Non-specific/specific SERS spectra concatenation for precise bacteria classifications with few samples using a residual neural network

Abstract

Deep learning neural network incorporating surface enhancement Raman Scattering technique (SERS) is becoming as a powerful tool for the precise classifications and diagnosis of bacterial infections. However, the large amount of sample requirement and time-consuming sample collection severely hinder its applications. We herein propose a spectral concatenation strategy for residual neural network using non-specific and specific SERS spectra for the training data augmentation, which is accessible to acquiring larger training dataset with same number of SERS spectra or same size of training dataset with fewer SERS spectra, compared with pure non-specific SERS spectra. With this strategy, the training loss exhibit rapid convergence, and an average accuracy up to 100% in bacteria classifications was achieved with 50 SERS spectra for each kind of bacterium; even reduced to 20 SERS spectra per kind of bacterium, classification accuracy is still > 95%, demonstrating marked advantage over the results without spectra concatenation. This method can markedly improve the classification accuracy under fewer samples and reduce the data collection workload, and can evidently enhance the performance when used in different machine learning models with high generalization ability. Therefore, this strategy is beneficial for rapid and accurate bacteria classifications with residual neural network.