RamanNet: a lightweight convolutional neural network for bacterial identification based on Raman spectra.

Area of Science:

• Microbiology
• Spectroscopy
• Machine Learning

Background:

• Raman spectroscopy combined with convolutional neural networks (CNNs) offers rapid bacterial identification.
• Existing CNN models for this task often involve complex hyperparameter design and high computational costs.

Purpose of the Study:

• To introduce RamanNet, a novel, simplified network architecture for bacterial species identification using Raman spectra.
• To reduce hyperparameter complexity and computational demands in bacterial Raman spectral analysis.

Main Methods:

• Developed RamanNet, a new, computationally efficient CNN architecture.
• Evaluated RamanNet on the Bacteria-ID and PKU-bacterial Raman spectral datasets.
• Compared RamanNet's performance and parameter count against existing CNN methods.

Main Results:

• RamanNet achieved comparable accuracy to previous CNN methods on both datasets.
• RamanNet utilizes significantly fewer network parameters (approx. 1/45 and 1/297).
• Achieved high accuracies: 84.7% isolate-level, 97.1% antibiotic treatment, 81.6% MRSA/MSSA differentiation on Bacteria-ID; 96.04% on PKU-bacterial dataset.

Conclusions:

• RamanNet provides a rapid, accurate, and computationally efficient solution for bacterial species identification via Raman spectra.
• The model's reduced parameter count allows for quick training, even on CPUs.
• RamanNet is adaptable for other Raman spectral classification tasks.