Rapid Detection Mechanism And Application Of Pathogens Based On Raman Optical Tweezer

Infectious diseases caused by pathogenic bacteria have become a significant challenge in the field of global public health,especially in the context of the continuous emergence of drug-resistant pathogenic bacteria,which pose an increasing threat to human health.However,existing pathogenic bacteria detection technologies have limitations,such as difficulty in single-cell detection,time-consuming cell culture,and destructive effects on cells,making it challenging to meet the needs of rapid detection.Therefore,the rapid detection of pathogenic bacteria and their drug resistance has become a vital issue.This study aims to develop a fast,single-cell level,and non-destructive detection method for pathogens and their drug resistance based on optical tweezers Raman technology and deep learning to improve detection speed and accuracy.This study conducted an in-depth study on the rapid detection of clinically common pathogenic bacteria and their drug resistance.First,to achieve this goal,it is necessary to establish a Raman spectral dataset covering a rich variety of pathogenic bacteria.However,there is still a lack of such data sets in China,which cannot meet the needs of clinical applications.Therefore,this study established a large-scale single-cell Raman spectral dataset for the first time in China and used a onedimensional residual network(Residual Network,ResNet)to realize the rapid identification of pathogenic bacteria.Second,drug-resistant pathogenic bacteria are the main pathogenic bacteria that cause serious hospital infections.Conventional antibiotic susceptibility testing require a long culture time or high cost,so we propose a rapid identification method for drug-resistant pathogens based on ROT and deep learning.Finally,due to the high cost of sample preparation and spectral acquisition time,the shortage of data has become a key issue hindering the further development of deep learning in the field of bacterial Raman spectroscopy.In addition,models trained on small-scale datasets are prone to overfitting,resulting in poor generalization ability.In order to deal with the above two challenges,this study adopted the method of transfer learning and carried out an important exploration of the small-scale data set for the spectral identification of pathogenic bacteria based on deep learning.These three research topics complement each other and provide a potential solution for the identification of clinical pathogens and the rapid detection of drug resistance.The main contents of this paper include the following aspects:The main contents of this thesis include the following:1.A rapid detection technique for pathogenic bacteria based on ROT and ResNet was established.(1)We optimized sample preparation and experimental conditions to achieve high-quality single-cell Raman spectra of pathogenic bacteria using ROT for in situ measurements and established the largest Raman spectral database of pathogenic bacteria in China.(2)Two neural network models,ResNet and RamanNet,were developed for the classification of pathogenic bacteria’s Raman spectra.Compared with the classical machine learning algorithms SVM(81.08%)and LDA(82.67%),the recognition accuracy of these two models was higher,at 94.53%and 96.04%,respectively.Moreover,RamanNet,as a lightweight model,greatly accelerates the training of the deep learning model.(3)We developed a data augmentation method applicable to pathogenic bacteria Raman spectra,which improves the recognition accuracy of Raman spectra of the ResNet model and RamanNet model by 2.16%and 1.92%,respectively.(4)We appied a CNN visualization method based on t-SNE and Grad-CAM,which makes the convolutional neural network for the spectral analysis field more interpretable.We believe that deep learning methods with strong interpretability will be a future trend in Raman spectral analysis.2.A rapid identification method of pathogenic bacteria resistance based on ROT and RamanU-Net was established.(1)High-quality and highly reproducible single-cell Raman spectral fingerprint information of drug-resistant and sensitive strains of Escherichia coli,Klebsiella pneumoniae and Staphylococcus aureus was obtained using ROT technique;(2)The single-cell Raman spectra of drug-resistant and sensitive strains of the three pathogenic bacteria were compared,and it was found that the differences between them might be closely related to the nucleic acid/protein ratio;(3)A U-Net-based architecture classification model RamanU-Net,which exhibited higher recognition accuracy of more than 98%compared to SVM,LDA,ResNet and RamanNet in three bacterial drug resistance detection tasks.The results in this chapter show that RamanU-Net combined with ROT technology has great potential for drug resistance detection of clinical pathogenic bacteria.3.A rapid detection method for pathogenic bacteria based on Raman spectroscopy and migration learning is developed.(1)Three deep learning models,ResNet,RamanNet,and RamanU-Net,developed in this study were evaluated on the Bacteria-ID public dataset and PKU-Bacterial selfbuilt dataset.(2)The results illustrate that the PKU-Bacterial dataset and the Bacteria-ID dataset can migrate to learn from each other,and the transfer learning can improve the classification performance of the three models on both datasets,especially when pre-training with a high signalto-noise ratio spectrum(PKU-Bacterial dataset).(3)RamanU-Net,designed for Raman spectralclassification for bacterial resistance detection,achieves the highest classification accuracy of 95.3%for MRSA/MSSA on the Bacteria-ID dataset after migration learning,which is the best performance so far.In contrast,RamanNet,which is better at differentiating Raman spectra of different species,achieves the highest accuracy of 87.2%in the 30-isolates task after transfer learning,which is also the best result so far.In conclusion,this paper investigates and develops three pathogenic bacterial detection and drug resistance identification techniques.These techniques are progressive and complementary in clinical pathogenic bacteria diagnosis.The relevant indices of bacterial identification have been reliably verified in each experiment,indicating that the results of this study have promising application scenarios and hold the potential to provide possible solutions for the accurate and rapid detection of clinical pathogenic bacteria.