Figure S1. Algorithm to exclude fibres that have been truncated by the cut lines of tiles. An annotated fascicle was divided into small tiles of 250 × 250 pixels with a 200-pixel stride and a 25-pixel overlap zone with adjacent tiles. Boxes with borders truncated by the edge of the tile and centres in the overlap zone (outside green rectangles) were excluded (fibres with black boxes). Only boxes centred within the overlapping zone (fibres with red boxes) were counted for further analysis to avoid double counting between adjacent tiles. Scale bar: 10 μm.

Figure S2. Detection of nerve fibres in specimens with various artefacts. (A-D) Representative images show samples exhibiting various artefacts: (A) crush artefact, (B) tissue shrinkage likely due to a hyperosmolar solution, (C) obliquely sectioned sample, and (D) pale staining. Our software effectively identified myelinated nerve fibres in samples with these artefacts except in obliquely sectioned samples, where fibre detection may be unreliable (arrows). Scale bars: 10 μm (top row), 2 μm (bottom row).

Figure S3. Fibre detection under different imaging conditions. (A-D) Detection of myelinated fibres on an original image (A) and images after different digital imaging processes: contrast enhancement (B), greyscale conversion (C), blurring and contrast reduction (D). Scale bars: 20 μm.

Figure S4. Consistency of automated and manual measurements. (A-E) The fibre diameter in the automated method is defined by its cross-sectional area assuming an ideal circle, while that in the manual method is defined as the average length of the maximum and its perpendicular diameter. To compare the methodological consistency of the automated and manual measurements, Bland–Altman plots of fibre diameter in pixels (A), fibre diameter in μm (B), axon diameter in pixels (C), axon diameter in μm (D), and g-ratio (E) were evaluated, with the average of the values from the two methods on the x-axes and the differences (automated minus manual measurement) on the y-axes (n = 260). Bias and LoA were defined as the mean of the differences, and bias ± 1.94 SD, respectively (dashed lines). (F) Representative fibres are shown with annotations. Fibre (red) and axon (orange) contours of myelin sheaths were segmented by machine learning software. Straight lines for fibre (red) and axon (orange) diameters were manually annotated. Each figure shows a representative example of good agreement between the two methods (top left), discrepancy in the diameter between the methods in a distorted fibre (top right), overestimation of g-ratio in the automated method (bottom left), and underestimation of g-ratio in the automated method due to an 8-shaped fibre (bottom right). For further analysis, fibres with myelin discontinuities (open ring fibres) are excluded by comparison with their fitted ellipses. Scale bars: 2 μm. LoA, limits of agreement.

Figure S5. Pathological subcategories of the non-specific axonopathy group. (A-D) Representative figures for the pathological subcategories of the non-specific axonopathy group: acute (A), acute and chronic (B), and chronic (C) axonal changes, and end-stage neuropathy (D). Acute axonal change is characterised by samples with ovoid formations (arrows) or phagocytosed myelin (asterisks) (A, B). Chronic axonal change is characterised by samples exhibiting mild to moderate fibre loss and clusters of regenerating nerve fibres (arrowheads) (B, C). Regardless of these acute or chronic axonal findings, a sample with remarkable fibre loss is categorised as an end-stage neuropathy (D). Scale bar: 20 μm.

Table S1. Inference time taken to perform whole slide morphometry.

Table S2. Pathological subcategories of the non-specific axonopathy.

Table S3. Measurements of sural nerve biopsy in pathological subcategories of the non-specific axonopathy.

Table S4. Morphometry of nerve fibres.

Table S5. Spatial analysis of each fascicle.