What it is
Bacteria are small, rod-shaped and packed together, so a fluorescence field of E. coli is mostly cells touching cells. FluoEcoli-Instance was built for that problem: every cell in every patch is outlined separately, so a model can learn where one cell ends and the next begins. The authors released it with a paper that trains a Hybrid Task Cascade model with CBAM attention on it.
| Fact | Value |
|---|---|
| Record | Zenodo 23389427, version 1, published 2026-10-15 |
| Imaging | high-content fluorescence microscopy (the record does not name the microscope) |
| Patches | 24,702 TIFF, 256 x 256 pixels |
| Instances | 199,419 cells, expert-annotated |
| Classes | one, "cell" (category_id 0) |
| Format | COCO JSON per split, with polygon, bounding box and area per cell |
| Split | 14,820 train, 4,941 validation, 4,941 test |
| Size | about 2.4 GB, one zip |
| License | CC BY 4.0 |
All counts in this table are the record's own. The record says nothing more about how the images were acquired or how the annotation was done, so the paper is the place to read that.
Browse it in DataTorch
Dataset · datatorch/fluoecoli-instance · 4,943 files
FluoEcoli-Instance (test split)
FluoEcoli-Instance (test split): high-content fluorescence microscopy of Escherichia coli for instance segmentation. This project holds the test split of the 24,702-image FluoEcoli-Instance set: 4,941 image patches (256x256 px, single-channel 16-bit grayscale TIFF) with 39,472 expert-annotated cell instances as polygon annotations, one class (cell). The train (14,820 images) and val (4,941 images) splits can follow. License: CC BY 4.0. Source: https://zenodo.org/records/23389427 (DOI 10.5281/zenodo.23389427). Please cite: Huynh, S. T. et al. 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.
Open →
The hosted copy is the test split: 4,941 patches with 39,472 cell polygons in the class "cell". The files are single-channel 16-bit grayscale TIFFs, and DataTorch decodes them for display at 8 bits. The training and validation splits can follow in the same project.
Open any patch to see each outline, zoom into crowded clusters, and correct or add outlines with your team. Exports come out as COCO.
What people use it for
Training and benchmarking instance segmentation models for bacteria: counting cells, measuring their size and shape, and separating touching cells in dense fields. The test split works on its own as a held-out set for a model trained on other bacterial data, since none of its source images appear in train or validation.
Source and citation
Source: https://zenodo.org/records/23389427 (read 2026-11-10). Cite: Huynh, S. T. et al. 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. Dataset DOI: 10.5281/zenodo.23389427.
Outline your own cells
Label microscopy images with your team, review each other's work and export COCO, in the browser.
Frequently asked questions
What does FluoEcoli-Instance label?
Every E. coli cell in each patch is outlined as a polygon with a bounding box and area, all in one class, "cell". It is an instance segmentation dataset, so touching cells are separate objects.
How big is FluoEcoli-Instance?
24,702 TIFF patches of 256 x 256 pixels with 199,419 cell instances. Train has 14,820 patches and 119,976 cells, validation 4,941 and 39,971, test 4,941 and 39,472. The zip is about 2.4 GB.
Can patches from one image end up in both train and test?
No. According to the record, full-size images were assigned to a split before they were cut into patches, so every patch of an image stays in one split, and the splits share no identical or near-duplicate images.
Why does the COCO file start category ids at 0?
The dataset numbers its single class "cell" as category_id 0, where the original COCO dataset starts at 1. Check this if your training code assumes 1-based ids.
What is the license?
Creative Commons Attribution 4.0 International (CC BY 4.0). Cite the IEEE JSTSP paper and the Zenodo record, DOI 10.5281/zenodo.23389427.
Open the hosted copy on DataTorch.
Browse the files in your browser, then download the dataset with a free account.
