Publication Date

Summer 7-8-2020

Abstract

Spectral quality classification is a vital step in data cleaning before the

analysis of magnetic resonance spectroscopy (MRS) data can be done. This

analysis compares five methods of quality classification; three of these are

legacy methods, Maudsley et al. (2006), Zhang et al. (2018), and

Bustillo et al. (2020), and two newly created methods that used a random forests

classifier (RFC) to inform their classifications. We found that the random forest

classifier was the most accurate at predicting spectra quality (balanced

accuracy for RF of 88% vs legacy of 70%, 72%, or 72%). A

Random-Forests-Informed Filtering method (RFIFM) for quality classification was

created by bounding four of the highest ranking features in the RFC to mimic

the classification methods of the legacy methodologies. The RFIFM had only

slightly decreased accuracy compared to the RFC (85% vs 88%), but still

outclassed the legacy methods. Overall, the top features in the RFC show that

the best measures of quality relate to the frequency of the metabolite peaks in

the spectra.

Degree Name

Statistics

Level of Degree

Masters

Department Name

Mathematics & Statistics

First Committee Member (Chair)

Erik Barry Erhardt

Second Committee Member

Rhoshel Lenroot

Third Committee Member

Fletcher G. W. Christensen

Language

English

Keywords

Spectroscopy, MRI, MRSI, Random Forests, Quality, Machine Learning

Document Type

Thesis

Download

Share

COinS