1. School of Biomedical Engineering （Suzhou）, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei 230026，China
2. Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou 215163, China
Abstract：Current methods for identifying pathogenic bacteria are time-consuming, leading to delays in optimal treatment and promoting antibiotic resistance. Therefore, developing a rapid, accurate, culture-free technique for this scenario has high clinical value. Raman spectroscopy can serve as a molecular fingerprint for rapid bacterial species identification, and computer-assisted classification is the current research hot spot. However, the classification methods based on machine learning and CNN in related works have poor generalization and insufficient feature mining ability, which leads to low classification accuracy. This study innovatively proposes a deep learning network named Raman Transformer (RaTR). RaTR can improve feature miningcapability and classification accuracy using kernel attention computation based on radial basis kernel function, and its model generalization is enhanced by introducing the diffusion process. Moreover, the discrete wavelet transform is proposed to address the excessive parameters and few-shot issues. Experimental validation on the Bacteria_ID and ATCC datasets shows that RaTR achieves classification accuracies of 85.83% and 84.73% respectively, demonstrating its accuracy and strong generalization. Visualizing key spectral features further confirms the effectiveness of feature extraction by the model. Finally, visualizing the spectral key features further confirms the effectiveness of RaTR feature extraction.
Key words：Raman spectroscopy; Deep learning; Classification of pathogenic bacteria
[1] Ikuta K S, Swetschinski L R, Aguilar G R, et al. The Lancet, 2022, 400(10369): 2221.
[2] Butler H J, Ashton L, Bird B, et al. Nature Protocols, 2016, 11(4): 664.
[3] Jones R R, Hooper D C, Zhang L, et al. Nanoscale Research Letters, 2019, 14: 231.
[4] Lussier F, Thibault V, Charron B, et al. TrAC Trends in Analytical Chemistry, 2020, 124: 115796.
[5] Nakar A, Pistiki A, Ryabchykov O, et al. Analytical and Bioanalytical Chemistry, 2022, 414(4): 1481.
[6] Ibtehaz N, Chowdhury M E H, Khandakar A, et al. Neural Computing and Applications, 2023, 35(25): 18719.
[7] Liu J, Osadchy M, Ashton L, et al. Analyst, 2017, 142(21): 4067.
[8] Ye J, Yeh Y T, Xue Y, et al. Proceedings of the National Academy of Sciences of the United States of America, 2022, 119(23): e2118836119.
[9] Ho C S, Jean N, Hogan C A, et al. Nature Communications, 2019, 10: 4927.
[10] Viel F, Maciel R C, Seman L O, et al. IEEE Access, 2023, 11: 24835.
[11] Chang M, He C, Du Y, et al. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2024, 305: 123475.
[12] Ali M N, El-Dahshan E S A, Yahia A H. Circuits Syst. Signal Process, 2017, 36: 4482.
[13] Zhang J, Zhou L, Wang L, et al. IEEE Transactions on Medical Imaging, 2022, 41(10): 2814.
[14] Xu J, Yi X, Jin G, et al. ACS Chemical Biology, 2022, 17(2): 376.
[15] Chefer H, Gur S, Wolf L. IEEE Conference on Computer Vision & Pattern Recognition (CVPR), 2021: 782.
