food-ai-nexus/microcolony-domain-adaptation

Microcolony Domain Adaptation (Foodborne Bacteria) is a microscopy image dataset for foodborne bacterial classification under varying imaging conditions. It was created to support research in adversarial domain adaptation, enabling models trained on standard phase contrast microscopy images to generalize across different optical configurations and biological conditions.

This dataset accompanies the publication: Bhattacharya, S., Wasit, A., Earles, M., Nitin, N., & Yi, J. (2025). Enhancing AI microscopy for foodborne bacterial classification using adversarial domain adaptation to address optical and biological variability. Frontiers in Artificial Intelligence, 8, 1632344. doi: 10.3389/frai.2025.1632344

Content

The dataset contains microscopy images of six foodborne bacterial species imaged under a source domain (standard conditions) and multiple target domains (varying optical and biological conditions). It is structured into four splits to support both standard supervised training and domain adaptation experiments.

Split	Description	Images
train	Standard phase contrast images (60x, 3h), with class subdirectories	539
train_fewshot	Small labeled samples from target domains for few-shot adaptation	150
test_standard	Held-out standard phase contrast images (same domain as train)	90
test_ood	Out-of-distribution images under varying imaging conditions	420

Classes

Code	Species	Full Name
Bc	Bacillus coagulans	Gram-positive spore-forming bacterium
Bs	Bacillus subtilis	Gram-positive model organism
Ec	Escherichia coli 1612	Gram-negative foodborne pathogen
Li	Listeria innocua	Non-pathogenic Listeria surrogate
SE	Salmonella enterica Enteritidis	Foodborne pathogen
ST	Salmonella enterica Typhimurium	Foodborne pathogen

Imaging Conditions

Domain	Objective	Incubation	Modality	Split
phase_contrast_60x_3h	60x	3 h	Phase contrast	train, train_fewshot, test_standard
20x-3h	20x	3 h	Phase contrast	train_fewshot, test_ood
20x-5h	20x	5 h	Phase contrast	train_fewshot, test_ood
20x	20x	3 h	Phase contrast	test_ood
brightfield	60x	3 h	Brightfield	train_fewshot, test_ood
defocus	60x	3 h	Phase contrast (defocused)	train_fewshot, test_ood
agar15	60x	3 h	Phase contrast (1.5% agar)	test_ood

Uses

This dataset is intended for:

• Image classification of foodborne bacterial microcolonies.
• Domain adaptation research, where models trained on the source domain (phase_contrast_60x_3h) are evaluated on target domains.
• Few-shot learning experiments using the train_fewshot split.

from datasets import load_dataset

# Load all splits
ds = load_dataset("food-ai-nexus/microcolony-domain-adaptation")

# Load only the standard train/test splits
train = ds["train"]
test_std = ds["test_standard"]
test_ood = ds["test_ood"]

License

This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) License.

Citation

@article{bhattacharya2025microcolony,
  title     = {Enhancing AI microscopy for foodborne bacterial classification using adversarial domain adaptation to address optical and biological variability},
  author    = {Bhattacharya, Siddhartha and Wasit, Aarham and Earles, J. Mason and Nitin, Nitin and Yi, Jiyoon},
  journal   = {Frontiers in Artificial Intelligence},
  volume    = {8},
  pages     = {1632344},
  year      = {2025},
  doi       = {10.3389/frai.2025.1632344}
}

Source

Original dataset: Zenodo 10.5281/zenodo.16741157
Code repository: GitHub food-ai-engineering-lab/microcolony-domain-adaptation-frai