A Lightweight Convolutional Neural Network for Bacterial Identification Based on Raman Spectra

Convolutional neural networks (CNN) combined Raman spectroscopy allows for fast and accurate identification of bacterial species. The current CNN, however, needs a complicated hyperparameters model design. Here, we propose a simple CNN network, namely Faster RamanNet, for quickly and precisely classifying bacteria at the species level using their Raman spectra. On the Bacteria-ID Raman spectral dataset and the PKU-bacterial Raman spectral datasets, we showed that the Faster RamanNet achieved comparable results when compared to the prior CNN models, although utilizing just roughly 1/45 and 1/297 network parameters, respectively. Faster RamanNet achieved an average isolate-level accuracy of 85.2±0.3%, antibiotic treatment identification accuracy of 97.0±0.3%, and accuracy of 95.44% on the PKU-bacterial dataset, respectively. Faster RamanNet model can be quickly trained even using CPU because its fewer model parameters. In this way, our method can quickly and accurately distinguish bacterial species in light of their Raman spectra and can be effectively reached out to other classification tasks.