C. elegans Worm Segmentation (BBBC010)

The C. elegans live/dead assay from the Broad Bioimage Benchmark Collection (BBBC010v1), with every hand-corrected per-worm mask converted to a browsable polygon annotation. The images are control wells from a screen for anti-infective compounds: worms were infected with Enterococcus faecalis and either left untreated (mostly dead: rod-like, uneven texture) or treated with ampicillin (mostly alive: curved, smooth). Each well was imaged in brightfield and in a GFP channel. The benchmark task is instance segmentation of worms that touch, overlap and cluster, which is why the original release ships one binary mask per worm rather than one label image per field (Wählby et al., 2012).

Contents

Images	200 (100 wells, one brightfield and one GFP image each; 696x520, 16-bit TIFF)
Annotations	2,814 worm polygons: the 1,407 per-worm masks, each placed on both channel images of its well (1,536 worm-ampicillin, 1,278 worm-untreated)
Classes	2 (worm-ampicillin, worm-untreated): the well's treatment from the plate layout, not a per-worm live/dead call
Dataset	worm-segmentation-bbbc010-v1 (the complete release)
Split	None official. 52 wells in columns 1-12 are ampicillin-treated positive controls and 48 wells in columns 13-24 are untreated negative controls; rows A-E give a natural 5-fold grouping. The whole set is one plate imaged in one session
Imaging	Discovery-1 automated microscope (Molecular Devices) at MGH, 384-well plate; channel w1 brightfield, channel w2 GFP; 11 to 15 worms per well

Channel pairing follows the source filenames: ..._<well>_w1_<uuid>.tif is brightfield and ..._<well>_w2_<uuid>.tif is GFP for the same well.

Using it

Browse any image to see the per-worm polygons; the brightfield and GFP images of a well carry identical 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/worm-segmentation-bbbc010

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

Conversion notes

• Images. The 200 TIFFs are the original files from BBBC010_v2_images.zip (the five __MACOSX resource-fork entries in the archive were discarded). 199 are stored uncompressed as shipped; one (A24_w2) was PackBits-compressed and was rewritten losslessly to an uncompressed 16-bit TIFF (pixel-identical, verified). Pixel values use a 12-bit range (maximum 4095), so viewers that assume full 16-bit scaling show them dark until auto-levelled.
• Polygons. Each of the 1,407 per-worm masks (<well>_<nn>_ground_truth.png, binary 0/255) was traced on a 4x nearest-neighbour upsampling so the polygon follows pixel edges in COCO coordinates (pixel x spans [x, x+1]). Outer contours and hole contours are both kept as paths of the one annotation, following the convention of DataTorch's own mask converter; 32 masks (6,373 pixels in total) contain holes where a worm curls around background. Re-rasterizing the polygons with even-odd fill reproduces every source mask with IoU 0.996 or better (mean 1.000). Polygons are not simplified.
• Fragments and specks. 10 masks consist of 2 to 4 disconnected pieces (worm parts separated by an overlapping neighbour); they are kept as multi-path annotations. 11 sub-pixel specks (area under one pixel after tracing) were dropped. No mask was empty.
• Mask numbering. Worm indices in the source filenames are not contiguous (for example A09 has no _05, and four wells start at _02). The 1,407 figure is the number of mask files actually present, which is 11 to 15 per well.
• Whole-field foreground masks. The release also ships one human-corrected foreground/background image per well (BBBC010_v1_foreground.zip). They are not imported because the union of the per-worm masks reproduces them exactly in 96 wells, after accounting for a consistent one-row vertical offset: the per-worm masks align with the GFP image to the pixel in all 100 wells, while the foreground masks sit one row lower. The per-worm geometry was used unchanged. In 4 wells (A09, B11, C02, C05) the two differ by 461 to 667 pixels, less than one worm.
• Classes. worm-ampicillin and worm-untreated encode the plate layout stated on the source page (columns 1-12 treated, 13-24 untreated). They say which control the well belongs to, not whether that particular worm is alive; both wells contain some of each phenotype.
• Image filenames match the original release, so results remain comparable with published work.

License and citation

CC0 1.0 (public domain). The contributors waived copyright and related rights; attribution is still good practice. The source asks that you cite it as:

"We used the C. elegans infection live/dead image set version 1 provided by Fred Ausubel and available from the Broad Bioimage Benchmark Collection [Ljosa, V., Sokolnicki, K.L., Carpenter, A.E. Annotated high-throughput microscopy image sets for validation. Nature Methods 9, 637 (2012)]."

The per-worm ground truth and the published benchmark results come from:

Wählby, C., Kamentsky, L., Liu, Z.H. et al. An image analysis toolbox for high-throughput C. elegans assays. Nature Methods 9, 714-716 (2012).

The underlying screen is described in Moy, T.I. et al., High-throughput screen for novel antimicrobials using a whole animal infection model. ACS Chemical Biology 4, 527-533 (2009).

Source: BBBC010v1, Broad Bioimage Benchmark Collection. Images originally acquired for a screen in Fred Ausubel's lab at MGH.

Related

• Cell Segmentation (BBBC007): another two-channel BBBC set with touching objects
• Nuclei Segmentation (BBBC039) and Nuclei Segmentation (DSB2018): nucleus instance segmentation companions
• More microscopy datasets on DataTorch
• All public datasets