Dissertation Defence: Raman spectroscopy and deep learning for tumour radiation response evaluation
January 15, 2025 at 10:00 am - 2:00 pm
Alejandra Maria Fuentes Nunez, supervised by Dr. Andrew Jirasek, will defend their dissertation titled “Raman spectroscopy and deep learning for tumour radiation response evaluation” in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Medical Physics.
An abstract for Alejandra Maria Fuentes Nunez’s dissertation is included below.
Examinations are open to all members of the campus community as well as the general public. Registration is not required for in-person exams.
ABSTRACT
Altered tumour cell metabolism contributes significantly to radiotherapy resistance. Identifying the biomolecular alterations driving a tumour’s radiation response may facilitate the development of personalized radiotherapy. Raman spectroscopy (RS) is a non-destructive optical technique that examines a sample’s molecular content and can help determine signatures of radiation response in tumour cells and tissues. Due to the multiplexed nature of RS, Convolutional Neural Networks (CNN) may optimize Raman data analysis by automating the feature extraction process. CNNs extract features directly from spectral data by fitting supervised predictive models, exceeding the classification performance of traditional machine learning. This dissertation aims to develop a CNN-based framework to enable automated RS-based evaluation of tumour radioresponse through three main objectives.
The first objective is to develop, train, and evaluate a CNN classification model for detecting post-irradiation biochemical changes in the RS of breast tumour xenografts. The CNN effectively classified irradiated versus nonirradiated tissue spectra, obtaining accuracies above 92.1% for data collected 3 days post-irradiation. Moreover, the CNN demonstrated higher classification accuracy than previously developed semi-supervised methods.
The second objective is to apply a CNN-based technique, Gradient-Weighted Class Activation Mapping (Grad-CAM), to visualize key discriminative RS signatures influencing the CNN predictions. A CNN was trained on a RS data set acquired from three irradiated tumour cell lines. The CNN classified radiosensitive versus radioresistant cell spectra with 99.8% accuracy. Grad-CAM heatmaps exhibited distinct class-specific patterns of critical Raman peaks. Heatmaps of radioresistant cell spectra displayed contributions from glycogen and amino acids peaks. Conversely, radiosensitive cell heatmaps showed contributions from lipid peaks, in agreement with previous findings.
The third objective is to develop a convolutional autoencoder (AE) architecture to perform single-step automated correction of noise, background and cosmic rays (CR) from RS acquired from irradiated tumour cells and tissues. The AE efficiently removed all three artifacts from 11000 spectra within 2.4 seconds, obtaining an overall percentage root mean squared difference (PRD) below 3.4% between the AE-corrected test spectra and their corresponding target data (pre-processed by current in-house algorithms).
Overall, this dissertation supports the development of a RS and CNNbased framework for automated characterization of tumour radiation response.
