Enhancing AI microscopy for foodborne bacterial classification using adversarial domain adaptation to address optical...

Area of Science:

• Microbiology
• Computer Science
• Biotechnology

Background:

• AI-enabled microscopy offers rapid bacterial classification but struggles with imaging variability in dynamic or resource-limited environments.
• Existing AI models often lack generalizability when applied to data from different microscopy setups or conditions.

Purpose of the Study:

• To enhance the generalizability and robustness of AI-powered bacterial classification using domain adaptation techniques.
• To develop a scalable framework for AI microscopy applicable to decentralized and resource-limited settings.

Main Methods:

• Employed domain-adversarial neural networks (DANNs) and multi-DANNs (MDANNs) to address single and multiple domain variations, respectively.
• Utilized an EfficientNetV2 backbone with few-shot learning for fine-grained feature extraction and scalability in data-limited scenarios.
• Trained models on a source domain of six bacterial species under optimal conditions and tested on target domains with varied modalities, magnifications, and incubation times, using minimal labeled data.

Main Results:

• DANNs significantly improved bacterial classification accuracy in target domains by up to 54.5% (e.g., 20x magnification: 34.4% to 88.9%), with minimal performance loss in the source domain.
• MDANNs further boosted accuracy in the brightfield (BF) domain from 73.3% to 76.7%.
• Feature visualization techniques (Grad-CAM, t-SNE) confirmed the model's ability to learn domain-invariant features.

Conclusions:

• Domain adaptation techniques effectively enhance the generalizability of AI microscopy for bacterial classification across diverse imaging conditions.
• The developed framework offers a scalable and adaptable solution for bacterial identification, expanding the utility of AI microscopy in challenging, real-world settings.