AxonPath datasets: annotated electron microscopy and brightfield images of myelinated nerve fibres

Annotated electron microscopy (EM) and brightfield images of myelinated nerve fibres, used to train and evaluate the AxonPath segmentation models.

Every myelinated fibre in the annotated region of each image is outlined, together with its inner cylinder (the axon plus the inner tongue of the myelin sheath) and its axon. The data are provided both as full-resolution images with their annotations, and as the training exports used for the released models.

Datasets

47 images and 9899 annotated fibres in total. TEM: transmission electron microscopy; STEM: scanning transmission electron microscopy. The origin of each dataset is given in SOURCES.csv.

Files

One zip per dataset, plus three tables:

• <Dataset>.zip
  • <Dataset>/raw/images/<image>.tif: full-resolution image
  • <Dataset>/raw/annotations/<image>.geojson: annotations (QuPath GeoJSON)
  • <Dataset>/training/images/<image>_roi0.tif: training export, image
  • <Dataset>/training/masks/<image>_roi0.tif: training export, semantic mask
  • <Dataset>/training/labels/<image>_roi0.tif: training export, fibre instance labels
• datasets.csv: one row per dataset: tissue, species, modality, downsample, nominal pixel size
• pixel_sizes.csv: one row per image: pixel size of the raw image and of its training export
• SOURCES.csv: origin and license of each dataset

Annotations (raw/annotations/)

GeoJSON exported from QuPath, in the pixel coordinates of the full-resolution image. Objects are classified as:

In brightfield images the inner tongue cannot be resolved, so there are no InnerCylinder objects and the Axon includes the inner tongue.

Training exports (training/)

The annotated region (ROI) of each image, downsampled by the dataset's factor, as exported for training with the AxonPath QuPath extension. Because only the ROI is exported, most training images are crops of the raw image; a few fibres annotated outside the ROI (sciatic nerve TEM) are not included.

• images/: the image, with the same bit depth as the raw image (8-bit, except CorpusCallosum_Mus_TEM_000: 32-bit float).
• masks/: semantic mask, 0 background, 1 myelin, 2 inner tongue, 3 axon. Classes are drawn in the order fibre, inner cylinder, axon, so 2 is the part of the inner cylinder not covered by the axon. For training, 2 and 3 together form the inner cylinder. Brightfield masks contain 0, 1 and 3 only.
• labels/: fibre instance labels, one integer per fibre, 0 background (8- or 16-bit, depending on the number of fibres).

Pixel sizes

pixel_sizes.csv gives the pixel size of every raw image, read from its TIFF calibration, and the effective pixel size of its training export (raw pixel size × downsample). The calibration can differ slightly between images of the same dataset; the pixel_size_um column of datasets.csv is a nominal, per-dataset value. The TIFF files carry their own calibration in µm.

Changes from the original files

• File names containing # were renamed (# → -, e.g. P#5 → P-5), for use in URLs.
• ImageJ metadata fields holding local file paths and original file names were removed from the TIFFs. Pixel data and calibration are unchanged.

Known issues

• OpticNerve_Mus_STEM_000/M06G02: one axon (848 pixels in the training export) has no surrounding fibre annotation. It is kept as is, because this is the data used to train and evaluate the released models, and will be corrected in a future version.

Related records

• AxonPath segmentation models: 10.5281/zenodo.22982902
• AxonPath training code: github.com/paucabar/axonpath
• AxonPath QuPath extension: github.com/paucabar/qupath-extension-axonpath
• AimSeg ground truth and classifiers: 10.5281/zenodo.8351731 (8 of the corpus callosum images, 4× downsampled; see SOURCES.csv)
• AimSeg paper: 10.1371/journal.pcbi.1010845

License

Creative Commons Attribution 4.0 International (CC BY 4.0).

Citation

A preprint describing AxonPath is in preparation. Until then, please cite this record.