Deep-Fish-Bone zebrafish larval skeleton segmentation

Deep-Fish-Bone is a brightfield microscopy benchmark for multi-label segmentation of the developing craniovertebral skeleton in zebrafish larvae. It is mirrored from its Zenodo record (V1.0, doi:10.5281/zenodo.23125004), with every structure mask converted to a polygon annotation you can browse. The original release ships one binary mask PNG per structure per record. Here each structure is a labelled polygon you can inspect in the annotator, filter by genotype or fish, and export. Code and the authors' evaluation splits are on GitHub.

Contents

Images	190 annotation records from 34 fish, ventral view, 2576 x 1932 px 8-bit RGB PNG
Annotations	4,395 skeletal-structure polygon annotations, plus 190 full-body outlines (full_mask)
Classes	25 structures: P, Op1, Op2, Oc1, Oc2, N, M1, M2, Hm1, Hm2, En1, En2, D1, D2, Cl1, Cl2, Ch1, Ch2, Cb1, Cb2, Br2a, Br2b, Br1a, Br1b, VC, plus the supporting full_mask
Genotypes	47 WT, 110 HET, 33 HOM records (9, 20 and 5 fish)
Dataset	deep-fish-bone-zebrafish-v1
Split	None in the archive. Each file carries cv_base_fold (0-4) from the authors' fish-grouped 5-fold CV: test = fold i, validation = fold i+1
Imaging	Brightfield microscopy of stained zebrafish larvae, ventral view

Each image's file metadata holds annotation_record_id, fish_id, genotype, quality (1-3), experiment_code, tif_filename, cv_base_fold and view. For the 68 image pairs that two annotators labelled independently, it also holds identical_image_as_record. Files live at images/ventral/<record id>/<record id>.png. Split by fish_id, never by image: several records come from the same fish, and 136 records are double annotations of 68 identical images.

Using it

Browse any image to see the structure polygons. 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/deep-fish-bone-zebrafish

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

Conversion notes

Every mask was read as foreground where its value is above 0. Each connected component's outer contour became a polygon, simplified to 1 pixel (OpenCV approxPolyDP). All components of one structure in one record form a single annotation, so VC has 914 polygons, one per vertebra. Holes are not represented. Rasterised back, the polygons match the source masks at mean IoU 0.982 (minimum 0.893). Labels use the README's canonical names: the files EN1/CL1/CH1/CB1 (and their 2 partners) are En1/Cl1/Ch1/Cb1 here. All 190 records were kept, nothing was subset, and the image pixels are unchanged.

Every file was checked, and the full list is in errata.txt in this project. The main points:

• Empty masks: 355 of 4,750 structure masks are empty, meaning the structure was not annotated as present, so they produce no annotation. Br2a and Br2b are present in 63 and 67 records, as documented. Missing elsewhere: Oc2 36, Oc1 32, M1 10, M2 5, Ch1 5, Ch2 5, Cb2 4, Cb1 3, Cl1 3, Cl2 2.
• Duplicates: 68 byte-identical image pairs. These are intentional double annotations, and each pair stays within one fish and one CV fold. Where the two annotators' masks differ, they agree at a median IoU of 0.69.
• Probable copy errors: three records have one non-empty mask filed under two structures (163428646 Ch2=En2, 163434162 Br2a=En1, 163434207 Br1b=En2). They are imported as distributed.
• Burned-in genotype text: 50 images show the genotype in red in the lower-left corner, for example "(+/-)". Crop or mask that corner before training a genotype classifier. Most images also have a 500 µm scale bar.
• Encoding: two masks are RGBA rather than 0/255 (163430975 Oc1 stores values 1-5; 163446505 Br2a is anti-aliased). The above-0 threshold keeps both.
• No corrupt files, orphan masks or records without metadata.

License and citation

CC BY 4.0. You may share and adapt the data for any purpose with attribution, a link to the license, and an indication of changes. Here the changes are the mask-to-polygon conversion and renamed image files. If you use this data, please cite:

Kumar, N., Raman, R., Muller, M., Maree, R., & Geurts, P. (2026). Deep-Fish-Bone: Brightfield Microscopy Dataset for Multi-Label Segmentation of Craniovertebral Skeletal Structures in Zebrafish Larvae (Version V1.0) [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.23125004

The README also asks you to cite the accompanying manuscript, whose publication details were not yet available: Kumar, Raman, Muller, Maree and Geurts, Deep Learning-Based Multi-Label Segmentation of Developing Craniovertebral Skeletal Structures in Zebrafish Larvae: A Brightfield Microscopy Benchmark Dataset and Systematic Evaluation.

Source: Zenodo record 23125004.

Related

• More microscopy datasets on DataTorch
• All public datasets