Optical diagnostics: deep-learning identification of bacterial species via sers spectra

Student thesis: Master's Thesis

Abstract

Bacterial infections are a major cause of death in both advanced and developing countries. Additionally, the cost of treating these infections is significant for the economy. Early detection of pathogenic bacteria in food, water, and bodily fluids holds paramount importance, although it remains a great challenge due to the complexities of the samples and the need for rapid screening of substantial sample volumes. In response to this issue, we have introduced a deep neural network (DNN) that showcases the capacity to distinguish different types of bacteria through the utilization of surface-enhanced Raman spectroscopy
(SERS). Our approach employs a vision transformers-based DNN to identify various types of bacteria via a label-free SERS technique. We trained both the vision transformer model and the Swin transformer (a modified version of vision transformers) using the high signalto-noise ratio (SNR) data from SERS. SERS is implemented using metal nanosubstrates and nanoparticles integrating in microfluidic systems for a variety of applications e.g. drug and cell detection. The enhancement factor of SERS can be as large as ~1011. This increased
sensitivity, makes SERS a suitable candidate for the analysis of small concentrations of (bio-)molecules/cells for diagnosing bacterial infections. These infections can be caused by the presence of different pathogenic bacteria and/or their toxic excretions or even by a body’s own immune response. These infections can lead to various health issues and complications, ranging from mild discomfort to severe illness or even life-threatening conditions.
Identifying the specific bacteria responsible for an infection is crucial for several reasons such as knowing the exact bacteria type helps healthcare professionals determine the appropriate treatment strategy. Secondly, identifying the specific bacteria allows Healthcare facilities to prevent the spread of these bacteria among patients, reducing the risk of outbreaks. Furthermore, bacterial identification aids in understanding the epidemiology of infections to monitor and manage disease outbreaks. However, many layers of complexity stand between these pathogenic bacteria and an accurate and sensitive detection. The main complexities are the following: 1 – the size of the bacteria relative to the nanostructures, 2 – limited chemical surface interactions, 3 – the matrix complexity of a bacteria’s natural environment, 4 – noise and overlap from non-pathogenic bacteria. The aim of this thesis is to address these complexities by utilizing SERS enhancement in conjunction with cuttingedge Vision Transformers models. The project starts with a SERS investigation of safe
bacterial samples in a suspension using a microplate with a quartz cover, with a Renishaw InVia confocal Raman microscope. using microplate will consequently lead to the on-chip study of different bacterial species which is for high-throughput analyses. Microplates consist of an array of wells in a compact format, which allows multiple samples or reactions to be conducted simultaneously. They offer benefits such as increased experimental efficiency, reduced reagent consumption, and the ability to automate processes. Obtained
results indicate that classification of bacteria is successfully done by our vision transformer and Swin transformer model, in this process, these results are compared with the simulations as well as with the state-of-the-art results. The final results of such research can help the doctors in fastening their decision when treating a patient, hence improving the rate of mortality and recovery of the patients. Furthermore, the wide-spread use of antibiotics is causing antibiotic resistance in common pathogenic bacteria. An accurate and rapid
diagnosis that will maximally limit the scope of prescribed antibiotics will have worldwide benefits.
Date of Award2023
Original languageEnglish

Cite this

'