Preview

PDF, English - main document Download (8MB) | Lizenz: Bayesian Deep Learning Framework for Radiogenomics Analysis of Cancer by Jayme, Alejandra underlies the terms of Creative Commons Attribution-NonCommercial 4.0

Abstract

Glioblastoma Multiforme (GBM) is a highly aggressive and heterogeneous brain tumor, requiring innovative approaches for prognosis and treatment planning. This thesis investigates multiview modeling of imaging-genetics data and the development of a Bayesian deep learning framework to address challenges in integrating multimodal data, quantifying uncertainty, and improving interpretability. Multiview modeling integrates imaging, genomics, and clinical data to uncover complementary insights. Deep learning methods leverage non-linear relationships across multimodal data and often outperform traditional ap- proaches. The Bayesian deep learning framework combines Bayesian Neural Networks (BNNs) to quantify uncertainty and Bayesian Belief Networks (BBNs) to enhance interpretability by revealing feature dependencies. Al- though the framework performed worse at prediction compared to con- ventional methods, it provided cautious predictions and valuable insights, making its outputs more actionable for clinical decision-making. This work highlights the potential of multiview modeling and Bayesian deep learning to improve data integration and interpretability for survival analysis in GBM, laying a foundation for future research on advanced mod- eling strategies and clinical applications.