SemiRaman: A self-supervised contrastive representation learning-based framework for semi-supervised Raman spectral...

Area of Science:

• Microbiology
• Spectroscopy
• Artificial Intelligence

Background:

• Raman spectroscopy is a key technique for identifying pathogenic bacteria.
• Current AI methods struggle with limited labeled spectral data and noisy spectra.
• Rapid detection is crucial for public health threats and biological contamination.

Purpose of the Study:

• To develop an efficient self-supervised framework (SemiRaman) for semi-supervised Raman spectral identification of pathogenic bacteria.
• To overcome data scarcity and spectral challenges in deep learning approaches.
• To achieve high accuracy with minimal labeled data.

Main Methods:

• Proposed SemiRaman framework combining unsupervised and supervised learning.
• Utilized redundancy reduction and multi-level contrastive learning on unlabeled data.
• Employed a multi-stage fine-tuning strategy with limited labeled data.

Main Results:

• Achieved 89.2% accuracy and 89.1% MF1-score on Bacteria-7 with 5% labeled data.
• Reached 92.0% accuracy and 91.0% MF1-score on Bacteria-14 with 10% labeled data.
• Demonstrated superior and stable performance compared to baseline methods, even with noise and high diversity.

Conclusions:

• SemiRaman offers a cost-effective and accurate solution for rapid pathogenic bacteria identification.
• The framework excels in semi-supervised settings with extremely limited labeled data.
• Enables reliable microbial detection for food safety, environmental monitoring, and public health.