Evaluating autoencoders for the dimensionality reduction of MRI-derived radiomics and classification of malignant brain tumors

Over the decades, machine learning has made great strides in medicine. Advances in medical imaging coupled with the booming data volumes pave the way for machine learning techniques (MLTs) in clinical applications from early disease detection to prognosis. Clinically, the diagnosis and characterization of disease often require the amalgamation of years of physician’s experience coupled with the timely analysis of multiple images. Modern machine learning algorithms can autonomously extract non-linear patterns from complex images like magnetic resonance imaging (MRI) and learn informative features, which could aid in a physician’s diagnosis. These characteristics of MLTs serve to fields plagued with high volumes of complex data. While the volume of images is vast, the radiomic features extracted from these images can be magnitudes higher, requiring dimensionality reduction for their use in MLTs. Numerous algorithms exist to reduce data dimensionality; autoencoders are a form of unsupervised representation learning designed to identify non-linear patterns to compress the input into a compact representation. This thesis proposes using autoencoders for dimensionality reduction of MRI-based radiomic features to improve the preoperative characterization of malignant brain tumors.