchers to approximate cluster membership using conventional manual morphometry, even without access to automated segmentation tools. Table 2 Mean morphometric values for each cluster of optic nerve axons-myelin units Feature Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Cluster 6 g-ratio 0.798ᵈ 0.821 0.641 0.786ᵃ 0.994 0.859 Axon area (µm²) 0.360ᶜ 1.010 0.300ᵃᵉ 1.590ᶠ 0.207ᶜ 1.900ᵈ Axon perimeter (µm) 2.850ᶜ 5.320 3.300ᵃ 9.360 1.850 7.260 Axon diameter (µm) 0.522 0.889ᵈ 0.403ᵉ 0.968ᵇ 0.412ᶜ 1.340 Fibre area (µm²) 0.607 1.820 1.320 3.240ᶠ 0.211 3.040ᵈ Solidity 0.930ᶠ 0.869ᶜᶠ 0.804ᵇᵈᶠ 0.704ᶜ 0.976 0.885ᵃᵇᶜ Eccentricity 0.742ᵉᶠ 0.778ᶠ 0.891ᵈ 0.909ᶜ 0.687ᵃᶠ 0.715ᵃᵇᵉ Orientation 0.103ᵇᶜᵈᵉ 0.190ᵃᶜᵈᵉ 0.143ᵃᵇᵈᵉᶠ 0.194ᵃᵇᶜᵈᵉᶠ 0.007ᵃᵇᶜᵈ 0.739 Axon/Fibre area ratio 0.592 0.557ᵈ 0.248 0.504ᵈ 0.988 0.626 Myelin area (µm²) 0.248 0.820ᶜ 1.020ᵇᶠ 1.650ᶠ 0.004ᶜᵈ 1.130 Note. Values represent mean morphometric measurements for each cluster derived from segmented TEM data of the ventral optic nerve. Superscripted letters indicate features not significantly different from the corresponding cluster (a = Cluster 1, b = Cluster 2, c = Cluster 3, d = Cluster 4, e = Cluster 5, f = Cluster 6). All other pairwise comparisons were significantly different across clusters ( p < 0.05; Games–Howell test). Mean g-ratio and axon diameter values provide accessible reference points for identifying comparable axon–myelin phenotypes using manually measured morphometry. Injury induced cluster-specific alterations in axonal phenotype To determine whether partial transection injury elicited distinct effects within specific axon populations, morphometric comparisons between injury and sham conditions were conducted for the more internally consistent Clusters 1 and 2 (Figs. 5 – 6 ). In Cluster 1, injury resulted in significant increases in g-ratio (t (8) = 2.837, p = 0.0219; Fig. 5 a), solidity (t(8) = 3.048, p = 0.0157; Fig. 5 b), axon-to-fibre area ratio (t (8) = 2.972, p = 0.0178), axon diameter (t(8) = 2.651, p = 0.0292; Fig. 5 a), and axon area (t(8) = 5.281, p = 0.0007; Fig. 5 e). In Cluster 2, injury similarly increased g-ratio (t (8) = 2.809, p = 0.0229; Fig. 6 a), solidity (t (8) = 2.755, p = 0.0248), and axon-to-fibre area ratio (t (8) = 2.734, p = 0.0257), though axon diameter and area were unaffected (p > 0.1). Neither cluster exhibited significant changes in axon perimeter, myelin area, fibre area, eccentricity, or orientation (p > 0.05). Together, these findings demonstrate that partial transection alters key ultrastructural features of axon–myelin integrity in a cluster-dependent manner, revealing differential vulnerability among optic nerve fibre subtypes. Discussion The trained AxonDeepSeg model achieved high segmentation accuracy and reproducibility, enabling precise quantification of optic nerve ultrastructure following injury. Automated morphometric profiling revealed subtle but significant injury-associated increases in g-ratio, solidity and axon-to-fibre area ratio, indicative of early myelin decompaction. Unsupervised clustering uncovered six morphologically distinct axon populations, two of which (Clusters 1 and 2) exhibited pronounced injury-specific alterations. These data validate AxonDeepSeg as a robust tool for large-scale ultrastructural analysis and demonstrate its capacity to resolve previously unrecognised, cluster-specific responses of axon–myelin units to focal injury. Conventional methods used to quantify axon and myelin ultrastructural features are typically based on manual measurement and visual assessment. These methods are subjective and labour-intensive, as measurements from hundreds of axons across multiple fields of view are required for statistically robust analyses. Furthermore, traditional metrics such as myelin thickness and g-ratio can only be measured where the myelin appears compact, excluding axons with partially or fully decompacted lamellae and thereby omitting potentially relevant pathological features. These limitations may contribute to inter-study variability, confounding comparisons across models that use different measurement or classification criteria. Here, we described the modification and use of an existing AxonDeepSeg model to automatically segment axons and myelin and measure a broad range of morphometric features in transmission electron micrographs of the rat optic nerve following PT or a sham control injury. The trained model achieved high segmentation accuracy across multiple validation metrics, comparable to or exceeding previously published benchmarks [ 33 , 35 ]. Automated axon diameter measurements were consistent with manual measurements [ 37 ], confirming that AxonDeepSeg accurately reproduced standard morphometric parameters with the added benefit of increased throughput and reduced bias. The automated analysis also detected significantly increased g-ratio (Fig. 3 a), which was not detected during manual measurement of the same images [ 37 ], implying AxonDeepSeg may be more sensitive to subtle and early injury effects than manual analysis. Additionally, AxonDeepSeg detected injury-effects in the non-traditional morphometrics of solidity and axon-to-fibre area ratio (Fig. 3 g,j), further underscoring its enhanced sensitivity to detect early ultrastructural changes and altered axon-myelin relationships, and provide new insights to the morphological changes following partial optic nerve injury. Furthermore, we show that unsupervised clustering of AxonDeepSeg-derived morphometric data can be used to identify distinct categories of axons, and that these axon populations respond differently to injury. This approach enables a more objective and biologically meaningful understanding of abnormal myelination in the context of injury and secondary degeneration. Unlike conventional methods, which cannot reliably quantify g-ratio or myelin thickness in partially decompacted fibres, AxonDeepSeg permitted analysis of both conventional and non-conventional morphometric features across all axons, including those with irregular or disrupted myelin. Although segmentation of myelin was slightly less precise - likely reflecting the morphological heterogeneity of compact and decompacted sheaths - automated segmentation allowed inclusion of axons that would typically be excluded from manual analyses. Automated analysis reveals early injury-related ultrastructural alterations AxonDeepSeg-based morphometry detected subtle but consistent injury-associated changes two weeks after partial optic nerve transection, a time point previously reported to lack overt morphological differences by manual analysis [ 37 ]. Injury increased g-ratio, solidity, and axon-to-fibre area ratio, whereas parameters such as axon area, fibre area, and perimeter remained unchanged. These findings indicate a mild reduction in relative myelin thickness and shape alterations suggestive of local structural stress rather than gross swelling. Increased solidity was evident in both the overall dataset and in the two largest fibre clusters (Clusters 1 and 2), while higher axon-to-fibre area ratio and g-ratio values suggest subtle thinning of the myelin sheath relative to the axon calibre. In the full dataset, these changes occurred without accompanying increases in axon diameter or fibre area, indicating a shift in axon-myelin geometry independent of swelling [ 42 ]. In Cluster 1, however, increased solidity coincided with small increases in axon diameter and area, suggesting that some fibres may undergo mild distension or cytoskeletal reorganisation [ 43 ]. The biological interpretation of solidity remains exploratory, but it likely reflects geometric changes in the axonal contour rather than size alone. Axon shape is determined by cytoskeletal architecture, organelle positioning, and extracellular constraints [ 44 ]. Subtle increases in solidity could therefore indicate reduced undulation or micro-beading of the axonal boundary, possibly due to early cytoskeletal disruption or redistribution of intracellular contents - changes known to occur after axonal stress or oxidative damage [ 45 , 46 ]. These features would not be detected by traditional morphometrics, highlighting the added sensitivity of automated segmentation. Consistent with our previous work, manual analyses of the same cohort found no overt changes in axon diameter or myelin appearance at this time point [ 37 ]. The present study confirmed agreement between manual and automated axon diameter measurements, while AxonDeepSeg revealed small but significant increases in g-ratio, solidity, and axon-to-fibre area ratio in the same images, demonstrating its superior sensitivity to early, pre-decompaction alterations in myelin geometry. Small increases in g-ratio and axon-to-fibre area ratio imply local thinning of the myelin sheath relative to the axon, possibly reflecting early lamellar separation or metabolic stress in oligodendrocytes [ 47 , 1 ]. Even subtle changes in myelin thickness or axonal geometry can influence conduction velocity and synchronisation within white-matter tracts [ 48 , 49 ]. Although the functional implications of altered solidity remain unclear, the convergence of these morphometric shifts suggests the onset of fine-grained structural adaptations within the axon–myelin unit that may precede overt decompaction or demyelination. AxonDeepSeg compared to existing segmentation tools A range of CNN-based segmentation frameworks have been developed for axon and myelin analysis, each varying in complexity, accessibility, and analytical breadth (Table 3 ). Table 3 Comparison of convolutional neural network (CNN)-based segmentation tools for axon analysis Software / Tool Reference Image type(s) Dimensionality Open-source availability GUI User friendliness Morphometric capability AxonDeepSeg [ 35 ] OM, (EM), (LM) 2D ✓ – ✓ ✓✓✓ AxonSeg / AxonSeg3D [ 32 ] EM 2D / 3D ✓ ✓ ✓✓ ✓✓ Microscopy Image Browser (MIB) [ 30 ] LM, EM 2D / 3D ✓ ✓ ✓✓ – Ilastik [ 34 ] EM 2D / 3D ✓ ✓ ✓ ✓ DeepAxon [ 50 ] OM 2D ✓ – ✓ ✓ ACSON / DeepACSON [ 51 ] EM 2D / 3D ✓ – – ✓✓ MiGA [ 33 ] EM 2D ✓ – ✓ ✓✓ AxonEM [ 52 ] EM 3D ✓ – – ✓ AxonDeepSeg with PairedImageTranslation [ 53 ] OM 2D – – – ✓ multi-class CNNs and deep encoder-decoder architecture [ 54 ] EM 2D – – – – AxonDeep [ 55 ] LM 2D – – – ✓ Note. OM = optical microscopy; LM = light microscopy; EM = electron microscopy. User friendliness was rated according to ease of adoption: ✓ = usable without prior expertise; ✓✓ = usable without detailed instruction. Morphometric capability was rated according to the range of measurements provided: ✓ = basic (classical) measures; ✓✓ = extended (modern) measures; ✓✓✓ = advanced or unique morphometric metrics. Many frameworks, such as Microscopy Image Browser [ 30 ], DeepACSON [ 51 ], and MiGA [ 52 ], achieve excellent segmentation fidelity but lack intuitive user interfaces or provide only basic geometric measurements. AxonDeepSeg offers several advantages for experimental neuroscientists: it is fully open-source, compatible with multiple microscopy modalities, and deployable through either command-line or graphical interfaces requiring minimal computational expertise. It also outputs both classical morphometric parameters (e.g., axon diameter, g-ratio) and non-traditional geometric descriptors such as solidity, eccentricity, and orientation, allowing multidimensional analysis of axon-myelin structure. However, as AxonDeepSeg defines continuous inner and outer myelin borders [ 35 ], it may smooth over regions of myelin decompaction or irregular lamellae, potentially underestimating severe pathology. Future training sets incorporating more pathological examples could further improve sensitivity to these features. Unsupervised clustering identifies distinct axon-myelin phenotypes The multidimensional dataset generated by AxonDeepSeg enabled unsupervised clustering of axons into six morphologically distinct groups based on combinations of traditional and geometric morphometrics. Most fibres were grouped in Clusters 1 or 2, which represented compactly myelinated and partially decompacted profiles, while Clusters 3 to 6 captured less frequent subtypes with thickened, disrupted, or absent myelin. This data-driven classification differed from earlier approaches that categorised axons solely by the proportion of compact versus decompacted myelin around their circumference [ 12 , 13 ] and instead used an integrated set of shape and size metrics to define axonal phenotypes. Comparisons between sham and injured nerves within each cluster revealed that injury effects were largely confined to Clusters 1 and 2, where significant increases in g-ratio, solidity, and axon/fibre area ratio were observed. These changes occurred in fibres that were mostly morphologically “normal,” implying that subtle ultrastructural remodelling may precede visible demyelination. Interestingly, Cluster 1, which visually appeared the most intact, showed the greatest number of altered parameters, suggesting that early axonal stress may manifest primarily through changes in axon geometry rather than overt myelin disruption. In contrast, clusters representing already abnormal or unmyelinated fibres showed minimal change, consistent with them representing stable or end-stage phenotypes at this time point. This pattern supports the idea that secondary degeneration affects distinct axonal subpopulations differently, in agreement with studies showing variable metabolic and stress resilience among oligodendrocyte–axon units [ 56 – 58 ]. It is possible that the relatively variable morphology of axons in Clusters 3–5 may have precluded detection of changes. Interpreting non-traditional morphometrics Among the non-traditional morphometrics, solidity emerged as particularly informative. It quantifies how closely an axon’s cross-section approximates its convex hull and may therefore reflect cytoskeletal tension or compaction of the axoplasmic contents. Increases in solidity, together with elevated g-ratio and axon-to-fibre area ratio, suggest coordinated microstructural adaptations in both the axon and its myelin sheath. Similar contour changes have been reported following traumatic brain injury and oxidative stress, where cytoskeletal disruption - including neurofilament reorganisation - precedes overt axon swelling and alters axonal shape rather than simply increasing size [ 42 , 59 ]. Together, these findings demonstrate that AxonDeepSeg can detect early morphological adaptations of axons and myelin that may precede overt degeneration or demyelination. Methodological and translational implications This workflow bridges a long-standing methodological gap in ultrastructural neurobiology by providing an objective, high throughput means of quantifying axon and myelin morphology. By eliminating manual bias and including all fibres within an image, irrespective of pathology, it allows comprehensive and reproducible characterisation of white-matter microstructure. The mean morphometric values for each cluster (Table 2 ) can serve as practical reference points for cross-dataset comparison, enabling other researchers to identify similar phenotypic classes using conventional measurements such as g-ratio and axon diameter. The approach can be readily extended to other models of neurotrauma and demyelination, including chronic traumatic encephalopathy and multiple sclerosis, where subtle axon-myelin mismatches are functionally important. Combining automated morphometry with molecular or electrophysiological data could clarify how specific morphometric signatures correspond to impaired conduction or altered axon-glia metabolism. Limitations and future directions Despite its strengths, automated myelin segmentation is not without challenges. The boundary-based algorithm of AxonDeepSeg may underestimate severe lamellar pathology, and segmentation accuracy depends on the diversity of images used for training. Incorporating a wider range of pathological examples and adopting hybrid three-dimensional models could further enhance fidelity. Longitudinal studies are also warranted to determine whether the early morphometric changes identified here predict later demyelination or functional impairment. Conclusion In summary, AxonDeepSeg provides a reliable and scalable method for quantifying axon and myelin ultrastructure. By detecting increases in g-ratio, solidity, and axon-to-fibre area ratio following partial optic nerve injury, AxonDeepSeg revealed early, subtle deviations in axon–myelin geometry that elude manual measurement. By integrating deep learning with unsupervised clustering, this approach enhances the precision and interpretive value of ultrastructural analysis, offering new opportunities to identify early morphological biomarkers of axonal stress and secondary degeneration. Abbreviations TEM transmission electron microscopy CNNs convolutional neural networks MIB Microscopy Image Browser OM optical microscopy LM light microscopy EM electron microscopy Declarations Competing Interests The authors declare that they have no competing interests. Funding The funding for this research project was provided by the Neurotrauma Research Program WA (NRP) and was funded by the State Government of Western Australia through the Department of Health, as well as the National Health & Medical Research Fund (APP1087114). The funders had no role in the study design, data collection, and analysis, decision to publish, or preparation of the manuscript. Author Contribution PP conceptualised the project and study design, conducted AxonDeepSeg model training,data analysis, data curation and validation, manual image analysis, figure preparation, data interpretation, and was a major contributor in writing the manuscript. BRL performed manual image analysis, data curation and validation, figure preparation, data interpretation, and was a major contributor in writing the manuscript. SCH performed manual image analysis, data curation and validation, data interpretation, and contributed to writing the manuscript. TM conducted animal experiments, tissue processing, image acquisition, manual image analysis, and reviewed the manuscript. AJW participated in project conceptualisation, conducted animal experiments, tissue processing, image acquisition, and reviewed the manuscript. NL conducted AxonDeepSeg model training and data analysis and reviewed the manuscript. ET conducted AxonDeepSeg model training and data analysis and reviewed the manuscript. CAB conducted animal experiments, tissue processing, and reviewed the manuscript. MF supervised the project, secured project funding, contributed to data interpretation, and edited and reviewed the manuscript. CCA conceptualised the project and study design, conducted project administration, supervision, animal experiments, tissue processing, image acquisition, AxonDeepSeg model training and data analysis, manual image analysis, data curation and validation, figure preparation, data interpretation, and was a major contributor in writing the manuscript. Acknowledgements Not applicable. Data Availability The datasets supporting the conclusions of this article are available at Zenodo [DOI: 10.5281/zenodo.18686326; https://zenodo.org/records/18686326?preview=1&token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6IjAxZmQ5NjU1LWZlM2ItNDE0MC04ZmIzLWQ1YTJlODRiODJjMyIsImRhdGEiOnt9LCJyYW5kb20iOiIzZGFkYjdmMGM0YzY2OTBmMTk0MWFmN2RkZjgwZWFiZSJ9.cVDQbJmtpryYW2x0Nwaphvt0VKaPtWYC7k8TVZgDSe7TXQoQesVIte9PbKfA73o0tXccmlRLG7lS-42GUV2CwQ].All custom code used for data processing, AxonDeepSeg model training, and clustering analyses is available at a GitHub repository, [https://github.com/CurtinNeurotrauma/ads-morphometrics-clustering]. References Stadelmann C, Timmler S, Barrantes-Freer A, Simons M (2019) Myelin in the central nervous system: Structure, function, and pathology. Physiol Rev 99(3):1381–1431. 10.1152/physrev.00031.2018 de Faria O, Pivonkova H, Varga B, Timmler S, Evans KA, Káradóttir RT (2021) Periods of synchronized myelin changes shape brain function and plasticity. Nat Neurosci 24(11):1508–1521. 10.1038/s41593-021-00917-2 McKenzie IA, Ohayon D, Li H, Paes de Faria J, Emery B, Tohyama Ket al et al (2014) Motor skill learning requires active central myelination. Science 346(6207):318–322. 10.1126/science.1254960 Nave K-A (2010) Myelination and the trophic support of long axons. Nat Rev Neurosci 11(4):275–283. 10.1038/nrn2797 Fitzgerald M, Payne SC, Bartlett CA, Evill L, Harvey AR, Dunlop SA (2009) Secondary retinal ganglion cell death and the neuroprotective effects of the calcium channel blocker lomerizine. Investig Ophthalmol Vis Sci 50(11):5456–5462. 10.1167/iovs.09-3717 Levkovitch-Verbin H, Quigley HA, Martin KRG, Zack DJ, Pease ME, Valenta DF (2003) A model to study differences between primary and secondary degeneration of retinal ganglion cells in rats by partial optic nerve transection. Investig Ophthalmol Vis Sci 44(8):3388–3393. 10.1167/iovs.02-0646 Raineteau O, Schwab ME (2001) Plasticity of motor systems after incomplete spinal cord injury. Nat Rev Neurosci 2(4):263–273. 10.1038/35067570 Bernardo A, Bianchi D, Magnaghi V, Minghetti L (2009) Peroxisome proliferator-activated receptor-γ agonists promote differentiation and antioxidant defenses of oligodendrocyte progenitor cells. J Neuropathology Experimental Neurol 68(7):797–808. 10.1097/NEN.0b013e3181aba2c1 Thorburne SK, Juurlink BHJ (1996) Low glutathione and high iron govern the susceptibility of oligodendroglial precursors to oxidative stress. J Neurochem 67(3):1014–1022. 10.1046/j.1471-4159.1996.67031014.x Koutsoudaki PN, Hildebrandt H, Gudi V, Skripuletz T, Škuljec J, Stangel M (2010) Remyelination after cuprizone induced demyelination is accelerated in mice deficient in the polysialic acid synthesizing enzyme st8siaiv. Neuroscience 171(1):235–244. 10.1016/j.neuroscience.2010.08.070 Sachs HH, Bercury KK, Popescu DC, Narayanan SP, Macklin WB (2014) A new model of cuprizone-mediated demyelination/remyelination. ASN Neuro 6(5):1759091414551955. 10.1177/1759091414551955 Payne SC, Bartlett CA, Harvey AR, Dunlop SA, Fitzgerald M (2011) Chronic swelling and abnormal myelination during secondary degeneration after partial injury to a central nervous system tract. J Neurotrauma 28(6):1077–1088. 10.1089/neu.2010.1665 Payne SC, Bartlett CA, Harvey AR, Dunlop SA, Fitzgerald M (2012) Myelin sheath decompaction, axon swelling, and functional loss during chronic secondary degeneration in rat optic nerve. Investig Ophthalmol Vis Sci 53(10):6093–6101. 10.1167/iovs.12-10080 Liewald D, Miller R, Logothetis N, Wagner H-J, Schüz A (2014) Distribution of axon diameters in cortical white matter: An electron-microscopic study on three human brains and a macaque. Biol Cybern 108(5):541–557. 10.1007/s00422-014-0626-2 Schindelin J, Arganda-Carreras I, Frise E, Kaynig V, Longair M, Pietzsch T al (2012) Fiji: An open-source platform for biological-image analysis. Nat Methods 9(7):676–682. 10.1038/nmeth.2019 Anthes DL, Theriault E, Tator CH (1995) Characterization of axonal ultrastructural pathology following experimental spinal cord compression injury. Brain Res 702(1):1–16. https://doi.org/10.1016/0006-8993(95)01028-6 Armstrong RC, Mierzwa AJ, Marion CM, Sullivan GM (2016) White matter involvement after tbi: Clues to axon and myelin repair capacity. Exp Neurol 275:328–333. https://doi.org/10.1016/j.expneurol.2015.02.011 Chen HS-M, Holmes N, Liu J, Tetzlaff W, Kozlowski P (2017) Validating myelin water imaging with transmission electron microscopy in a rat spinal cord injury model. NeuroImage 153:122–130. https://doi.org/10.1016/j.neuroimage.2017.03.065 Donovan V, Kim C, Anugerah AK, Coats JS, Oyoyo U, Pardo ACet al et al (2014) Repeated mild traumatic brain injury results in long-term white-matter disruption. J Cereb Blood Flow Metabolism 34(4):715–723. 10.1038/jcbfm.2014.6 Fehily B, Bartlett CA, Lydiard S, Archer M, Milbourn H, Majimbi M al (2019) Differential responses to increasing numbers of mild traumatic brain injury in a rodent closed-head injury model. J Neurochem 149(5):660–678. 10.1111/jnc.14673 Nashmi R, Fehlings MG (2001) Changes in axonal physiology and morphology after chronic compressive injury of the rat thoracic spinal cord. Neuroscience 104(1):235–251. https://doi.org/10.1016/S0306-4522(01)00009-4 Reeves TM, Smith TL, Williamson JC, Phillips LL (2012) Unmyelinated axons show selective rostrocaudal pathology in the corpus callosum after traumatic brain injury. J Neuropathology Experimental Neurol 71(3):198–210. 10.1097/NEN.0b013e3182482590 Sullivan GM, Mierzwa AJ, Kijpaisalratana N, Tang H, Wang Y, Song S-K al (2013) Oligodendrocyte lineage and subventricular zone response to traumatic axonal injury in the corpus callosum. J Neuropathology Experimental Neurol 72(12):1106–1125. 10.1097/nen.0000000000000009 Pham DL, Xu C, Prince JL (2000) Current methods in medical image segmentation. Annu Rev Biomed Eng 2(1):315–337. 10.1146/annurev.bioeng.2.1.315 Romero E, Cuisenaire O, Denef JF, Delbeke J, Macq B, Veraart C (2000) Automatic morphometry of nerve histological sections. J Neurosci Methods 97(2):111–122. 10.1016/S0165-0270(00)00167-9 Liu T, Jurrus E, Seyedhosseini M, Ellisman M, Tasdizen T (2012) Watershed merge tree classification for electron microscopy image segmentation. Proc IAPR Int Conf Pattern Recogn 2012 133–137 Wang Y-Y, Sun Y-N, Lin C-CK, Ju M-S (2012) Segmentation of nerve fibers using multi-level gradient watershed and fuzzy systems. Artif Intell Med 54(3):189–200. 10.1016/j.artmed.2011.11.008 More HL, Chen J, Gibson E, Donelan JM, Beg MF (2011) A semi-automated method for identifying and measuring myelinated nerve fibers in scanning electron microscope images. J Neurosci Methods 201(1):149–158. 10.1016/j.jneumeth.2011.07.026 Zhao X, Pan Z, Wu J, Zhou G, Zeng Y (2010) Automatic identification and morphometry of optic nerve fibers in electron microscopy images. Comput Med Imaging Graph 34(3):179–184. 10.1016/j.compmedimag.2009.08.009 Belevich I, Jokitalo E (2021) Deepmib: User-friendly and open-source software for training of deep learning network for biological image segmentation. PLoS Comput Biol 17(3):e1008374. 10.1371/journal.pcbi.1008374 Zaimi A, Duval T, Gasecka A, Côté D, Stikov N, Cohen-Adad J (2016) Axonseg: Open source software for axon and myelin segmentation and morphometric analysis. Front Neuroinform 10:37. 10.3389/fninf.2016.00037 Behanova A, Abdollahzadeh A, Belevich I, Jokitalo E, Sierra A, Tohka J (2022) Gacson software for automated segmentation and morphology analyses of myelinated axons in 3d electron microscopy. Comput Methods Programs Biomed. https://doi.org/10.1016/j.cmpb.2022.106802 . 220 106802 doi: Plebani E, Biscola NP, Havton LA, Rajwa B, Shemonti AS, Jaffey D al (2022) High-throughput segmentation of unmyelinated axons by deep learning. Sci Rep 12(1):1198. 10.1038/s41598-022-04854-3 Sommer C, Straehle C, Köthe U, Hamprecht FA, Ilastik Interactive learning and segmentation toolkit In 2011 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2011/03// 2011 (pp. 230–233). 10.1109/ISBI.2011.5872394 Zaimi A, Wabartha M, Herman V, Antonsanti P-L, Perone CS, Cohen-Adad J (2018) Axondeepseg: Automatic axon and myelin segmentation from microscopy data using convolutional neural networks. Sci Rep 8(1):3816. 10.1038/s41598-018-22181-4 Fitzgerald M, Bartlett CA, Evill L, Rodger J, Harvey AR, Dunlop SA (2009) Secondary degeneration of the optic nerve following partial transection: The benefits of lomerizine. Exp Neurol 216(1):219–230. 10.1016/j.expneurol.2008.11.026 Lins BR, Anyaegbu CC, McGonigle T, Hellewell SC, Patel P, Reagan H al (2023) Secondary degeneration impairs myelin ultrastructural development in adulthood following adolescent neurotrauma in the rat optic nerve. Int J Mol Sci 24(4):3343. 10.3390/ijms24043343 Bartlett CA, Fitzgerald M (2018) Partial transection of adult rat optic nerve as a model of secondary degeneration in the central nervous system. Bio Protoc 8(24):e3118. 10.21769/bioprotoc.3118 Hastie T, Tibshirani R, Friedman J (2009) The elements of statistical learning: Data mining, inference, and prediction. Springer-, New York Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B Grisel O (2011) Scikit-learn: Machine learning in python. J Mach Learn Res 12(85) 2825–2830 von Luxburg U (2007) A tutorial on spectral clustering. Stat Comput 17(4):395–416. 10.1007/s11222-007-9033-z Pozo Devoto VM, Lacovich V, Feole M, Bhat P, Chovan J, Čarna M al (2022) Unraveling axonal mechanisms of traumatic brain injury. Acta Neuropathol Commun 10(1):140. 10.1186/s40478-022-01414-8 Wang J, Hamm RJ, Povlishock JT (2011) Traumatic axonal injury in the optic nerve: Evidence for axonal swelling, disconnection, dieback, and reorganization. J Neurotrauma 28(7):1185–1198. 10.1089/neu.2011.1756 Leterrier C (2021) Putting the axonal periodic scaffold in order. Curr Opin Neurobiol 69:33–40. 10.1016/j.conb.2020.12.015 Giacci MK, Bartlett CA, Smith NM, Iyer KS, Toomey LM, Jiang H al (2018) Oligodendroglia are particularly vulnerable to oxidative damage after neurotrauma in vivo. J Neurosci 38(29):6491–6504. 10.1523/JNEUROSCI.1898-17.2018 Cummins N, Bartlett CA, Archer M, Bartlett E, Hemmi JM, Harvey AR al (2013) Changes to mitochondrial ultrastructure in optic nerve vulnerable to secondary degeneration in vivo are limited by irradiation at 670 nm. BMC Neurosci 14(1):98. 10.1186/1471-2202-14-98 Saab AS, Nave K-A (2017) Myelin dynamics: Protecting and shaping neuronal functions. Curr Opin Neurobiol 47:104–112. 10.1016/j.conb.2017.09.013 Arancibia-Cárcamo IL, Ford MC, Cossell L, Ishida K, Tohyama K, Attwell D (2017) Node of ranvier length as a potential regulator of myelinated axon conduction speed. eLife 6:e23329. 10.7554/eLife.23329 Dutta DJ, Woo DH, Lee PR, Pajevic S, Bukalo O, Huffman WC al (2018) Regulation of myelin structure and conduction velocity by perinodal astrocytes. Proc Natl Acad Sci U S A 115(46):11832–11837. 10.1073/pnas.1811013115 Moiseev D, Hu B, Li J (2019) Morphometric analysis of peripheral myelinated nerve fibers through deep learning. J Peripher Nerv Syst 24(1):87–93. 10.1111/jns.12293 Abdollahzadeh A, Belevich I, Jokitalo E, Sierra A, Tohka J (2021) Deepacson automated segmentation of white matter in 3d electron microscopy. Commun Biol 4(1):179. 10.1038/s42003-021-01699-w Wei D, Lee K, Li H, Lu R, Bae JA Liu Z Axonem dataset: 3d axon instance segmentation of brain cortical regions In de Bruijne M, Cattin PC, Cotin S, Padoy N, Speidel S, Y Zheng, (eds), 2021 2021 (pp. 175–185, Lecture Notes in Computer Science). Cham: Springer International Publishing. 10.1007/978-3-030-87193-2_17 Dweiri Y, Al-Zanina M, Durand D (2022) Enhanced automatic morphometry of nerve histological sections using ensemble learning. Electronics 11(14):2277. 10.3390/electronics11142277 Mesbah R, McCane B, Mills S Deep convolutional encoder-decoder for myelin and axon segmentation In (2016) International Conference on Image and Vision Computing New Zealand (IVCNZ), 2016/11// 2016 (pp. 1–6). 10.1109/IVCNZ.2016.7804455 Deng W, Hedberg-Buenz A, Soukup DA, Taghizadeh S, Wang K, Anderson MG al (2021) Axondeep: Automated optic nerve axon segmentation in mice with deep learning. Translational Vis Sci Technol 10(14):22. 10.1167/tvst.10.14.22 Giacci MK, Bartlett CA, Huynh M, Kilburn MR, Dunlop SA, Fitzgerald M (2018) Three dimensional electron microscopy reveals changing axonal and myelin morphology along normal and partially injured optic nerves. Sci Rep 8(1):3979. 10.1038/s41598-018-22361-2 Knöferle J, Koch JC, Ostendorf T, Michel U, Planchamp V, Vutova P (2010) al Mechanisms of acute axonal degeneration in the optic nerve in vivo. Proceedings of the National Academy of Sciences 107(13) 6064–6069 doi: 10.1073/pnas.0909794107 Lee Y, Morrison BM, Li Y, Lengacher S, Farah MH, Hoffman PN et al (2012) Oligodendroglia metabolically support axons and contribute to neurodegeneration. Nature 487(7408) 443–448 10.1038/nature11314 Siedler DG, Chuah MI, Kirkcaldie MTK, Vickers JC, King AE (2014) Diffuse axonal injury in brain trauma: Insights from alterations in neurofilaments. Front Cell Neurosci 8. 10.3389/fncel.2014.00429 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8951825","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":596944731,"identity":"a3121655-75ee-4c36-9389-6e0077e98deb","order_by":0,"name":"Parth Patel","email":"","orcid":"","institution":"Curtin University","correspondingAuthor":false,"prefix":"","firstName":"Parth","middleName":"","lastName":"Patel","suffix":""},{"id":596944732,"identity":"c000f8ab-1ac2-41aa-8ece-f49d7817aa92","order_by":1,"name":"Brittney R. Lins","email":"","orcid":"","institution":"Curtin University","correspondingAuthor":false,"prefix":"","firstName":"Brittney","middleName":"R.","lastName":"Lins","suffix":""},{"id":596944734,"identity":"65081bc3-7a75-4e87-86bf-21921fe0959e","order_by":2,"name":"Sarah C. Hellewell","email":"","orcid":"","institution":"Curtin University","correspondingAuthor":false,"prefix":"","firstName":"Sarah","middleName":"C.","lastName":"Hellewell","suffix":""},{"id":596944737,"identity":"0b1f5608-0381-4024-bac2-67886c29a680","order_by":3,"name":"Terence McGonigle","email":"","orcid":"","institution":"Curtin University","correspondingAuthor":false,"prefix":"","firstName":"Terence","middleName":"","lastName":"McGonigle","suffix":""},{"id":596944738,"identity":"7363c679-bdfd-43c2-b527-69d2d8021663","order_by":4,"name":"Andrew Warnock","email":"","orcid":"","institution":"Curtin University","correspondingAuthor":false,"prefix":"","firstName":"Andrew","middleName":"","lastName":"Warnock","suffix":""},{"id":596944739,"identity":"089a3476-0b50-4066-b8bb-8784398257ce","order_by":5,"name":"Naing Lin","email":"","orcid":"","institution":"Curtin University","correspondingAuthor":false,"prefix":"","firstName":"Naing","middleName":"","lastName":"Lin","suffix":""},{"id":596944740,"identity":"4e495816-9a4d-4ddc-a670-32770eb14819","order_by":6,"name":"Enoch Teo","email":"","orcid":"","institution":"Curtin University","correspondingAuthor":false,"prefix":"","firstName":"Enoch","middleName":"","lastName":"Teo","suffix":""},{"id":596944742,"identity":"88f842ef-d9ec-4ac2-b6c2-475b663e96b9","order_by":7,"name":"Carole Bartlett","email":"","orcid":"","institution":"Curtin University","correspondingAuthor":false,"prefix":"","firstName":"Carole","middleName":"","lastName":"Bartlett","suffix":""},{"id":596944746,"identity":"7f5baf3e-2583-4d9e-bab1-40d257542cf8","order_by":8,"name":"Melinda Fitzgerald","email":"","orcid":"","institution":"Curtin University","correspondingAuthor":false,"prefix":"","firstName":"Melinda","middleName":"","lastName":"Fitzgerald","suffix":""},{"id":596944747,"identity":"67451c76-7610-48f9-8c90-335edc1573f2","order_by":9,"name":"Chidozie C. Anyaegbu","email":"data:image/png;base64,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","orcid":"","institution":"Curtin University","correspondingAuthor":true,"prefix":"","firstName":"Chidozie","middleName":"C.","lastName":"Anyaegbu","suffix":""}],"badges":[],"createdAt":"2026-02-24 02:38:41","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8951825/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8951825/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107870752,"identity":"551cee36-9dd3-42c4-9ab2-124be27693b5","added_by":"auto","created_at":"2026-04-27 07:40:33","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":323646,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAutomated segmentation of optic nerve ultrastructure using AxonDeepSeg.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSchematic overview of the optic nerve partial transection model in Piebald Virol-Glaxo rat highlighting the surgical site behind the right eye. A pop-out TEM image of an injured optic nerve cross section showcases the characteristic ‘mushroom’ morphology at the primary injury site on dorsal aspect. A red frame in the ventral region represents the site of image collection and a second pop-out window shows an example of the high magnification TEM micrographs of axon and myelin cross sections. Myelin appears electron dense (dark) while axons appear lighter \u003cstrong\u003e(a)\u003c/strong\u003e. The workflow for training AxonDeepSeg illustrated by a ground truth label generated by manual segmentation of the original TEM micrograph and the corresponding automated segmentation generated by AxonDeepSeg.. The ground truth labels contain 3 colour values: white (axon), grey (myelin), and black (background). The trained segmentation images contain 3 colour values: blue (axon), red (myelin), and black (background; \u003cstrong\u003eb\u003c/strong\u003e).\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8951825/v1/2dfd9c1365e433a0dee1d349.jpg"},{"id":107865084,"identity":"431e51b8-476f-4844-acac-913e0d263a26","added_by":"auto","created_at":"2026-04-27 06:30:00","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":327882,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eValidation of AxonDeepSeg model performance. \u003c/strong\u003eQuantitative comparison of automated segmentation metrics in the test set images, and comparison between automated and manual segmentation method. \u003cstrong\u003e(a)\u003c/strong\u003e Bar graphs display the mean pixel-wise accuracy achieved for axons and myelin. \u003cstrong\u003e(b)\u003c/strong\u003eBar graphs display dice similarity coefficients achieved for axons and myelin. \u003cstrong\u003e(c)\u003c/strong\u003e Bar graphs display the false-positive rates for segmentation of the myelin and axon compartments. \u003cstrong\u003e(d)\u003c/strong\u003e Bar graphs display the false-negative rates for segmentation of the myelin and axon compartments. \u003cstrong\u003e(e)\u003c/strong\u003e A scatterplot displays both axon diameter measurements derived from AxonDeepSeg and manual measurements with no significant difference which implies quantitative agreement. Data are mean ± SEM; ** indicates p \u0026lt; 0.01; **** indicates p \u0026lt; 0.0001; ns indicates p \u0026gt; 0.05.\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8951825/v1/b4fedfffa77a4ce0921021b0.jpg"},{"id":107865082,"identity":"47061284-68ce-4a50-bfda-c8f172496888","added_by":"auto","created_at":"2026-04-27 06:30:00","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":295928,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparisons of optic nerve axons and myelin morphometrics 2 weeks after partial transection injury or sham surgery.\u003c/strong\u003e Each panel contains a representative image of a myelinated axon cross section, annotated to illustrate each morphometric parameter, and a scatterplot quantifying each parameter in sham and injured optic nerves. Injury resulted in increased g-ratio (\u003cstrong\u003ea\u003c/strong\u003e), greater solidity (\u003cstrong\u003eg\u003c/strong\u003e), and higher axon-to-fibre area ratio (\u003cstrong\u003ej\u003c/strong\u003e), while axon area (\u003cstrong\u003ed\u003c/strong\u003e), fibre area (\u003cstrong\u003ei\u003c/strong\u003e), and perimeter remained unchanged (\u003cstrong\u003ec\u003c/strong\u003e). No group differences were observed in myelin area (\u003cstrong\u003ee\u003c/strong\u003e), eccentricity (\u003cstrong\u003ef\u003c/strong\u003e), or orientation (\u003cstrong\u003eh\u003c/strong\u003e). Data are mean ± SEM (n = 5/group). \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05, Student’s t-test. * indicates p \u0026lt; 0.05; ns indicates p \u0026gt; 0.05.\u003c/p\u003e","description":"","filename":"Picture3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8951825/v1/8db23eeacba7d392b358f5c4.jpg"},{"id":107865087,"identity":"cae742ff-ab0b-4e4a-8613-c6779d50d615","added_by":"auto","created_at":"2026-04-27 06:30:00","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":316050,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eUnsupervised clustering of AxonDeepSeg-derived morphometrics reveals distinct axon–myelin phenotypes. \u003c/strong\u003ePrincipal component analysis and spectral clustering of morphometric features identified six phenotypically distinct clusters of axon–myelin units \u003cstrong\u003e(a)\u003c/strong\u003e. Representative TEM images from each cluster illustrate characteristic ultrastructural phenotypes: compact myelinated (Cluster 1), partially decompacted (Cluster 2), thickened or disrupted (Clusters 3 and 4), unmyelinated (Cluster 5), and large diameter, partially decompacted (Cluster 6) fibres. Scale bar = 1 µm\u003cstrong\u003e (b)\u003c/strong\u003e.\u003c/p\u003e","description":"","filename":"Picture4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8951825/v1/bb31d11f7eadd5540d8c73dd.jpg"},{"id":107865085,"identity":"836eb594-3165-4f5a-9ec7-2703241406d3","added_by":"auto","created_at":"2026-04-27 06:30:00","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":130950,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEffects of injury on morphometric parameters of Cluster 1. \u003c/strong\u003eComparison of sham and injury groups within Cluster 1 (n = 1080) axons. In Cluster 1, partial transection increased g-ratio, solidity, axon-to-fibre area ratio, axon diameter, and axon area (\u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05 for all). No significant effects were observed for fibre area, perimeter, or orientation. Data are mean ± SEM (n = 5/group). * indicates p \u0026lt; 0.05; *** indicates p \u0026lt; 0.001; ns indicates p \u0026gt; 0.05, unpaired Student’s t-test.\u003c/p\u003e","description":"","filename":"Picture5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8951825/v1/95c80146d7acc12e08c5ef8f.jpg"},{"id":107865086,"identity":"a6d8f7ff-885a-4d7c-97ed-1fc16577ede1","added_by":"auto","created_at":"2026-04-27 06:30:00","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":135762,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEffects of injury on morphometric parameters of Cluster 2. \u003c/strong\u003eScatterplots comparing sham and injury groups within Cluster 2 (n = 92) axons. In Cluster 2, injury elevated g-ratio (a), solidity (b), and axon-to-fibre area ratio (c), without changes in axon size-related metrics. No significant effects were observed for axon diameter (d), axon area (e), axon perimeter (f), myelin area (g), fibre area (h), eccentricity (i), or orientation (j). Data are mean ± SEM (n = 5/group). * indicates p\u0026lt;0.05; ns indicates p\u0026gt;0.05, unpaired Student’s t-test.\u003c/p\u003e","description":"","filename":"Picture6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8951825/v1/a84f709688964b2b0b366dae.jpg"},{"id":107872246,"identity":"9d8d1a19-067d-4ca8-9481-dc1b0258c376","added_by":"auto","created_at":"2026-04-27 07:56:25","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1964526,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8951825/v1/bc61b665-b547-4f77-95f8-2c2535d9a2e8.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Unsupervised clustering of AxonDeepSeg morphometrics uncovers early patterns of axon-myelin secondary degeneration following neurotrauma","fulltext":[{"header":"Introduction","content":"\u003cp\u003eMyelin, the lipid-rich sheath encasing axons, is essential for rapid and efficient transmission of action potentials within the nervous system [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Damage to myelin - whether caused by neurotrauma or neurodegenerative disease - contributes to the cognitive, sensory, and motor deficits characteristic of these conditions [\u003cspan additionalcitationids=\"CR3\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Following neurotrauma, a cascade of metabolic and cellular disturbances renders uninjured axons adjacent to the primary lesion vulnerable to progressive secondary degeneration [\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Myelinating oligodendrocytes are particularly susceptible to secondary degeneration due to their high metabolic demand and low antioxidant capacity [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], making them prone to oxidative stress and lipid peroxidation. The resulting dysmyelination and myelin decompaction can exacerbate axonal dysfunction and impair repair processes, underscoring the need for reliable methods to quantify myelin integrity and assess therapeutic strategies that preserve myelin structure [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTransmission electron microscopy (TEM) remains the gold standard for quantifying axonal and myelin morphology at ultrastructural resolution. Conventional morphometric analyses assess parameters such as axon diameter, myelin thickness, and g-ratio, with injury typically inducing reductions in myelin compaction and increased variability in axon calibre. For example, the extent of myelin decompaction can be quantified by measuring the percentage of the axonal circumference surrounded by decompacted laminae [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. However, these analyses are typically performed manually using image-processing software such as ImageJ [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], which are labour-intensive, subjective, and limited in throughput. Hundreds of axons across multiple regions must be measured to achieve statistical power, and such manual approaches are prone to user bias and inconsistent criteria across studies. This has contributed to inter-study variability and limits reproducibility when comparing injury models or therapeutic interventions [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan additionalcitationids=\"CR17 CR18 CR19 CR20 CR21 CR22\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. An automated approach capable of segmenting, measuring, and classifying axonal and myelin morphology, including irregularly compacted sheaths, would therefore provide a more objective and scalable method for ultrastructural analysis.\u003c/p\u003e \u003cp\u003eAutomated morphometric analysis relies on accurate image segmentation; the partitioning of microscopic images into regions corresponding to biological structures such as axons and myelin [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Traditional semi-automated segmentation techniques such as thresholding [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], watershed [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. or region growing [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] require manual pre- or post-processing, introducing user-dependent subjectivity. These approaches are also often restricted to specific imaging modalities or contrast conditions, limiting their broader applicability. In contrast, convolutional neural networks (CNNs) overcome many of these limitations by learning contextual features directly from training data, enabling fully automated segmentation without user intervention. Once trained, CNN-based models can segment images in seconds, offering major improvements in reproducibility, throughput, and efficiency.\u003c/p\u003e \u003cp\u003eA range of CNN-based segmentation frameworks have been developed for electron microscopy data, each varying in usability, accessibility, and analytical depth. While tools such as Microscopy Image Browser [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], AxonSeg [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], and others [\u003cspan additionalcitationids=\"CR33\" citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] have achieved high-quality segmentation, most are limited by one or more of the following factors: lack of an intuitive graphical interface, incomplete open-source availability, or morphometric outputs which are restricted to basic geometric features. Many of these models also require advanced coding expertise or provide only two-dimensional functionality, limiting adoption by experimental neuroscientists. AxonDeepSeg [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e] overcomes these constraints by combining full open-source accessibility, compatibility with multiple imaging modalities (electron-, optical-, and light-microscopy), and a streamlined command-line or graphical user interface-based workflow that requires minimal computational experience. Critically, AxonDeepSeg provides an expanded suite of morphometric parameters, from classical metrics (axon diameter, g-ratio) to higher-order geometric descriptors such as solidity, eccentricity, and orientation, enabling multidimensional quantification of axon and myelin ultrastructure. These features make AxonDeepSeg particularly suited for objective, high-throughput, and reproducible quantification of axonal pathology.\u003c/p\u003e \u003cp\u003eIn this study, we trained a custom AxonDeepSeg model using TEM images of the rat optic nerve following partial transection injury - a well-established model of secondary degeneration [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Beyond validating the performance of AxonDeepSeg for automated segmentation, we leveraged its diverse morphometric output to quantify both conventional features and additional geometric parameters such as solidity, eccentricity, and orientation. We further introduced the axon/fibre area ratio as a novel metric to capture subtle changes in myelin decompaction. Finally, by applying unsupervised clustering to these multidimensional morphometric data, we identified previously unrecognised axon\u0026ndash;myelin phenotypes and revealed distinct patterns of structural change following injury. This framework fills a key methodological gap by enabling objective, high-throughput, and reproducible analysis of axonal ultrastructure, providing a scalable basis for cross-study comparison of injury- and disease-related degeneration.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eAnimals\u003c/h2\u003e \u003cp\u003eFemale Piebald Virol-Glaxo rats (n\u0026thinsp;=\u0026thinsp;10; 3 months old; 160\u0026ndash;200 g; Animal Resource Centre, Murdoch, Western Australia, Australia) were randomly allocated to receive either optic nerve partial transection or sham control procedure (n\u0026thinsp;=\u0026thinsp;5/group), an adequate sample size for electron microscopy analyses [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. The experimental unit was individual rats. Randomisation was achieved at a cage level by allocating animals to groups in the order in which they were received from the supplier. Additional confounders were not controlled. Rats were housed two to three per standard polypropylene cage under specific-pathogen-free conditions, with food and water available \u003cem\u003ead libitum\u003c/em\u003e. Following arrival at the facility, rats were left to acclimatise for a minimum of one week prior to initiating procedures. All procedures were conducted in accordance with the \u003cem\u003eAustralian Code for the Care and Use of Animals for Scientific Purposes\u003c/em\u003e (National Health and Medical Research Council, 2013) and approved by both The University of Western Australia Animal Ethics Committee (RA/3/100/1485) and Curtin University Animal Ethics Committee (ARE2017-4).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003ePartial optic nerve transection\u003c/h3\u003e\n\u003cp\u003ePartial transection of the right optic nerve was performed as previously described [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e] (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea). Rats were anaesthetised via intraperitoneal injection of xylazine (10 mg/kg) and ketamine (50 mg/kg). An incision was made above the right eye to expose the optic nerve, and a 200 \u0026micro;m-deep dorsal incision was made 1 mm posterior to the globe using a diamond radial keratotomy knife. The incision above the eyelid was sutured post-operatively, and animals received analgesic (Norocarp, 2.8 mg/kg) and antibiotic (Neomycin, 10 mg/kg) treatment. Sham-operated rats underwent identical surgical procedures without the optic nerve incision. Each rat received one procedure, one time. Following recovery from anaesthesia rats were returned to their home cage. Optic nerve partial transection was used because it produces distinct and reproducible primary and secondary injury sites, ideal for investigating secondary degeneration [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. No animals were excluded from the study due to pre-defined inclusion or exclusion criteria (i.e. intolerability to anaesthesia, premature death, welfare concerns, surgical errors or complications, or anatomical abnormalities).\u003c/p\u003e\n\u003ch3\u003eTransmission electron microscopy\u003c/h3\u003e\n\u003cp\u003eTissue processing and imaging were conducted as previously described [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. Two weeks after injury or sham surgery, rats were euthanised with pentobarbitone sodium (850 mg/kg) and phenytoin sodium (125 mg/kg; intraperitoneal) and perfused transcardially with 0.9% saline followed by 2% paraformaldehyde/2.5% glutaraldehyde/2% sucrose in 0.1 M phosphate buffer (pH 7.2). Optic nerves were dissected, separated from the dural sheath, stored in 0.13 M Sorenson\u0026rsquo;s phosphate buffer (pH 7.2) and post-fixed in 1% osmium tetroxide (Electron Microscopy Sciences, ProSciTech, Townsville, QLD, Australia; Cat# C011). Samples were dehydrated through an ethanol gradient, transitioned to propylene oxide, and embedded in epoxy resin (Araldite Procure, ProSciTech; Cat# 039). Ultrathin cross-sections (100 nm) were prepared using an LKB Nova ultramicrotome (Bromma, Sweden), stained with uranyl acetate and lead citrate, and imaged using a JEOL 2100 transmission electron microscope (Tokyo, Japan) at 4000\u0026times; magnification. Ten images per nerve were captured from the ventral region of the optic nerve, spatially distant from the primary injury site but vulnerable to secondary degeneration, using Olympus iTEM Soft Imaging Solution software. A 3.05 mm copper support grid mounted beneath the sections ensured uniform segmentation and facilitated random sampling across the ventral optic nerve.\u003c/p\u003e\n\u003ch3\u003eTraining and evaluation of the AxonDeepSeg model\u003c/h3\u003e\n\u003cp\u003eThe AxonDeepSeg model (version 3.30) was developed on an NVIDIA GTX 1070 GPU using a dataset of 10 TEM images (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea for a representative original image). Model performance was assessed using 5-fold cross-validation [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. The dataset was randomly partitioned into five subsets, each containing two images. For each fold, eight images were used for model training, and the remaining two images were held out for evaluation. As AxonDeepSeg requires a validation dataset during training, one of the two held-out images was designated as a validation input. This validation image was used solely for satisfying the software workflow requirements and was not used for any hyperparameter tuning, model selection, or early stopping. Both held-out images were therefore independent of the training process and were used to compute performance metrics. As a result, five independent models were trained, and performance metrics were averaged across folds to obtain an overall estimate of model performance. We reported pixel-wise metrics, including False Negative Rate, False Positive Rate, Accuracy and Dice Similarity Coefficient and the ability of AxonDeepSeg to correctly identify axon versus myelin regions was compared using Student\u0026rsquo;s t-tests. Additionally, a two-way ANOVA was used to compare AxonDeepSeg-derived axon diameter measurements in sham and injured optic nerves with manually obtained measurements, which demonstrated consistency in morphometric estimation (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFollowing cross-validation, a final model was trained on the full dataset (all 10 images) using identical training parameters and used for the subsequent segmentation. Ground-truth labels (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb) were generated using the default AxonDeepSeg model, manually corrected in Adobe Photoshop CC, and reviewed independently by a second researcher before training. Blinding was achieved at the level of outcome assessment and data analysis by a third party assigning coded image identifications prior to manual measurements, with true identification revealed at completion. Blinding during allocation and conduct of the experiment was not possible due to the surgical requirements of the study.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSchematic overview of the optic nerve partial transection model in Piebald Virol-Glaxo rat highlighting the surgical site behind the right eye. A pop-out TEM image of an injured optic nerve cross section showcases the characteristic \u0026lsquo;mushroom\u0026rsquo; morphology at the primary injury site on dorsal aspect. A red frame in the ventral region represents the site of image collection and a second pop-out window shows an example of the high magnification TEM micrographs of axon and myelin cross sections. Myelin appears electron dense (dark) while axons appear lighter \u003cb\u003e(a)\u003c/b\u003e. The workflow for training AxonDeepSeg illustrated by a ground truth label generated by manual segmentation of the original TEM micrograph and the corresponding automated segmentation generated by AxonDeepSeg.. The ground truth labels contain 3 colour values: white (axon), grey (myelin), and black (background). The trained segmentation images contain 3 colour values: blue (axon), red (myelin), and black (background; \u003cb\u003eb\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eA suite of established and new morphometric features was assessed in AxonDeepSeg and manually as outcome measures (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). To ensure a robust morphological representation, all axon and myelin profiles within the selected frames were included in the training process, regardless of shape or size. The images were split into 256 \u0026times; 256-pixel patches and training was performed with a batch size of 4 for 300 epochs, taking approximately 2 hours per fold, for a total training time of approximately 10 hours. The final trained model was then used to perform automated segmentation (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb) on 68 images containing 1,213 total axons to generate the morphometric dataset. The segmented images and their corresponding morphometric outputs were visually quality checked. Decompacted interlamellar spaces that were incorrectly classified as axons were identified using their assigned object IDs, and the corresponding morphometric data were removed from the raw output prior to analysis.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDefinitions of morphometric features quantified by AxonDeepSeg as outcome measures\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeatures\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDefinition\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eg-ratio\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRatio of the axon diameter to the total fibre diameter (axon\u0026thinsp;+\u0026thinsp;myelin).\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAxon area (\u0026micro;m\u0026sup2;)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCross-sectional area of the axon compartment.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAxon perimeter (\u0026micro;m)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLength of the boundary enclosing the axon region.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMyelin area (\u0026micro;m\u0026sup2;)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDifference between the fibre area and the axon area, representing the myelin sheath and peri-axonal gap.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAxon diameter (\u0026micro;m)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDiameter of the axon, measured along the minor axis when modelled as an ellipse.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFibre area (\u0026micro;m\u0026sup2;)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCross-sectional area of the entire fibre (axon\u0026thinsp;+\u0026thinsp;myelin).\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSolidity\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCompactness of the axon, defined as the ratio of axon area to the area of a convex polygon containing the axon. Lower values indicate irregular or decompacted morphology.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eEccentricity\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMeasure of axonal circularity or flattening, where 0\u0026thinsp;=\u0026thinsp;perfect circle and values approaching 1 indicate elongation.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eOrientation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAngle between the major axis of the axon and the horizontal axis of the image.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAxon/Fibre area ratio\u003c/b\u003e *\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRatio of axon area to fibre area, indicating the relative thickness of the myelin sheath.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"2\"\u003e\u003cb\u003eNote.\u003c/b\u003e Definitions adapted from the AxonDeepSeg documentation (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://axondeepseg.readthedocs.io/en/latest/documentation.html#morphometrics-file)\u003c/span\u003e\u003cspan address=\"https://axondeepseg.readthedocs.io/en/latest/documentation.html#morphometrics-file)\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"2\"\u003e\u003cem\u003e* indicates metrics not exported by AxonDeepSeg but calculated subsequently.\u003c/em\u003e\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003eAnalysis of axon and myelin morphology\u003c/h3\u003e\n\u003cp\u003eTo determine whether AxonDeepSeg-derived morphometric data could detect injury-related changes in axon and myelin structure, unpaired Student\u0026rsquo;s \u003cem\u003et\u003c/em\u003e-tests were used to compare sham and injury groups for each independent morphological parameter. Normality and homogeneity of variance were verified using the Shapiro\u0026ndash;Wilk and Fligner\u0026ndash;Killeen tests, respectively. All assumptions were met except for the orientation parameter (W\u0026thinsp;=\u0026thinsp;0.7525, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.004), for which a non-parametric Mann\u0026ndash;Whitney \u003cem\u003eU\u003c/em\u003e test was applied.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eUnsupervised clustering of AxonDeepSeg-derived data\u003c/h2\u003e \u003cp\u003eTo explore fibre-level heterogeneity and identify morphologically distinct axon-myelin phenotypes, an unsupervised clustering analysis was performed on the complete morphometric dataset derived from AxonDeepSeg. The dataset comprised quantitative descriptors for each segmented axon, including both traditional ultrastructural features (axon diameter, g-ratio, myelin area) and non-traditional geometric parameters (solidity, eccentricity, orientation, and axon/fibre area ratio). These features were chosen to capture both the physical and topological variability of axons and their ensheathing myelin.\u003c/p\u003e \u003cp\u003eAll morphometric variables were standardised (z-score normalisation) prior to dimensionality reduction. Principal component analysis (PCA) was applied using the \u003cem\u003escikit-learn\u003c/em\u003e library [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e] to reduce data dimensionality while preserving the major sources of variance in the dataset. The first few principal components, which cumulatively explained most of the variance, were then used as inputs for spectral clustering. This method was selected for its robustness to non-linear relationships and ability to identify subgroups within datasets containing correlated features [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSpectral clustering was implemented in \u003cem\u003escikit-learn\u003c/em\u003e [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e], which computes a similarity matrix of samples based on feature proximity and transforms it into a low-dimensional embedding of the data, the labels are then assigned using the discretize method. The number of clusters was determined empirically by evaluating model stability and silhouette scores across different cluster counts. Six stable clusters were identified and retained for subsequent analyses.\u003c/p\u003e \u003cp\u003eTo verify the distinctness of each cluster, post-hoc statistical comparisons were conducted across all morphometric features using the Games\u0026ndash;Howell multiple comparisons test, which is appropriate for unequal sample sizes and heterogeneous variances. Bootstrapping was applied to confirm stability of the resulting group means. Clusters 1\u0026ndash;6 were then characterised by their mean g-ratio, axon diameter, and associated morphometric profiles.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eComparison of sham and injury groups within AxonDeepSeg-derived clusters\u003c/h3\u003e\n\u003cp\u003eTo examine injury effects within the phenotypic clusters defined by AxonDeepSeg, unpaired Student\u0026rsquo;s \u003cem\u003et\u003c/em\u003e-tests were used to compare axon and myelin features between sham and injury groups. Clusters 3\u0026ndash;6 were excluded from these comparisons due to insufficient sample sizes and incomplete animal representation. Clusters 1 and 2 contained sufficient samples to allow reliable statistical comparison.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eAutomated segmentation of optic nerve ultrastructure using AxonDeepSeg\u003c/h2\u003e \u003cp\u003eElectron micrographs of optic nerve cross sections were used to train the AxonDeepSeg model. A randomized dataset of 10 images of the ventral optic nerve collected 2 weeks after partial transection or sham surgery was used for model training and evaluation. Model performance was assessed using 5-fold cross-validation, and a final model was trained on all 10 images for downstream segmentation of axons and myelin in TEM images from the same optic nerves. Visual inspection confirmed that the trained AxonDeepSeg model effectively delineated axon and myelin boundaries across varying fibre morphologies, enabling reliable extraction of morphometric features for downstream analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eValidation of AxonDeepSeg model performance\u003c/h2\u003e \u003cp\u003eThe performance of the model was evaluated on the test image set. Pixelwise accuracy represents the proportion of correctly classified pixels within each compartment. The model achieved comparable pixelwise accuracy for axon segmentation (92.15%) and myelin segmentation (91.26%; t (18)\u0026thinsp;=\u0026thinsp;1.440, p\u0026thinsp;=\u0026thinsp;0.1671; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). Dice similarity achieved for axon segmentation (91.87%) was significantly higher than for myelin segmentation (85.74%; t (18)\u0026thinsp;=\u0026thinsp;9.429, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). In accordance with this, false-positive rates for myelin classification (7.67%) were significantly higher than axon classification (3.62%, t (18)\u0026thinsp;=\u0026thinsp;3.282, p\u0026thinsp;=\u0026thinsp;0.0041; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec). In contrast, false-negative rates were not significantly different regarding axon (11.84%) and myelin (11.47%) classification (t (18)\u0026thinsp;=\u0026thinsp;0.2024, p\u0026thinsp;=\u0026thinsp;0.8419; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ed). Importantly, axon diameter measurements derived from the AxonDeepSeg model did not differ from manually obtained values (F (1,16)\u0026thinsp;=\u0026thinsp;1.446, p\u0026thinsp;=\u0026thinsp;0.2466; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ee), validating the accuracy of the automated model against an established manual morphometric analysis.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eAxonDeepSeg identified injury-associated changes in axon and myelin morphology\u003c/h2\u003e \u003cp\u003eAutomated morphometric analyses revealed injury-specific alterations in optic nerve ultrastructure two weeks after partial transection. Automated analysis compared sham and injured optic nerve micrographs and revealed subtle but significant effects of injury, such as increased g-ratio (t (8)\u0026thinsp;=\u0026thinsp;2.661, p\u0026thinsp;=\u0026thinsp;0.0288; Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea), solidity (t (8)\u0026thinsp;=\u0026thinsp;3.040, p\u0026thinsp;=\u0026thinsp;0.0161; Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eg), and axon-to-fibre area ratio (t (8)\u0026thinsp;=\u0026thinsp;2.915, p\u0026thinsp;=\u0026thinsp;0.0194; Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ej), demonstrating sensitivity to subtle ultrastructural alterations and shifts in axon-myelin relationships following partial optic nerve injury. No significant differences were detected in axon diameter, axon perimeter, axon area, myelin area, eccentricity, orientation, or fibre area (all p\u0026thinsp;\u0026gt;\u0026thinsp;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb,c,d,e,f,h,i).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eUnsupervised clustering reveals distinct axonal phenotypes\u003c/h2\u003e \u003cp\u003eTo capture fibre-level heterogeneity, unsupervised clustering was applied to the morphometric dataset (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea). This approach, integrating both traditional (e.g., axon diameter, g-ratio) and non-traditional (e.g., solidity, eccentricity) morphometric descriptors, identified six distinct axon\u0026ndash;myelin phenotypes. Each cluster displayed a unique combination of quantitative features corresponding to recognisable ultrastructural profiles. Representative TEM images of each cluster (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb) illustrate visual differences between clusters.\u003c/p\u003e \u003cp\u003eCluster 1 (n\u0026thinsp;=\u0026thinsp;1080) comprised of compactly myelinated axons with moderate diameter and low g-ratio, representing structurally intact fibres. Cluster 2 (n\u0026thinsp;=\u0026thinsp;92) included myelinated axons of larger diameter with partially decompacted myelin and higher g-ratios. Cluster 3 (n\u0026thinsp;=\u0026thinsp;6) had reduced g-ratio and smaller fibre areas, consistent with axons surrounded by thickened, decompacted myelin. Cluster 4 (n\u0026thinsp;=\u0026thinsp;9) was characterised by large axon and fibre areas, extensive myelin disruption, and lower solidity. Cluster 5 (n\u0026thinsp;=\u0026thinsp;20) contained small, unmyelinated axons, while Cluster 6 (n\u0026thinsp;=\u0026thinsp;5) were large-diameter axons with moderately decompacted myelin and high g-ratio values. The mean values of morphometric features for each cluster (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) provide a practical reference for identifying similar phenotypic classes in independent datasets. Importantly, these values include routinely measurable parameters such as axon diameter and g-ratio, allowing researchers to approximate cluster membership using conventional manual morphometry, even without access to automated segmentation tools.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMean morphometric values for each cluster of optic nerve axons-myelin units\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCluster 1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCluster 2\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCluster 3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCluster 4\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCluster 5\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCluster 6\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eg-ratio\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.798ᵈ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.821\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.641\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.786ᵃ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.994\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.859\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eAxon area (\u0026micro;m\u0026sup2;)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.360ᶜ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.010\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.300ᵃᵉ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.590ᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.207ᶜ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.900ᵈ\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eAxon perimeter (\u0026micro;m)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2.850ᶜ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.320\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.300ᵃ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e9.360\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.850\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e7.260\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eAxon diameter (\u0026micro;m)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.522\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.889ᵈ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.403ᵉ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.968ᵇ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.412ᶜ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.340\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eFibre area (\u0026micro;m\u0026sup2;)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.607\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.820\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.320\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3.240ᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.211\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e3.040ᵈ\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eSolidity\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.930ᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.869ᶜᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.804ᵇᵈᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.704ᶜ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.976\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.885ᵃᵇᶜ\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eEccentricity\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.742ᵉᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.778ᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.891ᵈ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.909ᶜ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.687ᵃᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.715ᵃᵇᵉ\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eOrientation\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.103ᵇᶜᵈᵉ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.190ᵃᶜᵈᵉ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.143ᵃᵇᵈᵉᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.194ᵃᵇᶜᵈᵉᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.007ᵃᵇᶜᵈ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.739\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eAxon/Fibre area ratio\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.592\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.557ᵈ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.248\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.504ᵈ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.988\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.626\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eMyelin area (\u0026micro;m\u0026sup2;)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.248\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.820ᶜ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.020ᵇᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.650ᶠ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.004ᶜᵈ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.130\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003e\u003cb\u003eNote.\u003c/b\u003e Values represent mean morphometric measurements for each cluster derived from segmented TEM data of the ventral optic nerve. Superscripted letters indicate features not significantly different from the corresponding cluster (a\u0026thinsp;=\u0026thinsp;Cluster 1, b\u0026thinsp;=\u0026thinsp;Cluster 2, c\u0026thinsp;=\u0026thinsp;Cluster 3, d\u0026thinsp;=\u0026thinsp;Cluster 4, e\u0026thinsp;=\u0026thinsp;Cluster 5, f\u0026thinsp;=\u0026thinsp;Cluster 6). All other pairwise comparisons were significantly different across clusters (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05; Games\u0026ndash;Howell test). Mean g-ratio and axon diameter values provide accessible reference points for identifying comparable axon\u0026ndash;myelin phenotypes using manually measured morphometry.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eInjury induced cluster-specific alterations in axonal phenotype\u003c/h2\u003e \u003cp\u003eTo determine whether partial transection injury elicited distinct effects within specific axon populations, morphometric comparisons between injury and sham conditions were conducted for the more internally consistent Clusters 1 and 2 (Figs.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). In Cluster 1, injury resulted in significant increases in g-ratio (t (8)\u0026thinsp;=\u0026thinsp;2.837, p\u0026thinsp;=\u0026thinsp;0.0219; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea), solidity (t(8)\u0026thinsp;=\u0026thinsp;3.048, p\u0026thinsp;=\u0026thinsp;0.0157; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb), axon-to-fibre area ratio (t (8)\u0026thinsp;=\u0026thinsp;2.972, p\u0026thinsp;=\u0026thinsp;0.0178), axon diameter (t(8)\u0026thinsp;=\u0026thinsp;2.651, p\u0026thinsp;=\u0026thinsp;0.0292; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea), and axon area (t(8)\u0026thinsp;=\u0026thinsp;5.281, p\u0026thinsp;=\u0026thinsp;0.0007; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ee). In Cluster 2, injury similarly increased g-ratio (t (8)\u0026thinsp;=\u0026thinsp;2.809, p\u0026thinsp;=\u0026thinsp;0.0229; Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ea), solidity (t (8)\u0026thinsp;=\u0026thinsp;2.755, p\u0026thinsp;=\u0026thinsp;0.0248), and axon-to-fibre area ratio (t (8)\u0026thinsp;=\u0026thinsp;2.734, p\u0026thinsp;=\u0026thinsp;0.0257), though axon diameter and area were unaffected (p\u0026thinsp;\u0026gt;\u0026thinsp;0.1). Neither cluster exhibited significant changes in axon perimeter, myelin area, fibre area, eccentricity, or orientation (p\u0026thinsp;\u0026gt;\u0026thinsp;0.05). Together, these findings demonstrate that partial transection alters key ultrastructural features of axon\u0026ndash;myelin integrity in a cluster-dependent manner, revealing differential vulnerability among optic nerve fibre subtypes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe trained AxonDeepSeg model achieved high segmentation accuracy and reproducibility, enabling precise quantification of optic nerve ultrastructure following injury. Automated morphometric profiling revealed subtle but significant injury-associated increases in g-ratio, solidity and axon-to-fibre area ratio, indicative of early myelin decompaction. Unsupervised clustering uncovered six morphologically distinct axon populations, two of which (Clusters 1 and 2) exhibited pronounced injury-specific alterations. These data validate AxonDeepSeg as a robust tool for large-scale ultrastructural analysis and demonstrate its capacity to resolve previously unrecognised, cluster-specific responses of axon\u0026ndash;myelin units to focal injury.\u003c/p\u003e \u003cp\u003eConventional methods used to quantify axon and myelin ultrastructural features are typically based on manual measurement and visual assessment. These methods are subjective and labour-intensive, as measurements from hundreds of axons across multiple fields of view are required for statistically robust analyses. Furthermore, traditional metrics such as myelin thickness and g-ratio can only be measured where the myelin appears compact, excluding axons with partially or fully decompacted lamellae and thereby omitting potentially relevant pathological features. These limitations may contribute to inter-study variability, confounding comparisons across models that use different measurement or classification criteria.\u003c/p\u003e \u003cp\u003eHere, we described the modification and use of an existing AxonDeepSeg model to automatically segment axons and myelin and measure a broad range of morphometric features in transmission electron micrographs of the rat optic nerve following PT or a sham control injury. The trained model achieved high segmentation accuracy across multiple validation metrics, comparable to or exceeding previously published benchmarks [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. Automated axon diameter measurements were consistent with manual measurements [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], confirming that AxonDeepSeg accurately reproduced standard morphometric parameters with the added benefit of increased throughput and reduced bias. The automated analysis also detected significantly increased g-ratio (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea), which was not detected during manual measurement of the same images [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], implying AxonDeepSeg may be more sensitive to subtle and early injury effects than manual analysis. Additionally, AxonDeepSeg detected injury-effects in the non-traditional morphometrics of solidity and axon-to-fibre area ratio (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eg,j), further underscoring its enhanced sensitivity to detect early ultrastructural changes and altered axon-myelin relationships, and provide new insights to the morphological changes following partial optic nerve injury. Furthermore, we show that unsupervised clustering of AxonDeepSeg-derived morphometric data can be used to identify distinct categories of axons, and that these axon populations respond differently to injury. This approach enables a more objective and biologically meaningful understanding of abnormal myelination in the context of injury and secondary degeneration. Unlike conventional methods, which cannot reliably quantify g-ratio or myelin thickness in partially decompacted fibres, AxonDeepSeg permitted analysis of both conventional and non-conventional morphometric features across all axons, including those with irregular or disrupted myelin. Although segmentation of myelin was slightly less precise - likely reflecting the morphological heterogeneity of compact and decompacted sheaths - automated segmentation allowed inclusion of axons that would typically be excluded from manual analyses.\u003c/p\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eAutomated analysis reveals early injury-related ultrastructural alterations\u003c/h2\u003e \u003cp\u003eAxonDeepSeg-based morphometry detected subtle but consistent injury-associated changes two weeks after partial optic nerve transection, a time point previously reported to lack overt morphological differences by manual analysis [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. Injury increased g-ratio, solidity, and axon-to-fibre area ratio, whereas parameters such as axon area, fibre area, and perimeter remained unchanged. These findings indicate a mild reduction in relative myelin thickness and shape alterations suggestive of local structural stress rather than gross swelling. Increased solidity was evident in both the overall dataset and in the two largest fibre clusters (Clusters 1 and 2), while higher axon-to-fibre area ratio and g-ratio values suggest subtle thinning of the myelin sheath relative to the axon calibre. In the full dataset, these changes occurred without accompanying increases in axon diameter or fibre area, indicating a shift in axon-myelin geometry independent of swelling [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. In Cluster 1, however, increased solidity coincided with small increases in axon diameter and area, suggesting that some fibres may undergo mild distension or cytoskeletal reorganisation [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe biological interpretation of solidity remains exploratory, but it likely reflects geometric changes in the axonal contour rather than size alone. Axon shape is determined by cytoskeletal architecture, organelle positioning, and extracellular constraints [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. Subtle increases in solidity could therefore indicate reduced undulation or micro-beading of the axonal boundary, possibly due to early cytoskeletal disruption or redistribution of intracellular contents - changes known to occur after axonal stress or oxidative damage [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e, \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. These features would not be detected by traditional morphometrics, highlighting the added sensitivity of automated segmentation. Consistent with our previous work, manual analyses of the same cohort found no overt changes in axon diameter or myelin appearance at this time point [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. The present study confirmed agreement between manual and automated axon diameter measurements, while AxonDeepSeg revealed small but significant increases in g-ratio, solidity, and axon-to-fibre area ratio in the same images, demonstrating its superior sensitivity to early, pre-decompaction alterations in myelin geometry.\u003c/p\u003e \u003cp\u003eSmall increases in g-ratio and axon-to-fibre area ratio imply local thinning of the myelin sheath relative to the axon, possibly reflecting early lamellar separation or metabolic stress in oligodendrocytes [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Even subtle changes in myelin thickness or axonal geometry can influence conduction velocity and synchronisation within white-matter tracts [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. Although the functional implications of altered solidity remain unclear, the convergence of these morphometric shifts suggests the onset of fine-grained structural adaptations within the axon\u0026ndash;myelin unit that may precede overt decompaction or demyelination.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eAxonDeepSeg compared to existing segmentation tools\u003c/h2\u003e \u003cp\u003eA range of CNN-based segmentation frameworks have been developed for axon and myelin analysis, each varying in complexity, accessibility, and analytical breadth (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of convolutional neural network (CNN)-based segmentation tools for axon analysis\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSoftware / Tool\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReference\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eImage type(s)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDimensionality\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eOpen-source availability\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eGUI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUser friendliness\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eMorphometric capability\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAxonDeepSeg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOM, (EM), (LM)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e✓✓✓\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAxonSeg / AxonSeg3D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2D / 3D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e✓✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e✓✓\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMicroscopy Image Browser (MIB)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLM, EM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2D / 3D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e✓✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIlastik\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2D / 3D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDeepAxon\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eACSON / DeepACSON\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2D / 3D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e✓✓\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMiGA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e✓✓\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAxonEM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAxonDeepSeg with PairedImageTranslation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emulti-class CNNs and deep encoder-decoder architecture\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAxonDeep\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026ndash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e✓\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"8\" nameend=\"c8\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNote. OM\u003c/b\u003e\u0026thinsp;=\u0026thinsp;optical microscopy; \u003cb\u003eLM\u003c/b\u003e\u0026thinsp;=\u0026thinsp;light microscopy; \u003cb\u003eEM\u003c/b\u003e\u0026thinsp;=\u0026thinsp;electron microscopy.\u003c/p\u003e \u003cp\u003eUser friendliness was rated according to ease of adoption: ✓ = usable without prior expertise; ✓✓ = usable without detailed instruction.\u003c/p\u003e \u003cp\u003eMorphometric capability was rated according to the range of measurements provided: ✓ = basic (classical) measures; ✓✓ = extended (modern) measures; ✓✓✓ = advanced or unique morphometric metrics.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eMany frameworks, such as Microscopy Image Browser [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], DeepACSON [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e], and MiGA [\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e], achieve excellent segmentation fidelity but lack intuitive user interfaces or provide only basic geometric measurements. AxonDeepSeg offers several advantages for experimental neuroscientists: it is fully open-source, compatible with multiple microscopy modalities, and deployable through either command-line or graphical interfaces requiring minimal computational expertise. It also outputs both classical morphometric parameters (e.g., axon diameter, g-ratio) and non-traditional geometric descriptors such as solidity, eccentricity, and orientation, allowing multidimensional analysis of axon-myelin structure. However, as AxonDeepSeg defines continuous inner and outer myelin borders [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e], it may smooth over regions of myelin decompaction or irregular lamellae, potentially underestimating severe pathology. Future training sets incorporating more pathological examples could further improve sensitivity to these features.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eUnsupervised clustering identifies distinct axon-myelin phenotypes\u003c/h2\u003e \u003cp\u003eThe multidimensional dataset generated by AxonDeepSeg enabled unsupervised clustering of axons into six morphologically distinct groups based on combinations of traditional and geometric morphometrics. Most fibres were grouped in Clusters 1 or 2, which represented compactly myelinated and partially decompacted profiles, while Clusters 3 to 6 captured less frequent subtypes with thickened, disrupted, or absent myelin. This data-driven classification differed from earlier approaches that categorised axons solely by the proportion of compact versus decompacted myelin around their circumference [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] and instead used an integrated set of shape and size metrics to define axonal phenotypes.\u003c/p\u003e \u003cp\u003eComparisons between sham and injured nerves within each cluster revealed that injury effects were largely confined to Clusters 1 and 2, where significant increases in g-ratio, solidity, and axon/fibre area ratio were observed. These changes occurred in fibres that were mostly morphologically \u0026ldquo;normal,\u0026rdquo; implying that subtle ultrastructural remodelling may precede visible demyelination. Interestingly, Cluster 1, which visually appeared the most intact, showed the greatest number of altered parameters, suggesting that early axonal stress may manifest primarily through changes in axon geometry rather than overt myelin disruption. In contrast, clusters representing already abnormal or unmyelinated fibres showed minimal change, consistent with them representing stable or end-stage phenotypes at this time point. This pattern supports the idea that secondary degeneration affects distinct axonal subpopulations differently, in agreement with studies showing variable metabolic and stress resilience among oligodendrocyte\u0026ndash;axon units [\u003cspan additionalcitationids=\"CR57\" citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e]. It is possible that the relatively variable morphology of axons in Clusters 3\u0026ndash;5 may have precluded detection of changes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eInterpreting non-traditional morphometrics\u003c/h2\u003e \u003cp\u003eAmong the non-traditional morphometrics, solidity emerged as particularly informative. It quantifies how closely an axon\u0026rsquo;s cross-section approximates its convex hull and may therefore reflect cytoskeletal tension or compaction of the axoplasmic contents. Increases in solidity, together with elevated g-ratio and axon-to-fibre area ratio, suggest coordinated microstructural adaptations in both the axon and its myelin sheath. Similar contour changes have been reported following traumatic brain injury and oxidative stress, where cytoskeletal disruption - including neurofilament reorganisation - precedes overt axon swelling and alters axonal shape rather than simply increasing size [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e, \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e]. Together, these findings demonstrate that AxonDeepSeg can detect early morphological adaptations of axons and myelin that may precede overt degeneration or demyelination.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eMethodological and translational implications\u003c/h2\u003e \u003cp\u003eThis workflow bridges a long-standing methodological gap in ultrastructural neurobiology by providing an objective, high throughput means of quantifying axon and myelin morphology. By eliminating manual bias and including all fibres within an image, irrespective of pathology, it allows comprehensive and reproducible characterisation of white-matter microstructure. The mean morphometric values for each cluster (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) can serve as practical reference points for cross-dataset comparison, enabling other researchers to identify similar phenotypic classes using conventional measurements such as g-ratio and axon diameter. The approach can be readily extended to other models of neurotrauma and demyelination, including chronic traumatic encephalopathy and multiple sclerosis, where subtle axon-myelin mismatches are functionally important. Combining automated morphometry with molecular or electrophysiological data could clarify how specific morphometric signatures correspond to impaired conduction or altered axon-glia metabolism.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003eLimitations and future directions\u003c/h2\u003e \u003cp\u003eDespite its strengths, automated myelin segmentation is not without challenges. The boundary-based algorithm of AxonDeepSeg may underestimate severe lamellar pathology, and segmentation accuracy depends on the diversity of images used for training. Incorporating a wider range of pathological examples and adopting hybrid three-dimensional models could further enhance fidelity. Longitudinal studies are also warranted to determine whether the early morphometric changes identified here predict later demyelination or functional impairment.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn summary, AxonDeepSeg provides a reliable and scalable method for quantifying axon and myelin ultrastructure. By detecting increases in g-ratio, solidity, and axon-to-fibre area ratio following partial optic nerve injury, AxonDeepSeg revealed early, subtle deviations in axon\u0026ndash;myelin geometry that elude manual measurement. By integrating deep learning with unsupervised clustering, this approach enhances the precision and interpretive value of ultrastructural analysis, offering new opportunities to identify early morphological biomarkers of axonal stress and secondary degeneration.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTEM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003etransmission electron microscopy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCNNs\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003econvolutional neural networks\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMIB\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMicroscopy Image Browser\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eOM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eoptical microscopy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003elight microscopy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eEM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eelectron microscopy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eCompeting Interests\u003c/h2\u003e \u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThe funding for this research project was provided by the Neurotrauma Research Program WA (NRP) and was funded by the State Government of Western Australia through the Department of Health, as well as the National Health \u0026amp; Medical Research Fund (APP1087114). The funders had no role in the study design, data collection, and analysis, decision to publish, or preparation of the manuscript.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003ePP conceptualised the project and study design, conducted AxonDeepSeg model training,data analysis, data curation and validation, manual image analysis, figure preparation, data interpretation, and was a major contributor in writing the manuscript. BRL performed manual image analysis, data curation and validation, figure preparation, data interpretation, and was a major contributor in writing the manuscript. SCH performed manual image analysis, data curation and validation, data interpretation, and contributed to writing the manuscript. TM conducted animal experiments, tissue processing, image acquisition, manual image analysis, and reviewed the manuscript. AJW participated in project conceptualisation, conducted animal experiments, tissue processing, image acquisition, and reviewed the manuscript. NL conducted AxonDeepSeg model training and data analysis and reviewed the manuscript. ET conducted AxonDeepSeg model training and data analysis and reviewed the manuscript. CAB conducted animal experiments, tissue processing, and reviewed the manuscript. MF supervised the project, secured project funding, contributed to data interpretation, and edited and reviewed the manuscript. CCA conceptualised the project and study design, conducted project administration, supervision, animal experiments, tissue processing, image acquisition, AxonDeepSeg model training and data analysis, manual image analysis, data curation and validation, figure preparation, data interpretation, and was a major contributor in writing the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eNot applicable.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets supporting the conclusions of this article are available at Zenodo [DOI: 10.5281/zenodo.18686326; https://zenodo.org/records/18686326?preview=1\u0026amp;token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6IjAxZmQ5NjU1LWZlM2ItNDE0MC04ZmIzLWQ1YTJlODRiODJjMyIsImRhdGEiOnt9LCJyYW5kb20iOiIzZGFkYjdmMGM0YzY2OTBmMTk0MWFmN2RkZjgwZWFiZSJ9.cVDQbJmtpryYW2x0Nwaphvt0VKaPtWYC7k8TVZgDSe7TXQoQesVIte9PbKfA73o0tXccmlRLG7lS-42GUV2CwQ].All custom code used for data processing, AxonDeepSeg model training, and clustering analyses is available at a GitHub repository, [https://github.com/CurtinNeurotrauma/ads-morphometrics-clustering].\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eStadelmann C, Timmler S, Barrantes-Freer A, Simons M (2019) Myelin in the central nervous system: Structure, function, and pathology. Physiol Rev 99(3):1381\u0026ndash;1431. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1152/physrev.00031.2018\u003c/span\u003e\u003cspan address=\"10.1152/physrev.00031.2018\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ede Faria O, Pivonkova H, Varga B, Timmler S, Evans KA, K\u0026aacute;rad\u0026oacute;ttir RT (2021) Periods of synchronized myelin changes shape brain function and plasticity. Nat Neurosci 24(11):1508\u0026ndash;1521. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41593-021-00917-2\u003c/span\u003e\u003cspan address=\"10.1038/s41593-021-00917-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcKenzie IA, Ohayon D, Li H, Paes de Faria J, Emery B, Tohyama Ket al et al (2014) Motor skill learning requires active central myelination. Science 346(6207):318\u0026ndash;322. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1126/science.1254960\u003c/span\u003e\u003cspan address=\"10.1126/science.1254960\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNave K-A (2010) Myelination and the trophic support of long axons. Nat Rev Neurosci 11(4):275\u0026ndash;283. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/nrn2797\u003c/span\u003e\u003cspan address=\"10.1038/nrn2797\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFitzgerald M, Payne SC, Bartlett CA, Evill L, Harvey AR, Dunlop SA (2009) Secondary retinal ganglion cell death and the neuroprotective effects of the calcium channel blocker lomerizine. Investig Ophthalmol Vis Sci 50(11):5456\u0026ndash;5462. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1167/iovs.09-3717\u003c/span\u003e\u003cspan address=\"10.1167/iovs.09-3717\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLevkovitch-Verbin H, Quigley HA, Martin KRG, Zack DJ, Pease ME, Valenta DF (2003) A model to study differences between primary and secondary degeneration of retinal ganglion cells in rats by partial optic nerve transection. Investig Ophthalmol Vis Sci 44(8):3388\u0026ndash;3393. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1167/iovs.02-0646\u003c/span\u003e\u003cspan address=\"10.1167/iovs.02-0646\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRaineteau O, Schwab ME (2001) Plasticity of motor systems after incomplete spinal cord injury. Nat Rev Neurosci 2(4):263\u0026ndash;273. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/35067570\u003c/span\u003e\u003cspan address=\"10.1038/35067570\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBernardo A, Bianchi D, Magnaghi V, Minghetti L (2009) Peroxisome proliferator-activated receptor-γ agonists promote differentiation and antioxidant defenses of oligodendrocyte progenitor cells. J Neuropathology Experimental Neurol 68(7):797\u0026ndash;808. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/NEN.0b013e3181aba2c1\u003c/span\u003e\u003cspan address=\"10.1097/NEN.0b013e3181aba2c1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThorburne SK, Juurlink BHJ (1996) Low glutathione and high iron govern the susceptibility of oligodendroglial precursors to oxidative stress. J Neurochem 67(3):1014\u0026ndash;1022. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1046/j.1471-4159.1996.67031014.x\u003c/span\u003e\u003cspan address=\"10.1046/j.1471-4159.1996.67031014.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoutsoudaki PN, Hildebrandt H, Gudi V, Skripuletz T, Škuljec J, Stangel M (2010) Remyelination after cuprizone induced demyelination is accelerated in mice deficient in the polysialic acid synthesizing enzyme st8siaiv. Neuroscience 171(1):235\u0026ndash;244. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroscience.2010.08.070\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroscience.2010.08.070\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSachs HH, Bercury KK, Popescu DC, Narayanan SP, Macklin WB (2014) A new model of cuprizone-mediated demyelination/remyelination. ASN Neuro 6(5):1759091414551955. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/1759091414551955\u003c/span\u003e\u003cspan address=\"10.1177/1759091414551955\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePayne SC, Bartlett CA, Harvey AR, Dunlop SA, Fitzgerald M (2011) Chronic swelling and abnormal myelination during secondary degeneration after partial injury to a central nervous system tract. J Neurotrauma 28(6):1077\u0026ndash;1088. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1089/neu.2010.1665\u003c/span\u003e\u003cspan address=\"10.1089/neu.2010.1665\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePayne SC, Bartlett CA, Harvey AR, Dunlop SA, Fitzgerald M (2012) Myelin sheath decompaction, axon swelling, and functional loss during chronic secondary degeneration in rat optic nerve. Investig Ophthalmol Vis Sci 53(10):6093\u0026ndash;6101. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1167/iovs.12-10080\u003c/span\u003e\u003cspan address=\"10.1167/iovs.12-10080\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiewald D, Miller R, Logothetis N, Wagner H-J, Sch\u0026uuml;z A (2014) Distribution of axon diameters in cortical white matter: An electron-microscopic study on three human brains and a macaque. Biol Cybern 108(5):541\u0026ndash;557. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00422-014-0626-2\u003c/span\u003e\u003cspan address=\"10.1007/s00422-014-0626-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchindelin J, Arganda-Carreras I, Frise E, Kaynig V, Longair M, Pietzsch T al (2012) Fiji: An open-source platform for biological-image analysis. Nat Methods 9(7):676\u0026ndash;682. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/nmeth.2019\u003c/span\u003e\u003cspan address=\"10.1038/nmeth.2019\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnthes DL, Theriault E, Tator CH (1995) Characterization of axonal ultrastructural pathology following experimental spinal cord compression injury. Brain Res 702(1):1\u0026ndash;16. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/0006-8993(95)01028-6\u003c/span\u003e\u003cspan address=\"10.1016/0006-8993(95)01028-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArmstrong RC, Mierzwa AJ, Marion CM, Sullivan GM (2016) White matter involvement after tbi: Clues to axon and myelin repair capacity. Exp Neurol 275:328\u0026ndash;333. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.expneurol.2015.02.011\u003c/span\u003e\u003cspan address=\"10.1016/j.expneurol.2015.02.011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen HS-M, Holmes N, Liu J, Tetzlaff W, Kozlowski P (2017) Validating myelin water imaging with transmission electron microscopy in a rat spinal cord injury model. NeuroImage 153:122\u0026ndash;130. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.neuroimage.2017.03.065\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2017.03.065\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDonovan V, Kim C, Anugerah AK, Coats JS, Oyoyo U, Pardo ACet al et al (2014) Repeated mild traumatic brain injury results in long-term white-matter disruption. J Cereb Blood Flow Metabolism 34(4):715\u0026ndash;723. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/jcbfm.2014.6\u003c/span\u003e\u003cspan address=\"10.1038/jcbfm.2014.6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFehily B, Bartlett CA, Lydiard S, Archer M, Milbourn H, Majimbi M al (2019) Differential responses to increasing numbers of mild traumatic brain injury in a rodent closed-head injury model. J Neurochem 149(5):660\u0026ndash;678. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/jnc.14673\u003c/span\u003e\u003cspan address=\"10.1111/jnc.14673\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNashmi R, Fehlings MG (2001) Changes in axonal physiology and morphology after chronic compressive injury of the rat thoracic spinal cord. Neuroscience 104(1):235\u0026ndash;251. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/S0306-4522(01)00009-4\u003c/span\u003e\u003cspan address=\"10.1016/S0306-4522(01)00009-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReeves TM, Smith TL, Williamson JC, Phillips LL (2012) Unmyelinated axons show selective rostrocaudal pathology in the corpus callosum after traumatic brain injury. J Neuropathology Experimental Neurol 71(3):198\u0026ndash;210. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/NEN.0b013e3182482590\u003c/span\u003e\u003cspan address=\"10.1097/NEN.0b013e3182482590\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSullivan GM, Mierzwa AJ, Kijpaisalratana N, Tang H, Wang Y, Song S-K al (2013) Oligodendrocyte lineage and subventricular zone response to traumatic axonal injury in the corpus callosum. J Neuropathology Experimental Neurol 72(12):1106\u0026ndash;1125. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/nen.0000000000000009\u003c/span\u003e\u003cspan address=\"10.1097/nen.0000000000000009\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePham DL, Xu C, Prince JL (2000) Current methods in medical image segmentation. Annu Rev Biomed Eng 2(1):315\u0026ndash;337. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1146/annurev.bioeng.2.1.315\u003c/span\u003e\u003cspan address=\"10.1146/annurev.bioeng.2.1.315\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRomero E, Cuisenaire O, Denef JF, Delbeke J, Macq B, Veraart C (2000) Automatic morphometry of nerve histological sections. J Neurosci Methods 97(2):111\u0026ndash;122. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S0165-0270(00)00167-9\u003c/span\u003e\u003cspan address=\"10.1016/S0165-0270(00)00167-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu T, Jurrus E, Seyedhosseini M, Ellisman M, Tasdizen T (2012) Watershed merge tree classification for electron microscopy image segmentation. Proc IAPR Int Conf Pattern Recogn 2012 133\u0026ndash;137\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Y-Y, Sun Y-N, Lin C-CK, Ju M-S (2012) Segmentation of nerve fibers using multi-level gradient watershed and fuzzy systems. Artif Intell Med 54(3):189\u0026ndash;200. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.artmed.2011.11.008\u003c/span\u003e\u003cspan address=\"10.1016/j.artmed.2011.11.008\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMore HL, Chen J, Gibson E, Donelan JM, Beg MF (2011) A semi-automated method for identifying and measuring myelinated nerve fibers in scanning electron microscope images. J Neurosci Methods 201(1):149\u0026ndash;158. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jneumeth.2011.07.026\u003c/span\u003e\u003cspan address=\"10.1016/j.jneumeth.2011.07.026\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao X, Pan Z, Wu J, Zhou G, Zeng Y (2010) Automatic identification and morphometry of optic nerve fibers in electron microscopy images. Comput Med Imaging Graph 34(3):179\u0026ndash;184. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.compmedimag.2009.08.009\u003c/span\u003e\u003cspan address=\"10.1016/j.compmedimag.2009.08.009\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBelevich I, Jokitalo E (2021) Deepmib: User-friendly and open-source software for training of deep learning network for biological image segmentation. PLoS Comput Biol 17(3):e1008374. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pcbi.1008374\u003c/span\u003e\u003cspan address=\"10.1371/journal.pcbi.1008374\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZaimi A, Duval T, Gasecka A, C\u0026ocirc;t\u0026eacute; D, Stikov N, Cohen-Adad J (2016) Axonseg: Open source software for axon and myelin segmentation and morphometric analysis. Front Neuroinform 10:37. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fninf.2016.00037\u003c/span\u003e\u003cspan address=\"10.3389/fninf.2016.00037\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBehanova A, Abdollahzadeh A, Belevich I, Jokitalo E, Sierra A, Tohka J (2022) Gacson software for automated segmentation and morphology analyses of myelinated axons in 3d electron microscopy. Comput Methods Programs Biomed. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.cmpb.2022.106802\u003c/span\u003e\u003cspan address=\"10.1016/j.cmpb.2022.106802\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. 220 106802 doi:\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePlebani E, Biscola NP, Havton LA, Rajwa B, Shemonti AS, Jaffey D al (2022) High-throughput segmentation of unmyelinated axons by deep learning. Sci Rep 12(1):1198. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-022-04854-3\u003c/span\u003e\u003cspan address=\"10.1038/s41598-022-04854-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSommer C, Straehle C, K\u0026ouml;the U, Hamprecht FA, Ilastik Interactive learning and segmentation toolkit In \u003cem\u003e2011 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2011/03// 2011\u003c/em\u003e (pp. 230\u0026ndash;233). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ISBI.2011.5872394\u003c/span\u003e\u003cspan address=\"10.1109/ISBI.2011.5872394\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZaimi A, Wabartha M, Herman V, Antonsanti P-L, Perone CS, Cohen-Adad J (2018) Axondeepseg: Automatic axon and myelin segmentation from microscopy data using convolutional neural networks. Sci Rep 8(1):3816. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-018-22181-4\u003c/span\u003e\u003cspan address=\"10.1038/s41598-018-22181-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFitzgerald M, Bartlett CA, Evill L, Rodger J, Harvey AR, Dunlop SA (2009) Secondary degeneration of the optic nerve following partial transection: The benefits of lomerizine. Exp Neurol 216(1):219\u0026ndash;230. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.expneurol.2008.11.026\u003c/span\u003e\u003cspan address=\"10.1016/j.expneurol.2008.11.026\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLins BR, Anyaegbu CC, McGonigle T, Hellewell SC, Patel P, Reagan H al (2023) Secondary degeneration impairs myelin ultrastructural development in adulthood following adolescent neurotrauma in the rat optic nerve. Int J Mol Sci 24(4):3343. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/ijms24043343\u003c/span\u003e\u003cspan address=\"10.3390/ijms24043343\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBartlett CA, Fitzgerald M (2018) Partial transection of adult rat optic nerve as a model of secondary degeneration in the central nervous system. Bio Protoc 8(24):e3118. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21769/bioprotoc.3118\u003c/span\u003e\u003cspan address=\"10.21769/bioprotoc.3118\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHastie T, Tibshirani R, Friedman J (2009) The elements of statistical learning: Data mining, inference, and prediction. Springer-, New York\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B Grisel O (2011) Scikit-learn: Machine learning in python. J Mach Learn Res 12(85) 2825\u0026ndash;2830\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evon Luxburg U (2007) A tutorial on spectral clustering. Stat Comput 17(4):395\u0026ndash;416. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s11222-007-9033-z\u003c/span\u003e\u003cspan address=\"10.1007/s11222-007-9033-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePozo Devoto VM, Lacovich V, Feole M, Bhat P, Chovan J, Čarna M al (2022) Unraveling axonal mechanisms of traumatic brain injury. Acta Neuropathol Commun 10(1):140. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s40478-022-01414-8\u003c/span\u003e\u003cspan address=\"10.1186/s40478-022-01414-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang J, Hamm RJ, Povlishock JT (2011) Traumatic axonal injury in the optic nerve: Evidence for axonal swelling, disconnection, dieback, and reorganization. J Neurotrauma 28(7):1185\u0026ndash;1198. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1089/neu.2011.1756\u003c/span\u003e\u003cspan address=\"10.1089/neu.2011.1756\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeterrier C (2021) Putting the axonal periodic scaffold in order. Curr Opin Neurobiol 69:33\u0026ndash;40. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.conb.2020.12.015\u003c/span\u003e\u003cspan address=\"10.1016/j.conb.2020.12.015\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGiacci MK, Bartlett CA, Smith NM, Iyer KS, Toomey LM, Jiang H al (2018) Oligodendroglia are particularly vulnerable to oxidative damage after neurotrauma in vivo. J Neurosci 38(29):6491\u0026ndash;6504. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1523/JNEUROSCI.1898-17.2018\u003c/span\u003e\u003cspan address=\"10.1523/JNEUROSCI.1898-17.2018\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCummins N, Bartlett CA, Archer M, Bartlett E, Hemmi JM, Harvey AR al (2013) Changes to mitochondrial ultrastructure in optic nerve vulnerable to secondary degeneration in vivo are limited by irradiation at 670 nm. BMC Neurosci 14(1):98. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/1471-2202-14-98\u003c/span\u003e\u003cspan address=\"10.1186/1471-2202-14-98\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaab AS, Nave K-A (2017) Myelin dynamics: Protecting and shaping neuronal functions. Curr Opin Neurobiol 47:104\u0026ndash;112. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.conb.2017.09.013\u003c/span\u003e\u003cspan address=\"10.1016/j.conb.2017.09.013\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArancibia-C\u0026aacute;rcamo IL, Ford MC, Cossell L, Ishida K, Tohyama K, Attwell D (2017) Node of ranvier length as a potential regulator of myelinated axon conduction speed. eLife 6:e23329. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.7554/eLife.23329\u003c/span\u003e\u003cspan address=\"10.7554/eLife.23329\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDutta DJ, Woo DH, Lee PR, Pajevic S, Bukalo O, Huffman WC al (2018) Regulation of myelin structure and conduction velocity by perinodal astrocytes. Proc Natl Acad Sci U S A 115(46):11832\u0026ndash;11837. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1073/pnas.1811013115\u003c/span\u003e\u003cspan address=\"10.1073/pnas.1811013115\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoiseev D, Hu B, Li J (2019) Morphometric analysis of peripheral myelinated nerve fibers through deep learning. J Peripher Nerv Syst 24(1):87\u0026ndash;93. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/jns.12293\u003c/span\u003e\u003cspan address=\"10.1111/jns.12293\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbdollahzadeh A, Belevich I, Jokitalo E, Sierra A, Tohka J (2021) Deepacson automated segmentation of white matter in 3d electron microscopy. Commun Biol 4(1):179. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s42003-021-01699-w\u003c/span\u003e\u003cspan address=\"10.1038/s42003-021-01699-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWei D, Lee K, Li H, Lu R, Bae JA Liu Z Axonem dataset: 3d axon instance segmentation of brain cortical regions In de Bruijne M, Cattin PC, Cotin S, Padoy N, Speidel S, Y Zheng, (eds), \u003cem\u003e2021 2021\u003c/em\u003e (pp. 175\u0026ndash;185, Lecture Notes in Computer Science). Cham: Springer International Publishing. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-3-030-87193-2_17\u003c/span\u003e\u003cspan address=\"10.1007/978-3-030-87193-2_17\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDweiri Y, Al-Zanina M, Durand D (2022) Enhanced automatic morphometry of nerve histological sections using ensemble learning. Electronics 11(14):2277. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/electronics11142277\u003c/span\u003e\u003cspan address=\"10.3390/electronics11142277\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMesbah R, McCane B, Mills S Deep convolutional encoder-decoder for myelin and axon segmentation In (2016) \u003cem\u003eInternational Conference on Image and Vision Computing New Zealand (IVCNZ), 2016/11// 2016\u003c/em\u003e (pp. 1\u0026ndash;6). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/IVCNZ.2016.7804455\u003c/span\u003e\u003cspan address=\"10.1109/IVCNZ.2016.7804455\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeng W, Hedberg-Buenz A, Soukup DA, Taghizadeh S, Wang K, Anderson MG al (2021) Axondeep: Automated optic nerve axon segmentation in mice with deep learning. Translational Vis Sci Technol 10(14):22. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1167/tvst.10.14.22\u003c/span\u003e\u003cspan address=\"10.1167/tvst.10.14.22\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGiacci MK, Bartlett CA, Huynh M, Kilburn MR, Dunlop SA, Fitzgerald M (2018) Three dimensional electron microscopy reveals changing axonal and myelin morphology along normal and partially injured optic nerves. Sci Rep 8(1):3979. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-018-22361-2\u003c/span\u003e\u003cspan address=\"10.1038/s41598-018-22361-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKn\u0026ouml;ferle J, Koch JC, Ostendorf T, Michel U, Planchamp V, Vutova P (2010) al Mechanisms of acute axonal degeneration in the optic nerve in vivo. Proceedings of the National Academy of Sciences 107(13) 6064\u0026ndash;6069 doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1073/pnas.0909794107\u003c/span\u003e\u003cspan address=\"10.1073/pnas.0909794107\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee Y, Morrison BM, Li Y, Lengacher S, Farah MH, Hoffman PN et al (2012) Oligodendroglia metabolically support axons and contribute to neurodegeneration. Nature 487(7408) 443\u0026ndash;448 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/nature11314\u003c/span\u003e\u003cspan address=\"10.1038/nature11314\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSiedler DG, Chuah MI, Kirkcaldie MTK, Vickers JC, King AE (2014) Diffuse axonal injury in brain trauma: Insights from alterations in neurofilaments. Front Cell Neurosci 8. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fncel.2014.00429\u003c/span\u003e\u003cspan address=\"10.3389/fncel.2014.00429\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Transmission electron microscopy, Axons, Myelin, Machine learning, AxonDeepSeg","lastPublishedDoi":"10.21203/rs.3.rs-8951825/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8951825/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eReliable quantification of axon and myelin ultrastructure is essential for understanding white-matter pathology, yet conventional manual methods are subjective, labour-intensive, and limited to compactly myelinated fibres. Here, we adapted and validated a custom AxonDeepSeg model for automated segmentation of transmission electron micrographs from the rat optic nerve following partial transection or sham surgery.\u003c/p\u003e \u003cp\u003eThe trained model achieved accuracy and Dice similarity indices comparable to published benchmarks, and automated axon diameter measurements closely matched manual quantification. Beyond reproducing traditional morphometrics, AxonDeepSeg detected subtle yet significant increases in g-ratio, solidity, and axon-to-fibre area ratio two weeks post-injury-changes that were undetectable by manual analysis-indicating early alterations in axon-myelin geometry preceding overt demyelination.\u003c/p\u003e \u003cp\u003eUnsupervised clustering of multidimensional morphometric data identified six distinct axon\u0026ndash;myelin phenotypes, two of which (Clusters 1 and 2) exhibited pronounced injury-related remodelling not apparent on assessment of the full, un-clustered dataset. These findings demonstrate that automated morphometry not only enhanced throughput and reproducibility but also increased sensitivity to otherwise undetectable microstructural changes in subsets of myelinated axons. This integrative workflow provides a scalable framework for identifying early and subtle indicators of axonal pathology and secondary degeneration in neurotrauma and other demyelinating conditions.\u003c/p\u003e","manuscriptTitle":"Unsupervised clustering of AxonDeepSeg morphometrics uncovers early patterns of axon-myelin secondary degeneration following neurotrauma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-27 06:29:56","doi":"10.21203/rs.3.rs-8951825/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e738342d-abf5-4a33-bc05-8af5fc55e151","owner":[],"postedDate":"April 27th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-27T06:29:56+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-27 06:29:56","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8951825","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8951825","identity":"rs-8951825","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}