Nicholas Nuechterlein, Qualifying Project Presentation

In this paper, we use standard magnetic resonance (MR) imaging to predict specific unfavorable genomic aberrations, called copy-number variations (CNVs), in glioblastoma multiforme (GBM) patients. We consider 46 GBM patients with MR data from The Cancer Imaging Atlas (TCIA) and CNV labels from The Cancer Genome Atlas (TCGA). To address the high-dimensionality of the MR volumes, we take a radiomic approach. We extract shape, histogram, and texture features using classical computer vision techniques. We develop several feature selection methods to choose a small meaningful subset of these extracted features to feed into a set of traditional machine learning classifiers. Our primary feature selection method achieves a cross-validated area under the curve (AUC) score of 0.82 when evaluated with penalized logistic regression, roughly 50% higher than human experts. We also use transfer feature selection to repurpose features selected from different classification tasks on a separate MR dataset for our CNV classification task: this method archives an AUC score of 0.728. Finally, we focus on interpretability. We trace features found during feature selection to the tumor regions from which they were extracted. We suggest this as a method to direct the human eye toward possible visual patterns that may distinguish CNV classes.