Towards causal prediction on Magnetic Resonance Imaging including non-imaging data

Louisa Fay, Florian Marencke, Bin Yang, Sergios Gatidis, Thomas Kuestner

Keywords: Causality, MRI, deep learning, multi-modal pre-training, tabular data

Abstract: Deep learning methods can detect correlations in data but they cannot determine underlying causal relationships. Understanding causality, however, is essential because spurious correlations can obscure the true relationships in the data. In many large studies, imaging data is accompanied by additional tabular (non-imaging) clinical data. Our aim is to use the non-imaging information to learn a multi-modal feature representation that can make predictions based on learned causal dependencies while avoiding spurious correlations. This work presents our first preliminary results and outlines our future investigations.

Submission Number: 96