Keynote Talk #2: Multimodal Data Analysis with Applications in Imaging Genomics

Abstract

Multi-modality matched datasets in healthcare capture information about the disease of the same patient from multiple views, often at different physical scales, thereby providing a more complete picture of complex diseases like cancer. We present a novel extension of the canonical correlation analysis (CCA) framework that takes into account underlying dependencies within individual modalities to better capture correlations between two modalities. We demonstrate the utility of the resulting embedding space as a fusion module in survival prediction for breast cancer patients using histology imaging and genomics data.