AxonPath Myelinated Nerve Fibre Segmentation

The AxonPath datasets: electron microscopy (TEM and STEM) and brightfield images of myelinated nerve fibres from mouse and rat central and peripheral nervous tissue, in which every myelinated fibre in the annotated region is outlined together with its inner cylinder and its axon. They were made to train and evaluate the AxonPath segmentation models and ship as QuPath GeoJSON. Here every outline is a browsable polygon annotation, and each file carries its sub-dataset, species, tissue, modality and pixel size as metadata.

Contents

Images	47 source images (full-resolution raw TIFFs), hosted as 77 files: 27 large images split into 2 to 4 strips, see Conversion notes
Annotations	Source: 9,899 fibres (25,025 objects). Hosted: 25,765 polygons (Fibre 10,201, InnerCylinder 5,457, Axon 10,025, Nerve 8, ROI 74), higher only because objects crossing strip borders were clipped into pieces
Classes	5: Fibre (whole myelinated fibre), InnerCylinder (axon plus inner tongue, EM only), Axon, Nerve (nerve outline, brightfield only), ROI (the annotated region; fibres were outlined only inside it)
Dataset	axonpath-myelinated-fibres-v1, one folder per sub-dataset under images/
Split	None in the source (all 7 sub-datasets were released for training and evaluation)
Imaging	TEM, STEM and brightfield; mouse and rat; 8-bit, except corpus callosum (32-bit float)

Sub-dataset	Tissue	Species	Modality	Images (hosted files)	Fibres	Pixel size (µm)
CorpusCallosum_Mus_TEM_000	Corpus callosum (CNS)	Mouse	TEM	20 (40)	2,074	0.00175–0.00211
Fimbria_Rat_STEM_000	Fimbria (CNS)	Rat	STEM	1 (2)	420	0.0061
OpticNerve_Mus_STEM_000	Optic nerve (CNS)	Mouse	STEM	1 (2)	433	0.0061
OpticNerve_Mus_TEM_000	Optic nerve (CNS)	Mouse	TEM	10 (10)	1,084	0.0043
SciaticNerve_Mus_STEM_000	Sciatic nerve (PNS)	Mouse	STEM	3 (6)	470	0.01519
SciaticNerve_Mus_TEM_000	Sciatic nerve (PNS)	Mouse	TEM	9 (9)	921	0.01704–0.01708
TibialNerve_Mus_BF_000	Tibial nerve (PNS)	Mouse	Brightfield	3 (8)	4,497	0.07142

Fibre counts are those of the source annotations, and they match the record exactly. In brightfield the inner tongue cannot be resolved, so there is no InnerCylinder and the Axon includes the inner tongue.

Using it

Browse any image to see the fibre, inner cylinder and axon outlines. File metadata gives subdataset, species, tissue, system, modality, pixel_size_um, source_image and tile_offset_px. To download everything (free account required):

• ZIP: use the Download button on this project.
• Python SDK: pip install datatorch, then:

datatorch login
datatorch pull datatorch/axonpath-myelinated-fibres

Annotations export in COCO format through the project's export schema.

Conversion notes

• Source and checks. All 7 zips and the 4 tables were downloaded from Zenodo and matched the published md5. Every image decodes, there are no byte-identical duplicates, every image has its GeoJSON, and the per-sub-dataset image and fibre counts match the record. The full list is in errata.txt in this project.
• Decode method. Only raw/ is hosted: the full-resolution TIFFs and their QuPath GeoJSON, in full-resolution pixel coordinates. The authors' training/ exports (downsampled ROI crops with semantic masks and instance labels) are not included; get them from the Zenodo record. GeoJSON classes map to labels of the same name, and the unclassified object (the annotated region) becomes ROI. MultiPolygons (49 objects) become one annotation with several polygons. 6 interior rings (holes) were dropped, because COCO polygons cannot hold them. Vertices were reduced losslessly (only collinear points on the pixel staircase were removed).
• Strips. DataTorch's upload limit is about 40 MiB per file, so the 27 images whose pixel data exceeds 32 MiB were cut into the fewest equal, non-overlapping strips along the long axis: 2 per image for the corpus callosum and STEM images, 2 or 4 for brightfield. Pixel values, bit depth (including 32-bit float) and calibration are unchanged; no resampling or compression. Objects crossing a strip border were clipped (604 fibre pieces, 1,481 pieces in all), and 9 zero-area slivers on borders were dropped. A strip's name gives its pixel offset in the source image (e.g. A65A_tile2of4_x3051_y0.tif), which is also in its metadata, so full-resolution coordinates can be rebuilt. Strips outside the annotated ROI carry no annotations.
• Discrepancies. (1) One Axon in OpticNerve_Mus_STEM_000/M06G02 has no surrounding fibre, as the authors note. (2) Per-class counts differ from the fibre count: some fibres have no inner cylinder or axon object and a few have two. (3) The ROI of SciaticNerve_Mus_TEM_000/L51L1_12 extends 1 px above the frame and was clipped.
• Subset. None: all 7 sub-datasets and all 47 images.

License and citation

CC BY 4.0, for all 7 sub-datasets. The record, its README and every row of its SOURCES.csv give CC BY 4.0, and 8 of the corpus callosum images come from the AimSeg ground truth, also CC BY 4.0. You may share and adapt the data, including commercially, with attribution. 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

A preprint describing AxonPath is in preparation; until it appears the authors ask that this record be cited. For the AimSeg images, see also Carrillo-Barberà et al., PLoS Comput Biol (2023).

Source: Zenodo record 23084690.

Related

• More microscopy datasets on DataTorch
• All public datasets