Date of Graduation
8-2018
Document Type
Thesis (MS)
Program Affiliation
Biostatistics, Bioinformatics and Systems Biology
Degree Name
Masters of Science (MS)
Advisor/Committee Chair
Arvind Rao
Committee Member
Richard Wendt
Committee Member
Ankit Patel
Committee Member
Ganesh Rao
Committee Member
David Fuentes
Abstract
A complete codeletion of chromosome 1p/19q is strongly correlated with better overall survival of diffuse glioma patients, hence determining the codeletion status early in the course of a patient’s disease would be valuable in that patient’s care. The current practice requires a surgical biopsy in order to assess the codeletion status, which exposes patients to risks and is limited in its accuracy by sampling variations. To overcome such limitations, we utilized four conventional magnetic resonance imaging sequences to predict the 1p/19q status. We extracted three sets of image-derived features, namely texture-based, topology-based, and convolutional neural network (CNN)-based, and analyzed each feature’s prediction performance. The topology-based model (AUC = 0.855 +/- 0.079) performed significantly better compared to the texture-based model (AUC = 0.707 +/- 0.118) while comparably against the CNN-based model (0.787 +/- 0.195). However, none of the models performed better than the baseline model that is built with only clinical variables, namely, age, gender, and Karnofsky Performance Score (AUC = 0.703 +/- 0.256). In summary, predicting 1p/19q chromosome codeletion status via MRI scan analysis can be a viable non-invasive assessment tool at an early stage of gliomas and in follow-ups although further investigation is needed to improve the model performance.
Keywords
machine learning, deep learning, magnetic resonance imaging, diffuse glioma, 1p/19q codeletion, radiomics, topology, data mining