Open Collections

Suwito, Jason Samuel

According to the 2023 Canadian Cancer Statistics, 1 in 4 Canadians will die from cancer. For males, prostate cancer accounts for 1 in 5 new diagnoses. One of the ways to reduce cancer mortality is through early diagnosis. Medical physicists have been developing different diagnostic methods for cancer detection and radiation treatment analysis including using Raman spectroscopy. However, one of the main limitations of Raman spectroscopy is that the data is very prone to saturation and cosmic rays making some of its spectra unusable. Hence, there is an opportunity to utilize machine learning for imputing the missing spectra. This study aims to compare the performance of known imputation methods for spectrometry-based data such as Random Forest, Quantile Regression Imputation of Left-Censored Data (QRILC), and K-Nearest Neighbour (KNN) with alternative imputation methods that involve weights that incorporate the spatial component of where on the tissue the spectra are measured. The results reveal that spatial imputation methods outperform the regular imputation method. This implies that there is a spatial relationship between spectra in a Raman spectroscopy matrix where spectra that are closer are more correlated than spectra that are further away.

Text

University of British Columbia. DATA 449

Vancouver : University of British Columbia Library

10.14288/1.0444822

Science, Irving K. Barber Faculty of (Okanagan); Computer Science, Mathematics, Physics and Statistics, Department of (Okanagan)

Undergraduate

DSpace