What it is
A cross-section of nerve is packed with myelinated fibres: an axon wrapped in a myelin sheath, pressed against its neighbours. Measuring them one by one, how thick the myelin is and how big the axon is, means outlining each fibre and the axon inside it. AxonPath is a set of such outlines, made in QuPath to train and evaluate the AxonPath segmentation models, and it spans three kinds of microscope and both the central and peripheral nervous system.
| Sub-dataset | Tissue | Species | Microscope | Images | Fibres |
|---|---|---|---|---|---|
| CorpusCallosum_Mus_TEM_000 | corpus callosum (CNS) | mouse | TEM | 20 | 2,074 |
| Fimbria_Rat_STEM_000 | fimbria (CNS) | rat | STEM | 1 | 420 |
| OpticNerve_Mus_STEM_000 | optic nerve (CNS) | mouse | STEM | 1 | 433 |
| OpticNerve_Mus_TEM_000 | optic nerve (CNS) | mouse | TEM | 10 | 1,084 |
| SciaticNerve_Mus_STEM_000 | sciatic nerve (PNS) | mouse | STEM | 3 | 470 |
| SciaticNerve_Mus_TEM_000 | sciatic nerve (PNS) | mouse | TEM | 9 | 921 |
| TibialNerve_Mus_BF_000 | tibial nerve (PNS) | mouse | brightfield | 3 | 4,497 |
| All | 47 | 9,899 |
| Fact | Value |
|---|---|
| Record | Zenodo 23084690, version 1, published 2026-10-01 |
| Classes | Fibre, Axon, InnerCylinder (EM only), Nerve (brightfield only), and an unclassified ROI marking the annotated region |
| Format | QuPath GeoJSON in full-resolution pixel coordinates, plus training exports (semantic masks and instance labels) |
| Calibration | pixel size per image in pixel_sizes.csv, from about 0.002 µm (corpus callosum TEM) to 0.071 µm (tibial nerve brightfield) |
| Size | about 1.8 GB in seven zips, one per sub-dataset |
| License | CC BY 4.0, all seven sub-datasets (SOURCES.csv) |
All counts in these tables are the record's own. Only fibres inside each image's ROI were annotated, so the edges of an image can hold unlabelled fibres.
Browse it in DataTorch
Dataset · datatorch/axonpath-myelinated-fibres · 93 files
AxonPath Myelinated Nerve Fibre Segmentation
Myelinated nerve fibre segmentation from the AxonPath datasets: 47 electron microscopy (TEM/STEM) and brightfield images of mouse and rat CNS and PNS tissue in 7 sub-datasets, with 9,899 annotated fibres. Hosted copy: 77 files (27 large images split into strips) and 25,765 polygon annotations (Fibre, InnerCylinder, Axon, Nerve, ROI) imported from the authors' QuPath GeoJSON. License: CC BY 4.0 for all 7 sub-datasets (record, README and SOURCES.csv; the 8 AimSeg corpus callosum images are CC BY 4.0 too). Source: https://zenodo.org/records/23084690. Please cite: Carrillo-Barberà, P., Poveda Sabuco, A., Sgattoni, C., Gomez-Sanchez, J. A., Rondelli, A. M., & Williams, A. (2026). AxonPath datasets: annotated electron microscopy and brightfield images of myelinated nerve fibres (v1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23084690
Open →
The hosted copy holds all seven sub-datasets: 77 image files and 25,765 polygons in the classes Fibre, InnerCylinder, Axon, Nerve and ROI, imported from the authors' QuPath GeoJSON. The 27 largest images were cut into strips to fit the upload limit, with pixel values, bit depth and calibration unchanged and each strip's offset kept in its metadata. A shape that crossed a strip border became two shapes, so the copy has more polygons than the source's 25,025.
The import checks found three small things, kept in errata.txt in the project: one ROI that runs a pixel outside its image, one Axon with no Fibre around it (in OpticNerve_Mus_STEM_000, which the authors already list as a known issue), and six polygon holes that were dropped.
Open any image to see each fibre and its axon, zoom into a dense bundle, and correct or add outlines with your team. Exports come out as COCO.
What people use it for
Training and benchmarking instance segmentation of myelinated fibres, and measuring fibres one by one: axon size, myelin thickness and the g-ratio that relates them, in studies of myelination, demyelination and nerve injury. Because the sub-datasets cover three microscopes and two nervous systems, it is also a way to test whether a model trained on one tissue or microscope holds up on another.
Source and citation
Source: https://zenodo.org/records/23084690 (read 2026-11-11). Cite: Carrillo-Barberà, P., Poveda Sabuco, A., Sgattoni, C., Gomez-Sanchez, J. A., Rondelli, A. M., & Williams, A. (2026). AxonPath datasets: annotated electron microscopy and brightfield images of myelinated nerve fibres (v1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23084690. The training code and the QuPath extension are at github.com/paucabar/axonpath and github.com/paucabar/qupath-extension-axonpath.
Outline your own fibres
Label microscopy images with your team, review each other's work and export COCO, in the browser.
Frequently asked questions
What does AxonPath label?
Inside an annotated region (ROI) of each image, every myelinated fibre is outlined as a Fibre (axon, inner tongue and myelin sheath together), with its Axon outlined inside it. EM images also have an InnerCylinder (axon plus inner tongue). Brightfield images have a Nerve outline instead, and no InnerCylinder, because the inner tongue cannot be resolved there.
Which tissues and microscopes are in it?
Seven sub-datasets. Mouse corpus callosum (TEM, 20 images), rat fimbria (STEM, 1), mouse optic nerve (STEM, 1, and TEM, 10), mouse sciatic nerve (STEM, 3, and TEM, 9) and mouse tibial nerve (brightfield, 3). Four are central nervous system and three peripheral.
Can I measure g-ratio from these labels?
The labels pair each fibre with the axon inside it, which is what a g-ratio (axon size over fibre size) is computed from, and the record lists g-ratio among its keywords. The record does not publish g-ratio values, so any you compute are your own.
Why does the hosted copy have 77 files when the source has 47 images?
27 of the source images were larger than the upload limit, so each was cut into strips. Pixel values, bit depth and calibration are unchanged, and each strip records its offset in the original image in its metadata. Shapes that crossed a strip border were split, which is why the copy has 25,765 shapes against 25,025 in the source.
What is the license?
Creative Commons Attribution 4.0 International (CC BY 4.0), for all seven sub-datasets. A preprint is in preparation, so until it appears the authors ask you to cite the Zenodo record, DOI 10.5281/zenodo.23084690.
Open the hosted copy on DataTorch.
Browse the files in your browser, then download the dataset with a free account.
