Fig 1.
The myelin sheath plays a crucial role in the nervous system, making imaging essential for studying myelin formation, degeneration, and regeneration.
(A) Schematic representation of the process of myelin formation in the central nervous system. (B) Myelin formation is an ongoing process, starting during development (myelination) and continuing throughout lifespan. Myelin regeneration (remyelination) can occur in response to demyelination. Failure in remyelination contributes to axonal degeneration. (C) Conventionally for myelin g-ratio analysis, the axon area (green) has been annotated at the inner edge of the compact myelin, thus ignoring the contribution of the area occupied by the inner tongue (diagonal orange stripes). The myelin g-ratio is determined by assimilating the axon and fibre areas to circles to estimate their respective diameters. This conventionally disregards any area contribution from the inner tongue.
Fig 2.
Validation ground truth for the segmentation of fibre cross-sections on electron microscopy images separating the myelin sheath components (compact myelin and inner tongue) from each other and from the axon.
(A, top) Examples of transmission electron microscopy (TEM) images of the corpus callosum from adult mice undergoing remyelination after inducing a demyelinating lesion. Technical artefacts of no interest, degraded myelin debris and degenerated dark axons (red asterisk) were not included. Scale bar (red line) = 1 μm. (A, bottom) Manual segmentation of the compacted myelin (blue), the inner tongue (orange), and the axon (green). (B-D) Diversity of axon/fibre size, shape or myelin thickness. (B) Histograms representing different metrics determined from the manual annotations. (C) The fibres are colour-coded based on the histogram bins to represent the distribution of fibre eccentricity, describing how much a fibre section diverges from a circle, with 0.0 representing a perfect circle. (D) The fibres are colour-coded based on the histogram bins that illustrate their myelin g-ratio distribution (ratio of diameter of the area enclosed by the innermost compact myelin border and the diameter of the whole fibre). Higher g-ratios correspond to thinner myelin, with 1.0 representing the complete absence of myelin sheath. It is worth noting that this metric does not account for the presence of the inner tongue. As a result, two fibres, one with a shrunken inner tongue and another with an enlarged inner tongue, can exhibit similar myelin g-ratios (white asterisks).
Fig 3.
Main steps of the AimSeg bioimage analysis workflow.
(A-D) AimSeg combines two machine learning (ML) classifiers for pixel and object classification. First, the pixel classifier uses (A) the electron microscopy (EM) data to generate a (B) probability map. Then, (C) potential axon instances are segmented from the axoplasm probabilities. (D) The object classifier scores each instance as an axon or inner tongue. Objects outside myelinated fibre cross-sections (marked with an asterisk) will be eliminated in the next steps. (E-H) AimSeg Stage 1 uses (E) the myelin probability channel to get an (F) inverted mask. (G) This mask is then analysed to identify the elements within the innermost compact myelin border, which we call the ‘inner region’, and to exclude those representing the background. These identified elements are categorised as either selected or rejected ROIs, respectively. Running the supervised mode (optional), the user can easily toggle the ROI selection group (selected/rejected) or use the ImageJ’s selection tools to add/edit ROIs. (H) Semantic segmentation at the end of Stage 1. (I-L) AimSeg Stage 2 (I) uses the inner region labels as seeds that expand to fill myelin regions generating (J) a label mask for the fibres, which is processed to get (K) the fibre ROIs. (L) Semantic segmentation at the end of Stage 2. (M-P) AimSeg Stage 3 combines (M) the prediction of axon and inner tongue instances with (N) the fibre binary mask. (O) This ensures that only myelinated axons are selected. Instances classified as inner tongue are marked as rejected ROIs in the supervised mode. (P) Semantic segmentation at stage 3. (Q-T) AimSeg combines the gathered sets of ROIs to conduct a thorough analysis of myelinated axons. In this process, AimSeg assigns labels to the instances of (Q) the fibre, (R) the inner region and (S) the axon establishing a hierarchical relationship among instances within the same myelinated axon. (T) Additionally, AimSeg generates a semantic mask, where each pixel is categorised as background, axon, inner tongue, or compact myelin. Scale bar (red line) = 1 μm.
Fig 4.
AimSeg segmentation performance, assessed independently for the three detections performed sequentially by AimSeg: the inner region (i.e., the axon plus the inner tongue), the fibre and the axon.
Evaluation of the instance segmentation performed either in (A, B) automated or (C) supervised modes. At first, since no user intervention was allowed, results included both (A) accurate and (B) loose segmentations. Note that skipping an inner region at Stage 1 caused the myelin mask of the surrounding fibres to overflow at Stage 2. (C) The supervised mode allows the user to curate the AimSeg selection during the user-edited stages. Note that toggling the rejected inner region at Stage 1 solves the overflowing issue at Stage 2 and facilitates the automated detection of the corresponding fibre and axon. (D) Quantitation of the segmentation performance is based on the F1 score, an object-based metric, plotted for increasing intersection over union (IoU) thresholds for estimating the shape matching accuracy in both the automated and the human-supervised results. Scale bar (white line) = 0.5 μm.
Fig 5.
Agreement between the validation ground truth and the AimSeg measurements for the analysis of myelin properties.
(A) Example image for the semantic segmentation of the fibres. Scale bar (white line) = 1 μm. (B) Comparison of the fibre areas obtained by manually segmenting the images or using AimSeg. (B, left) The measurement agreement is calculated as the Lin’s concordance correlation coefficient (CCC), (B, right) while the measurement bias is assessed by means of a Bland-Altman analysis. Diagonal line in the CCC plot represents perfect agreement (y = x). (C) Illustration of the myelin g-ratio measurement, calculated as the ratio of the inner region diameter to the fibre diameter. (D) Illustration of the axon g-ratio measurement, calculated as the ratio of the axon diameter to the fibre diameter. CCC and Bland-Altman plot for (E) the myelin and (F) the axon g-ratios.
Fig 6.
Segmentation metrics and agreement between the control ground truth and the AimSeg measurements for the analysis of myelin properties.
(A) Segmentation performance for the instances detected at each AimSeg stage. (A, left) F1 score plotted for increasing intersection over union (IoU) thresholds. (A, right) Average F1 score, precision, recall, and Jaccard index. (B) Comparison of the fibre areas obtained by manually segmenting the images or using AimSeg. (B, left) The measurement agreement is calculated as the Lin’s concordance correlation coefficient (CCC), (B, right) while the measurement bias is assessed by means of a Bland-Altman analysis. Diagonal line in the CCC plot represents perfect agreement (y = x). CCC and Bland-Altman plot for (C) the myelin and (D) the axon g-ratios.