Variational Autoencoders for Dimension Reduction of Radiomic Data

Precision medicine is a medical model that proposes the customization of healthcare, with medical decisions and treatments being tailored a specific patient. Radiomics is an emerging form of precision medicine that uses machine learning on quantitative biomarkers extracted from computed tomography (CT) or magnetic resonance imaging images and is fast, low-cost, and non-invasive. However, due to a small dataset and large number of features, radiomic data suffers from the small-n-large-p problem. Thus, improvements in dimension reduction will allow for radiomic data to enhance medical diagnostics. In this study, Variational Autoencoders (VAE) were tested as a potential form of radiomic feature reduction in the radiomics machine learning pipeline in an offline and supervised manner. This serves to reduce the dimensions of the radiomic feature map and improve diagnostic performance of binary classification of two-year survival for CT scans of non-small cell lung cancer patients. It was found that a supervised VAE performed as well, if not slightly better than leading dimension reduction methods, with the added benefit of diminished class overlap in the reduced features.