Accurate Identification of Bacteria Using a Spectral Transformer Machine Learning Model for Hyperspectral Raman Images

In recent years Raman spectroscopy (RS) has become a promising tool for bacterial identification and antimicrobial resistance testing [1]. In this regard, multiple academic studies have shown excellent results, demonstrating high classification accuracy on multiple bacterial species [1]–[2]. These studies have typically applied complex, and time-consuming sample preparation, making their approaches unsuitable for clinical implementation. It has been a challenging task in realizing clinically relevant accuracies for bacteria identification using RS. Recent efforts involving machine-learning (ML) have demonstrated a large potential to reach high accuracies [1]–[2]. However, to bridge the gap between academic proof-of-concept and clinical application and for the purpose of introducing ML to data-limited domains. We propose a new ML transformer model, the spectral transformer (ST) [3]–[4]. We show that the ST can be used for fast and accurate classification of hyperspectral Raman images of bacteria. We Specifically demonstrate that our ST model can outperform state-of-the-art convolutional neural networks (CNNs) in scientific domains with limited availability of data [3]–[4]. We demonstrate the abilities of our ST ML model on Raman hyper-spectral images from 20 classes of reference bacteria as shown in figure 1. To further test the performance of the ST ML model in vivo we tested the concept on Raman data recorded on experimentally infected wounds in mice, with the aim to identify the species of the infecting organism.