AI Center Seminar - AI for Health Series - Dr. Jonas Richiardi

The talk is organized by the EPFL AI Center as part of the AI for Health seminar series.

Hosting professor: Dr. Dorina Thanou

Title
From brains to hearts, graph-based machine learning as a unifier for prediction on spatio-temporal medical imaging data

Abstract
Interactions between parts of a biological system are a central concept in biomedical sciences. These interaction networks can be mathematically modelled as labelled graphs, and this type of approach has been used across scales for genes, proteins, cells, organs, or individuals. Coupled with machine learning, graph representations have many promising applications for precision medicine, including differential diagnosis, treatment planning, survival modelling, and prognosis.

In this talk, we will discuss how graph-based learning can be developed and applied to spatio-temporal medical imaging data, with a focus on brain and heart imaging. We will start with an introduction to spatio-temporal imaging data, then show how "organ graphs" can be defined and computed for both brain and heart, yielding an expressive and compact meso-scale representation of each patient. We'll introduce a new statistical estimator for spatio-temporal correlation in brain imaging, the local correlation of averages, which exhibits superior theoretical and empirical properties.

With representations addressed, we will then transition to defining the machine learning tasks that can be addressed given an organ graph representation, and give a rapid overview of existing approaches. As an example of a novel approach, we'll focus on a newly proposed graph neural process model, which combines multiplex graphs with neural ordinary differential equations and neural processes to yield promising performance in spatio-temporal trajectory reconstruction, interpolation, and graph classification for cardiac imaging data, as well as a latent space that reflects known underlying pathophysiological features of cardiac disease. 

Finally, we will show some recent empirical benchmark results on graph neural networks for regression on brain graphs, with applications to multimodal graphs, (where edges are vector-valued), transfer learning between regression tasks, and hyperbolic graph neural networks.

Bio
Jonas Richiardi is a Principal Investigator and Senior Lecturer at the Department of Radiology, Lausanne University Hospital, Switzerland, and heads the Translational Machine Learning Laboratory (https://unil.ch/tml). He is also the section head of the Imaging for Precision Medicine section (https://cibm.ch/research/projects/imaging-for-precision-medicine/), part of the Data Science Module of the CIBM Center for Biomedical Imaging.

Previously, he was Clinical Research Lead at Siemens Healthcare, a Marie Curie fellow in Neurology at Stanford University and the University of Geneva, and a post-doctoral researcher in the Medical Image Processing Lab (EPFL/UNIGE). He obtained his Ph.D. at EPFL in the Laboratory of the Dalle Molle Institute for Perceptual Artificial Intelligence, Signal Processing Institute, and and his M.Phil. from the university of Cambridge's Engineering Department and Computer Laboratory.

His research interests include machine learning for complex multimodal biological data, in particular magnetic resonance brain imaging data and its combination with -omics data. Methods development are focused on graph-based machine learning approaches for spatio-temporal imaging data, learning from scarce and heterogeneous data, and multimodal approaches. Applications to precision medicine include diagnosis, treatment selection, prognosis, and treatment response prediction, in particular for stroke, cardiovascular disease, and oncology. In parallel, he leads efforts to develop imaging data science infrastructure so that these techniques can be applied to messy, hospital-scale clinical routine data.