A Unified Approach to Analysis of MRI Radiomics of Glioma Using Minimum Spanning Trees

Abstract

Radiomics has shown great promise in detecting important genetic markers involved in cancers such as gliomas, as specific mutations produce subtle but characteristic changes in tumor texture and morphology. In particular, mutations in IDH (isocitrate dehydrogenase) are well-known to be important prognostic markers in glioma patients. Most classification approaches using radiomics, however, involve complex hand-crafted feature sets or “black-box” methods such as deep neural networks, and therefore lack interpretability. Here, we explore the application of simple graph-theoretical methods based on the minimum-spanning tree (MST) to radiomics data, in order to detect IDH mutations in gliomas. This is done using a hypothesis testing approach. The methods are applied to an fMRI dataset on n = 413 patients. We quantify the significance of the group-wise difference between mutant and wild-type using the MST edge-count testing methodology of Friedman and Rafsky. We apply network theory-based centrality measures on MSTs to identify the most representative patients. We also propose a simple and rapid dimensionality-reduction method based on k-MSTs. Combined with the centrality measures, the latter method produces readily interpretable 2D maps that reveal distinct IDH, non-IDH, and IDH-like groupings.