Deep learning-assisted Raman spectroscopy for rapid lactic acid bacteria identification at the colony level.

To achieve rapid and accurate identification at the colony level and improve the efficiency of colony selection, we proposed an adaptive colony Raman acquisition method based on signal-to-noise ratio screening (ACRA-SNR). This method enables in situ spectral acquisition of colonies, effectively mitigating the impact of spatial heterogeneity within colonies on classification and identification. By combining this acquisition method with the Raman Swin Transformer (Ra-ST) model, we achieved fast and accurate classification and identification of colonies. In this work, we analyzed the spectral data of fourteen lactic acid bacteria (LAB) strains, and the Ra-ST model achieved a classification accuracy of 98.2 %. To evaluate the predictive performance of the model, we used it to identify two LAB strains from the same species as those in the classification model but from different sources. The results showed identification accuracy above 70 %, demonstrating good generalization. Additionally, we conducted a comparative analysis between the Ra-ST model and other models, which demonstrated that the Ra-ST model outperformed the others in both classification and prediction performance. Therefore, the combination of the ACRA-SNR technique and the Ra-ST model is expected to facilitate the accurate classification and identification of LAB and other functional bacteria in industrial production, thereby significantly enhancing production efficiency and output.