Enhancing robustness and generalization in microbiological few-shot detection through synthetic data generation and...

Area of Science:

• Biotechnology
• Machine Learning
• Microbiology

Background:

• Continuous hygiene monitoring in medical and pharmaceutical processes relies on manual microorganism detection.
• Automating colony detection with deep learning is hindered by a lack of sufficient training data.

Purpose of the Study:

• To develop a novel pipeline combining generative data augmentation and few-shot detection for improved colony detection performance with limited data.
• To address the challenge of insufficient training data in automated microorganism detection.

Main Methods:

• Utilized a diffusion-based generator model for inpainting synthetic bacterial colonies onto real agar plate backgrounds.
• Implemented a decoupled feature classification strategy with a feed-forward network for class-agnostic detection and lightweight classification.
• Trained the model effectively using only 25 real images.

Main Results:

• Achieved a +0.45 mAP improvement over training from scratch and a +0.15 mAP advantage over current state-of-the-art synthetic data augmentation.
• Obtained an AP50 score of 0.7 in a few-shot scenario on the AGAR dataset.
• Demonstrated robustness to image corruptions like noise and blur.

Conclusions:

• The proposed pipeline offers a scalable and efficient solution for colony detection, reducing the need for large labeled datasets.
• The method is applicable to real-world scenarios in hygiene monitoring and biomedical research.
• Potential for broader applications in detecting new colony types rapidly.