Cell Segmentation (BBBC007)

Two-channel segmentation of Drosophila melanogaster Kc167 cells, mirrored from the Broad Bioimage Benchmark Collection (BBBC007v1) with every hand-drawn nucleus and cell outline converted to a browsable polygon annotation. Cells were stained for DNA (nuclei) and actin (cell body); the original release ships the two channels as separate grayscale TIFFs plus matching outline rasters. Here each outline is a polygon you can inspect in the annotator, filter, and export. The benchmark task is finding the boundaries between touching cells, with the nuclei available as seeds.

Contents

Images	32 (16 fields of view, one DNA image and one actin image each)
Annotations	2,485 polygons: 1,264 nucleus on the DNA images, 1,221 cell on the actin images
Classes	2 (nucleus, cell)
Dataset	bbbc007-v1 (the complete release; the source has no train/test split)
Split	None. Images come from four acquisition batches (folders A9, f113, f96 (17), f9620 in the source), which is the natural grouping for a leave-one-batch-out split
Imaging	Widefield fluorescence, Zeiss Axioplan 2 with Axiocam MRm; 8-bit grayscale, 400x400, 450x450 and 512x512 px; cells about 25 px across

Channel pairing follows the source filenames: ..._D_1UL / ...d0 / A9 p*d are DNA and ..._F_2UL / ...d1 / A9 p*f are actin. Everything else in the two names is identical.

Using it

Browse any DNA image to see the nucleus polygons and any actin image to see the cell 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/cell-segmentation-bbbc007

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

Conversion notes

• Images. The source TIFFs were written losslessly to PNG, one per channel, with spaces in filenames replaced by underscores (A9 p10d.tif becomes A9_p10d.png). 12 of the 32 TIFFs are stored as RGB with a faint tint on about 150 to 200 fixed pixels (fiducial marks); those were reduced to grayscale by taking the per-pixel maximum over channels, which equals the dominant channel everywhere else.
• Outlines. The source outline rasters are monochrome with value 0 on outline pixels and 255 elsewhere (not the reverse). An object was taken as a 4-connected region of non-outline pixels that is fully enclosed by outline pixels and does not touch the image border. Its polygon is the external contour of that region dilated by one pixel, so the hand-drawn outline ring belongs to the object. Polygons are not simplified.
• Nuclei. The DNA-channel rasters yielded 1,301 enclosed regions; 37 of 1 to 9 pixels (slivers between neighbouring outlines) were dropped, leaving 1,264 nucleus polygons. No nucleus outline touches an image border.
• Cells. The actin-channel rasters yielded 1,515 enclosed regions. 160 slivers under 10 pixels were dropped. A further 134 enclosed regions contain no outlined nucleus; inspection shows these are predominantly dark background pockets ringed by cells, so following the source protocol (one cell was outlined per nucleus) they are not exported. That leaves 1,221 cell polygons.
• Border cells are missing. The annotators did not close the cell outlines along the image edge, so clipped cells at the border merge with the background in the raster and cannot be recovered as individual objects. 121 such border-touching regions were excluded. As a result 38 nucleus polygons have no enclosing cell polygon, and 1,221 cells is fewer than the roughly 80 cells per image the source quotes. The source metric ignores boundary pixels adjacent to background, so this does not change the benchmark definition.
• Image filenames match the original release (up to the TIFF to PNG extension and the space to underscore change), so results remain comparable with published work.

License and citation

CC0 1.0 (public domain). Anne Carpenter waived copyright and related rights to the images and ground truth; attribution is still good practice. If you use this data, please cite:

Jones, T.R., Carpenter, A.E., Golland, P. Voronoi-based segmentation of cells on image manifolds. Proc. ICCV Workshop on Computer Vision for Biomedical Image Applications (CVBIA), 2005.

Ljosa, V., Sokolnicki, K.L., Carpenter, A.E. Annotated high-throughput microscopy image sets for validation. Nature Methods 9, 637 (2012).

Source: BBBC007v1, Broad Bioimage Benchmark Collection. Images courtesy of the laboratory of David Sabatini, Whitehead Institute for Biomedical Research; outlines by Chris Gang.

Related

• Nuclei Segmentation (DSB2018) and Nuclei Segmentation (BBBC039): nuclei-only companions to this two-channel set
• More microscopy datasets on DataTorch
• All public datasets