Master of science Mathias Novik Jensen will Thursday December 7th, 2023, at 12:15 hold his disputas for the PhD degree in Science. The title of his thesis is:
"Raman-spectroscopy of extracellular vesicles and self-supervised deep learning"
This thesis explores the prospect of a waveguide device for optical trapping and Raman spectroscopy of single biological nanoparticles. The aim of the work is to develop a Raman-on-chip device capable of training and measuring single particles in individual trapping sites, such that the throughput of the device can be increased through parallelization. The challenge posed by the induced Raman scattering in the waveguide device, producing a background in the measured Raman spectra of the nanoparticles, is investigated. It is found that the candidate UV-written silica waveguides produce a Raman spectrum lower than -107.4 dB in a 1 cm waveguide, which is 15 dB lower than silicon nitride. The background is found to exhibit no prominent features in the biological fingerprint region (800-1700 cm-1). Machine learning is explored for mitigating the significance of the background. A general platform for a machine learning model for this purpose is developed, first as a convolutional neural network is developed for analysis of tomographic scans of silicon boules. The developed neural network demonstrates an accuracy of 98.7% and robustness to a noise increase of up to 18 dB. The convolutional neural network is further expanded into a convolutional autoencoder for Raman spectra. The autoencoder model is based on the convolutional neural network is shown to be able to recover Raman spectra of extracellular vesicles with high fidelity, even in the presence of stochastic noise and an emulated waveguide background, recovering the spectra from a signal-to-noise ratio of -18±3 dB to 5.4 dB. The model is also shown to be able to differentiate the extracellular vesicles by their biological origins through self-supervised learning. Further developments of the model are demonstrated to enable it to differentiate 13 different origins, allowing a classifier to classify the nanoparticles to their known origins with an accuracy of 92.2%.
The disputas will be led by Professor Arne Smalås, Dean at the Faculty of Science and Technology.