FluoEcoli-Instance (test split)

High-content fluorescence microscopy of Escherichia coli for instance segmentation, mirrored from the FluoEcoli-Instance release on Zenodo with every cell as a browsable polygon annotation. This project holds the test split of the 24,702-image set; the train (14,820 images) and val (4,941 images) splits can follow.

Contents

Images	4,941 (the official test split)
Annotations	39,472 cell polygons (1 to 91 per image, median 6)
Classes	1 (cell)
Dataset	fluoecoli-instance-v1
Split	test only, files under images/test/
Imaging	Fluorescence microscopy, 256 × 256 px patches cut from full-size images; single-channel 16-bit grayscale TIFF

The full-size microscopy images were assigned to train, val and test before being cut into patches, so every patch from one image is in the same split and the splits share no identical or near-duplicate images.

The TIFFs are raw 16-bit intensities (a typical patch's 1st to 99th percentile runs from about 1,900 to 15,300, out of 65,535), so they look nearly black in viewers that don't rescale. Stretch the contrast (for example, 1st to 99.5th percentile) before viewing or training on 8-bit inputs.

Using it

Browse any image to see the per-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/fluoecoli-instance

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

Conversion notes

Images are the original TIFFs, unchanged, with the original file names. Polygons come straight from the release's coco/test/test.json: one polygon per cell, imported as-is with DataTorch's COCO importer. The source uses category_id 0 for cell, so for the import the category was remapped to 1 and image and annotation ids were shifted by +1. Geometry and file names are untouched. Subset rule: only coco/test/images/*.tiff and coco/test/test.json from the archive (md5 7014ab9749f6d96ae8dcf22efca86db0).

Integrity checks against the record all passed: 4,941 images and 39,472 instances, no undecodable files, no byte-identical duplicates, no orphan annotations or images, and no polygons outside the frame. Discrepancies, all listed in errata.txt in this project:

• 24 annotations are degenerate (zero area or a zero-width box, mostly 1 to 3 px slivers at patch edges). They were kept.
• 418 annotations have a COCO bbox 1 to 18 px smaller than their polygon, almost all at the frame edge. Only polygons are imported here, so this copy isn't affected.
• In a visual spot check, some visible cells have no polygon. This may be the authors' labelling protocol; the record does not say.
• Train, val and test reuse the same file names (000000.tiff, ...). Keep the split folder in mind when combining them.

License and citation

CC BY 4.0 (Creative Commons Attribution 4.0 International). Reuse and redistribution are allowed with credit. If you use this data, please cite:

Huynh, S. T., Nguyen, H. X., Duong, D. Q., Nguyen, T. K., Nguyen, N. K. P., Nguyen, H. T. L., Vu, P. N., Nguyen, N. P. G., Vu, K. D., Ho, H. Q. N., Tran, T.-A., Baker, S., Nguyen, B. T. FluoEcoli-Instance: a high-content fluorescence microscopy dataset and CBAM-enhanced hybrid task cascade for E. coli instance segmentation. IEEE Journal of Selected Topics in Signal Processing 19(8), 1967-1982 (2025). doi:10.1109/JSTSP.2025.3648286

Source: FluoEcoli-Instance on Zenodo (DOI 10.5281/zenodo.23389427). Training code: huynhthanh98/jove_mmdetection.

Related

• More microscopy datasets on DataTorch
• All public datasets