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

Machine learning has immense potential to enhance diagnostic research in a wealth of medical applications. Advances in medical imaging have made machine learning applications in clinical oncology possible, including cancer diagnosis and prognosis. Radiomics is the extraction of quantitative imaging features to characterize tumor intensity, shape, and texture. Image-based features can then be used for data mining and precision medicine. Strong predictive models can be used for early detection and treatment planning which can be critical in a patient’s outcome. Malignant brain tumors account for 29.7% of brain cancers [40]. Parenchymal metastatic (MET) lesions and glioblastomas (GBM) account for the majority of malignant brain tumors. Lymphomas (LYM) are less common with an incidence of less than 2%. However, characterization of these tumors from MRI imaging is difficult due to the similarity of their radiologically observed image features. Machine learning could aid diagnostics by improving the classification of these common brain tumors and providing a more accurate preoperative diagnosis without additional invasive testing. While the number of patients is in the hundreds, the number of radiomic features is an order of magnitude larger, making these classification tasks subject to the curse of dimensionality. Therefore, dimensionality reduction is needed to balance feature dimensionality and model complexity. Several techniques have been designed to extract useful features and reduce the dimensionality that burdens machine learning tasks. The most widely used are principal component analysis (PCA), factor analysis, and, more recently, autoencoders. Autoencoders are a form of unsupervised representation learning that can be used for dimensionality reduction. It is similar to PCA but uses a more complex and non-linear model to learn a compact latent space. This thesis examines the effectiveness of autoencoders for dimensionality reduction on the radiomic feature space of multiparametric MRI images. Specifically, autoencoders’ effect on predictive performance for identifying different malignant brain tumors: GBM, LYM, and MET. This thesis further aims to address the class imbalances imposed by the rarity of lymphomas by examining different approaches to increase overall predictive performance through multiclass decomposition strategies.