Rapid and accurate bacteria identification through deep-learning-based two-dimensional Raman spectroscopy.

Area of Science:

• Spectroscopy
• Biosensing
• Artificial Intelligence

Background:

• Surface-enhanced Raman spectroscopy (SERS) provides molecular fingerprints for applications in medicine and biology.
• AI-enabled Raman spectroscopy enhances bacteria identification, but often faces a trade-off between accuracy and processing time.

Purpose of the Study:

• To develop an efficient bacteria identification strategy combining deep learning with spectrogram encoding.
• To overcome the limitations of high-resolution spectrograms and reduce data processing time.

Main Methods:

• Utilized wavelet packet transform for spectral compression (1/15th) and Gramian angular field techniques to amplify subtle spectral differences.
• Integrated deep learning models for enhanced bacteria identification from SERS data.

Main Results:

• Achieved 99.64% identification accuracy for two bacterial types and 90.55% for thirty types.
• Demonstrated a 90% reduction in training time compared to conventional methods.
• Showcased model stability and generalization ability with superimposed Gaussian noise.

Conclusions:

• The developed algorithm offers an efficient approach for accurate and rapid bacteria identification.
• This method has potential for on-site testing, is updatable, and contributes to spectroscopy understanding.
• Paves the way for advancements in environmental monitoring, food safety, and public health diagnostics.