A Comprehensive Radiogenomic Feature Characterization of 19/20 Co-gain in Glioblastoma

Abstract

The prognosis and treatment planning of glioblastoma multiforme (GBM) involves a holistic analysis of imaging, clinical, and molecular data. The correlation of imaging and molecular features has garnered much interest due to its potential to reduce the number of invasive procedures on a patient and resource utilization of the overall prognostic and treatment planning process. This article detects and characterizes the impact of tumor biomarkers (such as shape, texture, location, and the tissue surrounding the tumor) in detecting a prognostic mutation – the concurrent gain of 19 and 20 chromosomes, and proposes two novel ideas for this analysis. First, to address the challenges associated with the limited, diverse, and complex nature of medical data, this article proposes a novel generative model – the realistic radiogenomic design using disentanglement in generative adversarial networks (R2D2-GAN), designed to recreate highly subtle, unapparent manifestations of mutations in magnetic resonance imaging. It generates high-resolution, diverse data that captures the discriminatory visual features of the molecular markers while tackling the high diversity, unbalanced, and limited GBM data with rare mutations correlating with patient survival such as 19/20 co-gain. Second, this study proposes a quantitative metric called the synthetic image fidelity (SIF) score to evaluate the performance of GANs in learning visually unapparent prognostic features through the use of gradient-based model explanations. Results are compared with current methods.